Cloud

Ethics
Cloud
Louise Amoore

Ethics
Algorithms
and the Attributes
of Ourselves
 
and Others

Duke University Press

Durham and London
2020
© 2020 Duke University Press
All rights reserved
Printed in the United States of America on acid-­free paper ∞
Designed by Drew Sisk
Typeset in Portrait Text, SangBleu Kingdom, and Helvetica Neue
by Copperline Book Services

Library of Congress Cataloging-­in-­Publication Data

Names: Amoore, Louise, author [date].
Title: Cloud ethics : algorithms and the attributes of ourselves
and others / Louise Amoore.
Description: Durham : Duke University Press, 2020. | Includes
bibliographical references and index.
Identifiers: lccn 2019033257 (print) | lccn 2019033258 (ebook) |
isbn 9781478007784 (hardcover) | isbn 9781478008316 (paperback) |
isbn 9781478009276 (ebook)
Subjects: lcsh: Cloud computing—Moral and ethical aspects. |
Decision making—Moral and ethical aspects. | Algorithms—Moral
and ethical aspects.
Classification: lcc qa76.585 .a46 2020 (print) |
lcc qa76.585 (ebook) | ddc 004.67/82—dc23
lc record available at https://lccn.loc.gov/2019033257
lc ebook record available at https://lccn.loc.gov/2019033258

Cover art: James Turrell, Skyspace Seldom Seen, 2004. Houghton Hall,
King’s Lynn, Norfolk, England. Courtesy of Houghton Hall Archives.
Photograph by Louise Amoore.
 For my family, with love
&
For my students, with thanks for their curiosity
This page intentionally left blank
Any true scientific knowing is always, like feeling, only partial.
 — John Fowles

The possibility of the ethical lies in its impossibility.

 — Drucilla Cornell
This page intentionally left blank
C ON T EN T S

xi
Acknowledgments

1 Introduction
Politics and Ethics in the Age of Algorithms

Part 1. Condensation

29 1 The Cloud Chambers:

Condensed Data and Correlative Reason

56 2 The Learning Machines:

Neural Networks and Regimes of Recognition

Part 2. Attribution

85 3 The Uncertain Author:

Writing and Attribution

108 4 The Madness of Algorithms:

Aberration and Unreasonable Acts

Part 3. Ethics

133 5 The Doubtful Algorithm:

Ground Truth and Partial Accounts

154 6 The Unattributable:

Strategies for a Cloud Ethics

173
Notes

197
Bibliography

212
Index
This page intentionally left blank
ACK NOW LED GM EN T S

The writing of this book coincided with what has felt like the most difficult
and dangerous of political times I have experienced. The UK stands on the
brink of leaving the community of European peoples that it joined in my first
year of life, and far-­right political forces are gathering at every turn. So many
of the things in the world that I hold dear seem to be under threat: the right
of all young people to have an education funded by those who have already
benefited in the same way; the obligation on all people and governments to
welcome and support those who flee from violence and hardship; the capacity
of a democratic society to challenge hate and prejudice; the rights of women,
black and ethnic minorities, and lgbtq+ people to make political claims in
the world without risk to their lives and well-­being. Following the rising pow-
ers of algorithms, as they have begun to dominate economic and political life,
my fears for the future are amplified. Let me be clear, machine learning algo-
rithms that anticipate our future propensities are seriously threatening the
chances that we have to make possible alternative political futures. In this con-
text, the writing of this book has been challenging for its constant reminder
of the limiting of potential political futures. But it has also been intensely up-
lifting for the incredible generosity of others in encouraging my work. One of
the overriding feelings I have as I complete this book is that young people have
the full measure of what is at stake ethicopolitically in an age of algorithms.
After every lecture I have given, over every discussion in corridors and cafes,
I have been struck by the depth and force of the understanding and innova-
tive thought of the next generation of young scholars and students. There is
much to be hopeful for, and I dedicate this book to this sense of hopefulness
and curiosity.
I have worked in the Department of Geography at Durham University for
fourteen years. It is not easy to describe how a group of people comes to matter
so greatly to whether it will be possible to conduct the work one most wishes
to. I can only say that the mood of experimentation, the quiet support of cu-
rious juxtapositions of theory and method, the standing together on a snowy
union picket line, and the quite breathtaking generosity in reading work has
been vital to me. Thank you all — you are pretty incredible: Ben Anderson, An-
drew Baldwin, Oliver Belcher, Mike Bentley, Harriet Bulkeley, Sydney Calkin,
Kate Coddington, Rachel Colls, Mike Crang, Niall Cunningham, Jonny Dar-
ling, Alex Densmore, Nicky Gregson, Paul Harrison, Sarah Hughes, Elizabeth
 Johnson, Sarah Knuth, Paul Langley, Lauren Martin, Cheryl McEwan, Colin
McFarlane, Joe Painter, and Jeremy Schmidt. The past few years have also been
extraordinary for the PhD students I have been so lucky to supervise: Alice
Cree, Arely Cruz Santiago, Andrei Dinca, Pete Forman, Sarah Hughes, Olivia
Mason, Anna Okada, and Vanessa Schofield. I have never known any period
so exciting for the discussions you have provoked and the experimental meth-
ods you are pioneering. You all have so much courage, and you make me feel
xii more courageous.
When I reflect on the many forks in the path of writing this book, I can
Acknowledgments

remember quite clearly how particular questions from an audience, or com-

ments over lunch at a workshop, have opened avenues I had not imagined. I
am immensely grateful to the many people who have so generously hosted me
and engaged with my work. There are too many debts to others to recount
fully here, but I thank you profoundly, and I will signal some of the branching
points of my journey. A visit I made to the University of Wisconsin – Madison,
at the invitation of Eric Schatzberg, Alan Rubel, and their colleagues, had a
lasting effect on my work. The lecture audiences and workshop participants
at the Holtz Center for Science and Technology Studies and the Interna-
tional Information Ethics Conference have influenced my thought more than
they could possibly know — thank you. My time spent at Australian National
University, Canberra, brought conversations that have endured — thank you,
Fiona Jenkins and colleagues, for inviting me during your Divided Authorities
theme, and Fiona and William for hosting my visit. For discussions on robot-
ics, point clouds, and machine learning, I thank Jutta Weber, Oliver Leistert,
and their colleagues at University of Paderborn, Germany. Thanks to Nanna
Thylstrup and her collaborators on the Uncertain Archives project at Univer-
sity of Copenhagen, and to discussions with Luciana Parisi, Ramon Amaro,
and Antoinette Rouvroy — for experiments with algorithms and archives. Pe-
ter Niessen and Christine Hentschel at the University of Hamburg probably
did not realize how much their invitation to give a lecture in their Revenge
of the Bots series would indelibly mark the final stages of my writing; thank
you, generous audience, especially the computer scientists. Marc Boeckler, Till
Straube, and their colleagues at Goethe University, Frankfurt, generously in-
vited me as part of their Digital Disorders theme and made possible the most
inspiring interdisciplinary conversations. Thank you. Sometimes, the process
of writing can be a lonely affair, but these past years have felt full of conver-
sation and inspiration. I am grateful for the many times I have felt that my
work reached new audiences and different perspectives: thank you, Peter Bur-
gess and his students at the École Nationale Supérieure, Paris; Pip Thornton,
Andrew Dwyer, Joe Shaw, and others present in the audience at the Turing
Institute, London; Jeroen de Kloet and colleagues at the Amsterdam Centre
for Globalisation Studies (and Erna Rijsdijk and Mike Shapiro for asking ques-
tions that I can still recall); Nathaniel Tkacz, Celia Lury, and their colleagues
at the Warwick Centre for Interdisciplinary Methodologies for discussions of
data streams; Huub Dijstelbloem and the University of Amsterdam’s Politics
of Visualization group; Martin Kornberger, Ann-­Christina Lange, and their
Copenhagen Business School colleagues working on digital infrastructures; xiii
Gunes Tavmen and her colleagues at the University of London, Birkbeck, and

Acknowledgments
the Vasari Institute, for conversations on cloud ethics; Susan Schuppli, Luci-
ana Parisi, Jussi Parrika, and Eyal Weizman, for memorable discussions of the
architecture of algorithms; N. Katherine Hayles and all the participants in
the Thinking with Algorithms conference in Durham; Jamie Woodward, Maja
Zefuss, Noel Castree, Martin Coward, Veronique Pin-­Fat, and those in the au-
dience at the Manchester University 125 Years Anniversary lecture. You have
no idea how much your generosity has supported my work.
I am grateful for the patience and time afforded to me by many people whose
daily working lives bring them close to the lifeworlds of algorithms. Without
the insights of computer scientists, doctors, software developers, lawyers — 
and many others who were willing to share their accounts — it would not have
been possible to pursue my questions. It can be difficult to engage with new
vocabularies and practices, but this kind of difficulty has been so productive
in following threads in the dark and finding insight in the multiple meanings
of bias, attribute, decision, weight, and judgment. It seems to me that the basis
for ethics in an age of algorithms is not strictly a shared language and common
understanding but precisely the persistence of difference and the inexplicable.
The writing of this book was funded by the award of a Leverhulme Major
Research Fellowship, “Ethics of Algorithms” (2016 – 2018). I have been extraor-
dinarily fortunate to be afforded the time and resources to do the work that
I so wanted to be able to do. It has become much more difficult in the UK to
fund the aspects of scholarship that involve reading, thinking, presenting and
discussing work, and writing. Thank you, Leverhulme Trust, and may you
long continue to act against the grain of prevailing logics of research funds
in our times. During my fellowship, two extraordinary women have under-
taken my teaching duties: Sarah Hughes and Sydney Calkin. They are a tour
de force, and I thank them for everything they have done. Much of the field-
work research for the book was funded by the Economic and Social Research
Council (esrc) for research on “Securing against Future Events” (2012 – 16).
Throughout the project I worked in collaboration with Volha Piotukh — a post-
 doctoral researcher of simply astounding talent. Volha, my thanks will always
fall short. You simply never gave up — you got us into places for interviews and
observations that I would have said were impossible. You asked questions that
challenged, exposed, and opened new pathways. You are the human embodi-
ment of a whole new methodology for the social sciences and humanities to
engage science. Hanging out with you (long-haul flights, broken trains) has
been inspirational.
xiv I owe an immeasurable debt to Duke University Press editor Courtney
Berger, who has once more offered encouragement and thoughtful critique in
Acknowledgments

equal measure. Thank you to Jenny Tan and Sandra Korn for guiding me pa-
tiently through the editorial and production process, and Lalitree Darnielle
for working on the art and illustrations with such care. I am especially grate-
ful to the three anonymous reviewers for the incredible investment of their
time and insight in making this a better book. A version of chapter 1 appeared
previously as “Cloud Geographies: Computing, Data, Sovereignty,” Progress in
Human Geography 42, no. 1 (2018): 4 – 2 4.
Finally, friends and family. You know who you are, and you know how you
tolerated the tunnel vision that can be necessary to see something through.
Marieke de Goede, we have walked and talked in countless places from Turin
to San Diego over twenty-­two years, and every conversation has left its mark.
My gratitude to you is distilled to a dinner in Boston in April 2016 when you
put the whole show back on the road — thank you. Jonathan, you listened pa-
tiently to all the nonsense and supplied a different kind of distillation, craft
gin — a finer brother there cannot be. Grace and Tom — you are older now, and
though the words “mum is finishing a book” bring necessary shrugs of the
shoulders (and some giggling), you are so very kind about it. Paul, I only real-
ized in these past few weeks that we never work on completing manuscripts at
the same time. Like everything in our lives, they are shared across time, with
love and support. Thank you for making it possible, and it is your turn again
now.
IN T RODUC T ION

Politics and Ethics

in the Age of Algorithms

The mathematical proposition has been given the stamp of incontestability.

I.e.: “Dispute about other things; this is immovable — it is a hinge on which your
dispute can turn.”
— Ludwig Wittgenstein, On Certainty

“A Hinge on Which Your Dispute Can Turn”

It is March 2016, and I am seated in a London auditorium, the gray curve of the
river Thames visible from the windows. A tech start-­up business, specializing
in developing machine learning algorithms for anomaly detection, is present-
ing its latest algorithmic innovations to the assembled government and corpo-
rate clients. The projection screen displays a “protest monitoring dashboard”
as it outputs risk scores of “upcoming threats of civil unrest” in cities around
the globe, their names scrolling: Chicago, London, Paris, Cairo, Lahore, Is-
lamabad, Karachi. The score that the analyst reads from the dashboard is the
singular output of deep neural network algorithms that have been trained to
recognize the attributes of urban public life, the norms and anomalies latent
in the data streams extracted from multiple sources, from Twitter and Face-
book to government databases. As the presenter explains to the audience of
national security, policing, and border officials, “We train our algorithm to
understand what a protest is and is not,” and “it gets better,” “adapting day by
day,” as it iteratively learns with humans and other algorithms.1 The process of
learning “what a protest is” from the clustered attributes in data and modify-
ing the model continues when the algorithm is later deployed in the city or at
the border: “We give you the code,” he pledges, “so that you can edit it.” How
does an algorithm learn to recognize what a protest is? What does it mean to
cluster data according to the attributes and propensities of humans to gather
in protest or in solidarity? At the London event, as the presenter displays a still
 2
Introduction

Figure I.1  An image from Stephen Spielberg’s film adaptation of Philip K.

Dick’s novel Minority Report appears in a technology company’s presentation
to government and corporate analysts. Author’s photograph.

from Minority Report (figure I.1), my thoughts turn to the protests that took
place one year earlier, in the US city of Baltimore.
On April 12, 2015, Freddie Gray, a twenty-­five-­year-­old African American
man, sustained fatal injuries in the custody of the Baltimore Police Depart-
ment. The profound violence of Gray’s murder is an all-­too-­familiar event in
the racialized architectures of our contemporary cities. During the days that
followed his death, however, as people gathered on Baltimore’s streets to pro-
test the killing, the violence of the act extended into the plural actions of a set
of machine learning algorithms that had been supplied to the Baltimore Po-
lice Department and the US Department of Homeland Security by the tech
company Geofeedia. With the use of techniques similar to those described in
the London protest-­monitoring software, the Geofeedia algorithms had been
trained on social media data, analyzing the inputs of Twitter, Facebook, You-
Tube, Flickr, and Instagram and producing scored output of the incipient pro-
pensities of the assembled people protesting Gray’s murder. “Several known
sovereign citizens have begun to post social media attempting to rally per-
sons to demonstrate,” recorded the Baltimore Police Department in a memo
that promised to “continue to evaluate threat streams and follow all actionable
leads.”2 Indeed, Geofeedia went on to market its algorithms to other states on
the basis of a Baltimore “case study” (figure I.2) in which Freddie Gray is said
to have “passed away,” the city to have “braced itself for imminent protests,”
and the police to have seized “opportunities” to analyze “increased chatter
from high school kids who planned to walk out of class.”3
During those days in April, terabytes of images, video, audio, text, and bio- 3
metric and geospatial data from the protests of the people of Baltimore were

Politics and Ethics

rendered as inputs to the deep learning algorithms. Even the written text em-
bedded within social media images — such as the “police terror” placards car-
ried aloft and captured on Instagram — was extracted by a neural network and
became features in the algorithm.4 People were arrested and detained based
on the outputs of a series of algorithms that had — as the London scene also
proposed — learned how to recognize what a protest is, what a gathering of

Figure I.2  Geofeedia’s account of the Baltimore protests in the marketing

of software analyzing social media data for the detection of incipient public
protests. American Civil Liberties Union, 2016.
 people in the city might mean. As Simone Browne has argued in her compel-
ling account of the “digital epidermalization” of biometric algorithms, what
is at stake is the recognizability of a body as human, as fully political.5 Among
Baltimore’s arrests and detentions were forty-­nine children, with groups of
high school students prevented from boarding buses downtown because the
output of the algorithm had adjudicated on the high risk they posed in the
crowd.6 Based on the so-­called ground truth of features that the algorithms
4 had learned in the training data, the algorithms clustered the new input data
of people and objects in the city, grouping them according to their attributes
Introduction

and generating a numeric scored output.7

The profound violence of the killing of one man, and the residue of all
the past moments of claims made in his name, and in the name of others be-
fore him (note that the names Freddie Gray and Michael Brown persist in the
training of subsequent algorithms to arbitrate protest), becomes lodged within
the algorithms that will continue to identify other faces, texts, and signs in
future crowds. Understood as the principal architecture of what N. Kather-
ine Hayles calls the “computational regime,” what matters to the algorithm,
and what the algorithm makes matter, is the capacity to generate an action-
able output from a set of attributes.8 What kind of new political claim, not yet
registered as claimable, could ever be made if its attributes are recognizable in
advance? The very capacity to make a political claim on the future — even to
board a bus to make that claim — is effaced by algorithms that condense mul-
tiple potential futures to a single output.
At the level of the algorithm, it scarcely matters whether the clustered
attributes are used to define the propensities of consumers, voters, dna se-
quences, financial borrowers, or people gathering in public space to make a
political claim.9 Thus, when in 2016 Cambridge Analytica deployed its deep
learning algorithms to cluster the attributes of voters in the UK EU referen-
dum and the US presidential election, or when Palantir’s neural networks
supply the targets for the US ice deportation regime, what is at stake eth-
icopolitically is not only the predictive power of algorithms to undermine
the democratic process, determine the outcomes of elections, decide police
deployments, or make financial, employment, or immigration decisions. Of
greater significance than these manifest harms, and at the heart of the con-
cerns of this book, algorithms are generating the bounded conditions of what
a democracy, a border crossing, a social movement, an election, or a public pro-
test could be in the world.
Ethics of Algorithms

At first sight, the potential for violent harm precipitated by algorithms that
learn to recognize human propensities appears to be a self-­evident matter for
critique. Surely, one could say, the ethical terrain of the algorithm resides in
the broader political landscape of rights and wrongs, good and evil. After all,
one could readily identify a set of rights, already apparently registered as be-
longing to rights-­bearing subjects, that has been contravened by algorithms 5
that generate targets, adjudicating which people may peaceably assemble, or

Politics and Ethics

which people are worthy of credit or employment, and on what terms. Indeed,
on this terrain of delineating the rights and wrongs of algorithmic actions is
precisely where many critical voices on the harms of the algorithm have been
heard. Writing in the New York Times, for example, Kate Crawford identifies
machine learning’s “white guy problem,” arguing that “we need to be vigilant
about how we design and train machine learning systems.”10 The dominant
critical perspectives on algorithmic decisions have thus argued for remov-
ing the “bias” or the “value judgements” of the algorithm, and for regulating
harmful and damaging mathematical models.11 Within each of these critical
calls, the ethical problem is thought to dwell in the opacity of the algorithm
and in its inscrutability, so that what Frank Pasquale has called the “black
box society” is addressed through remedies of transparency and accountabil-
ity.12 In sum, the rise of algorithmic power in society has been overwhelmingly
understood as a problem of opaque and illegible algorithms infringing or un-
dercutting a precisely legible world of rights belonging to human subjects. In
such a framing, there is an outside to the algorithm — an accountable human
subject who is the locus of responsibility, the source of a code of conduct with
which algorithms must comply. To call for the opening of the black box, for
transparency and accountability, then, is to seek to institute arrangements
that are good, ethical, and normal, and to prevent the transgression of societal
norms by the algorithm.
Yet, when people gathered to protest on Baltimore streets, or when Face-
book users’ data fueled the political and commercial models of Cambridge An-
alytica (figure I.3), legible rights to peaceable assembly or to electoral due pro-
cess were not violated primarily by illegible algorithms. Rather, the means by
which people could appear in a political forum, the conditions of their appear-
ance, and the capacities they had to make a recognizable political claim in the
world were subject to algorithmic regimes of what Michel Foucault calls truth
telling and wrongdoing.13 In short, what matters is not primarily the identifi-
cation and regulation of algorithmic wrongs, but more significantly how algo-
Figure I.3  Cambridge Analytica advertises how “data drives all we do” in the
fused commercial and political pursuit of ways “to change audience behavior.”
Screenshot archived by the author in May 2018, when the firm ceased trading
under that name.

rithms are implicated in new regimes of verification, new forms of identify-

ing a wrong or of truth telling in the world. Understood in these terms, the
algorithm already presents itself as an ethicopolitical arrangement of values,
assumptions, and propositions about the world. One does not need to look
beyond the algorithm for an outside that is properly political and recogniz-
ably of ethics. Indeed, there can be no legible human outside the algorithm
and underwriting its conduct, for as John Cheney-­Lippold reminds us, we are
enmeshed in the data that produce each “freshly minted algorithmic truth.”14
One cannot sustain a search for codes of ethics that instill the good, the law-
ful, or the normal into the algorithm. Contemporary algorithms are not so
much transgressing settled societal norms as establishing new patterns of good
and bad, new thresholds of normality and abnormality, against which actions
are calibrated.
Actions one might consider harmful, as William Connolly notes, are not
merely “actions by immoral agents who freely transgress the moral law” but
are “arbitrary cruelty installed in regular institutional arrangements taken to
embody the Law, the Good, and the Normal.”15 Amid the widespread search
for new ethical arrangements for the scrutiny and regulation of algorithms,
what becomes of the arbitrary harms lodged within embodied algorithmic ar-
rangements? One could imagine a world in which the deep neural networks
used in cities like Baltimore are scrutinized and rendered compliant with rules
and yet continue to learn to recognize and misrecognize people and to infer
intent, to generate rules from the contingent and arbitrary data of many past
moments of associative life on the city streets, to refine and edit the code for
future uses in unknown future places. I may feel that some notion of legible
rights is protected, and yet the attributes generated from my data, in correla- 7
tion with yours and others’, continue to supply the conditions for future arbi-

Politics and Ethics

trary actions against unknown others. I draw a distinction here between eth-
ics as code, or what Michel Foucault describes as “the code that determines
which acts are permitted or forbidden,” and ethics as the inescapably political
formation of the relation of oneself to oneself and to others.16 My argument
is that there is a need for a certain kind of ethical practice in relation to algo-
rithms, one that does not merely locate the permissions and prohibitions of
their use. This different kind of ethical practice begins from the algorithm as
always already an ethicopolitical entity by virtue of being immanently formed
through the relational attributes of selves and others. My desire for a different
mode of critique and ethical account is animated not by the question, How
ought the algorithm be arranged for a good society?, but by the question, How
are algorithmic arrangements generating ideas of goodness, transgression, and
what society ought to be?
In this book I propose a different way of thinking about the ethicopolitics
of algorithms. What I call a cloud ethics is concerned with the political forma-
tion of relations to oneself and to others that is taking place, increasingly, in
and through algorithms. My use of the term cloud here is not confined to the
redefined sovereignties and technologies of a “cloud computing era,” as un-
derstood by Benjamin Bratton and others, but refers to the apparatus through
which cloud data and algorithms gather in new and emergent forms.17 The
cloud in my cloud ethics is thus closer to that envisaged by John Durham Pe-
ters, for whom clouds are media in the sense that they are “containers of pos-
sibility that anchor our existence and make what we are doing possible.”18 To
consider algorithms as having ethics in formation is to work with the propen-
sities and possibilities that algorithms embody, pushing the potentials of their
arrangements beyond the decisive moment of the output.
A cloud ethics acknowledges that algorithms contain, within their spa-
tial arrangements, multiple potentials for cruelties, surprises, violences, joys,
distillations of racism and prejudice, injustices, probabilities, discrimination,
and chance. Indeed, many of the features that some would like to excise from
 the algorithm — bias, assumptions, weights — are routes into opening up their
politics. Algorithms come to act in the world precisely in and through the rela-
tions of selves to selves, and selves to others, as these relations are manifest in
the clusters and attributes of data. To learn from relations of selves and others,
the algorithm must already be replete with values, thresholds, assumptions,
probability weightings, and bias. In a real sense, an algorithm must necessarily
discriminate to have any traction in the world. The very essence of algorithms
8 is that they afford greater degrees of recognition and value to some features
of a scene than they do to others. In so doing, algorithms generate themselves
Introduction

as ethicopolitical beings in the world. If to have ethics is not merely to have a

code prohibiting, for example, bias or assumptions, but to work on oneself via
relations, then the ethicopolitics of algorithms involves investigations of how
they learn to recognize and to act, how they extract assumptions from data
relations, and how they learn what ought to be from relations with other hu-
mans and algorithms.
To be clear, the cloud ethics I propose here does not belong to an episteme
of accountability, transparency, and legibility, but on the contrary begins with
the opacity, partiality, and illegibility of all forms of giving an account, human
and algorithmic. To advance a cloud ethics is to engage the ungrounded poli-
tics of all forms of ethical relations. The significant new ethical challenges that
algorithms seem to present to society actually manifest novel features of some
profoundly old problems of the grounds for ethical action. As Judith Butler ex-
plains in her Spinoza lectures, the demand to give an account of oneself will
always fall short, for “I cannot give an account of myself without accounting
for the conditions under which I emerge.”19 If one assumes that the determi-
nation of an unequivocal I who acts is a necessary precondition of ethics, as
Butler cautions, then this identifiable self is “dispossessed” by the condition of
its emergence in relation to others. For Butler, this persistent failure to give a
clear-­sighted account does not mark the limit point of ethics. On the contrary,
the opaque and unknowable nature of making all kinds of acting subjects is
the condition of possibility of having an ethicopolitical life.20
In short, and in contrast to the equation of ethics with transparency and
disclosure, ethical responsibility is sustained by conditions of partiality and
opacity. My notion of a cloud ethics extends the opacity of the human sub-
ject, envisaging a plurality of venues for ethical responsibility in which all
selves — human and algorithmic — proceed from their illegibility. The apparent
opacity and illegibility of the algorithm should not pose an entirely new prob-
lem for human ethics, for the difficulty of locating clear-­sighted action was al-
ready present. The I who forms the ethical relation was always in question and
is now, with algorithms, in question in new ways. Though the mathematical
propositions of algorithms cannot be made fully legible, or rendered account-
able, they can be called to give accounts of the conditions of their emergence.
These conditions include some relations that are identifiably between humans
and algorithms — such as the selection and labeling of training data, the set-
ting of target outputs, or the editing of code “in the wild,” for example — but
others still are relations of algorithms to another algorithm, such as a classi-
fier supplying the training data from which a neural network will learn. In all 9
such instances of iterative learning, the significant point is that the conditions

Politics and Ethics

of an algorithm’s emergence — a composite of human-­algorithm relations — are
venues for ethicopolitics.
In a discussion on the impossibility of the transparent algorithm, the bril-
liant and generous scholar of black studies and machine learning Ramon Am-
aro once said, “Well what would it be if we even could open it? It’s just math.”21
Of course, he intended the comment as a provocation, but mathematics is
never only “just math,” as Amaro’s work vividly shows. To reflect on the con-
ditions of an algorithm’s emergence is also to consider how, as mathematical
knowledge forms, algorithms have achieved the status of objective certainty
and definiteness in an uncertain world.22 Ludwig Wittgenstein observed math-
ematical propositions to be “given the stamp of incontestability,” a mark of the
“incontrovertible” and an “exemption from doubt” that other propositions,
such as “I am called,” are not afforded.23 For Wittgenstein, mathematics as
practice — like all other language games — is concerned with particular uses of
propositions, where “what a proposition is, is in one sense determined by the
rules of sentence formation, and in another sense by the use of the sign in
the language game.”24 His concern is that the mathematical proposition has
achieved a particular status of certainty in an otherwise uncertain world, so
that it becomes “a hinge on which your dispute can turn.”25 For Wittgenstein,
the mathematical proposition should be regarded as no less doubtful or uncer-
tain than the “empirical propositions” otherwise made about the world. In-
deed, Wittgenstein’s point is to address mathematical propositions as empirical
actions that are “in no way different from the actions of the rest of our lives,
and are in the same degree liable to forgetfulness, oversight and illusion.”26
Following Wittgenstein, the use of mathematical propositions is profoundly
social and, in my reading, ethicopolitical. An algorithm is formulated through
a series of truth claims about its match to the world, and yet, in its use in the
world it is as prone to forgetfulness, oversight, misrecognition, and illusion as
any other language game.
Algorithms such as those used to detect latent social unrest in the city
 may appear in the world as Wittgenstein’s “hinge on which your protest can
turn” in the most direct sense that the hinge delimits and circumscribes the
arc of the politics of protest. But the algorithm as hinge does not merely mark
the limit point of resistance; rather, it presents something as a singular opti-
mal output, when it is actually generated through multiple and contingent
relations. My cloud ethics considers the algorithmic hinge to be akin to Karen
Barad’s scientific apparatus, which decides what matters in the world, what
10 or who can be recognized, what can be protested, and which claims can be
brought.27 Understood in this way, the algorithm is not the hinge as an incon-
Introduction

trovertible axis, exempted from doubt, on which all social, political, and eco-
nomic life turns. “The hinge point,” as Foucault differently identifies, can also
be the point of “ethical concerns and political struggle,” as well as the point of
“critical thought against abusive techniques of government.”28
So, a principal ethicopolitical problem lies in the algorithm’s promise to
render all agonistic political difficulty as tractable and resolvable. Where poli-
tics expresses the fallibility of the world and the irresolvability of all claims,
the algorithm expresses optimized outcomes and the resolvability of the claim
in the reduction to a single output. In the following pages, I specify how this
book approaches what an algorithm is and what it does. Among the many
problems of studying algorithms is the matter of specifying which type of al-
gorithm one is addressing. Though my primary focus in this book is on ma-
chine learning algorithms, and predominantly deep neural networks, in most
of the instances I discuss, these algorithms are used in conjunction with some
much older and less fashionable rules-­based and decision tree algorithms. The
form of the algorithm is not delimited by its name but by its coming into be-
ing, its use in the wild. As Nick Seaver has argued, “Rather than offering a
‘correct’ definition,” a critical study of algorithms could begin from “their em-
pirical profusion and practical existence in the wild.”29 For example, an ad-
vanced deep neural network for object recognition is intimately connected to
some much older classifiers that are used in the preparation of the data, and
it meaningfully comes into being as it is modified through its deployment in
the world. It is not possible to identify a secure boundary around one specific
named algorithm because, as a calculative device, it is a composite creature.
I propose three routes into understanding what algorithms are in the world,
each with its own distinctive implications for what is at stake ethicopolitically:
algorithms as arrangements of propositions; algorithms as aperture instruments; al-
gorithms as giving accounts of themselves.
“Generate All the Rules and Debate Which Is Interesting”:
Algorithms as Arrangements of Propositions

The spatiality of the logic of algorithms is most commonly figured as a series

of programmable steps, a sequence, or a “recipe,” governed by “precise instruc-
tions” within a “finite procedure.”30 The calculative logic is thus most widely
represented as a logic of the series, where each step in a calculative procedure
is defined by its position in a finite series.31 This notion of a recipe, or a series
of steps, contributes to limiting the imagination of what the ethicopolitics of 11
algorithms could be. The spatial imaginary of the algorithm as a series of steps

Politics and Ethics

nurtures a particular set of ideas about how to intervene in the series to change
the outcomes.32 So, for example, if the negative outcome of a credit-­checking
algorithm was found to include a racial or gendered bias in one of the steps,
then the removal of this element in the recipe could be considered significant
in notions of accountable and responsible algorithmic decisions. Similarly, to
envisage a “kill switch” in the algorithms of autonomous weapons systems, for
example, is to imagine a sequence of steps in which a human in the loop could
meaningfully intervene and prevent a lethal decision.33 Where algorithms are
represented as a sequence, or what Manuel DeLanda calls “mechanical recipes
specified step by step,” the addition or deletion of a step destroys the outcome,
halts the decision in its tracks.34 In this sense the spatial imagination of the
algorithm as series appears to make possible all kinds of human oversight of
otherwise automated decisions.
The representation of algorithms as a logical series, however, seriously
overlooks the extent to which algorithms modify themselves in and through
their nonlinear iterative relations to input data. The machine learning al-
gorithms that are so categorically redefining our lives are characterized less
by the series of steps in a calculation than by the relations among functions.
Within computer science these relations are understood to be recursive func-
tions, whereby the output of one calculation becomes the defining input for
another, and so on, with each function nested within others like an infinite
nesting of Russian dolls. These recursive functions, as Paulo Totaro and Do-
menico Ninno have argued, are having specific and durable effects on contem-
porary society.35 Significantly, the removal or deletion of one function does
not destroy the overall arrangement. Indeed, intrinsic to the logic of machine
learning algorithms is their capacity to learn which outputs from which of
their layers to pay greater attention to and which to bypass or discard.36 This
matters greatly, because the removal of a step one assumed to contain sensitive
data on race, for example, would not remove or delete the process of learning
via proxies how to recognize by means of racialized attributes. Consider, for
 example, the machine learning algorithms that are being used in sentencing
decisions by the courts to anticipate the optimal outcome of a prison sentence
versus noncustodial measures.37 Such technologies do not deploy a sequential
logic that could be amenable to oversight and intervention in the steps. On the
contrary, the optimal future outcome is defined solely in relation to an array of
recursive functions, a “different mode of knowing,” as Adrian Mackenzie ex-
plains, in which the input data on past convictions and sentencing outcomes
12 of hundreds of thousands of unknown others supply the contingent probabili-
ties to all the layers within the algorithm.38
Introduction

Algorithms are not merely finite series of procedures of computation but

are also generative agents conditioned by their exposure to the features of data
inputs. As Luciana Parisi has argued, this is “a new kind of model,” which
“derives its rules from contingencies and open-­ended solutions.”39 When al-
gorithms learn by inductively generating outputs that are contingent on their
input data, they are engaging experimentally with the world. As computer sci-
entist Rakesh Agrawal explains the shift from rules-­based to generative learn-
ing algorithms, past forms “used a statistical notion of what was interesting” so
that “the prevailing mode of decision making was that somebody would make
a hypothesis, test if it was correct, and repeat the process.” With machine
learning algorithms, such as recursive neural networks, Agrawal notes that
“the decision process changed,” and algorithms would “generate all rules, and
then debate which of them was interesting.”40 In their contemporary form, al-
gorithms generate output signals that open onto uncertainty as to what is in-
teresting, useful, or optimal. These output signals are not mere mathematical
abstractions but are actionable propositions, such as “this person poses high
risk of overstaying their visa,” or “this object is threatening the security of the
street.” Often, when algorithm designers describe how they work with clients
on a particular application, they suggest that they “tune” the algorithm as part
of a discussion with the client of what is useful or optimal. This experimental
tuning enacts the process Agrawal describes as debating which of the outputs
is interesting, where the observation of the output of the model modifies and
adjusts the weightings and thresholds of the algorithm. A kind of science of
emergent properties, as I describe in my book The Politics of Possibility, such
techniques significantly transcend and undercut traditional statistical notions
of what matters, what is interesting, and what is optimal.41
In this book I understand the spatial logic of algorithms to be an arrange-
ment of propositions that significantly generates what matters in the world. In
contrast to the spatiality of the series or recipe, the arrangement of proposi-
tions articulates the algorithm’s capacity to engage experimentally with the
world, to dwell comfortably with contingent events and uncertainties, and yet
always to be able to propose, or output, an optimal action. Practically, in the
research for this book, I have studied the algorithm not as a finite series of pro-
grammable steps but as perennially adjustable and modifiable in relation to a
target output. What does it mean to understand algorithms as arrangements
of propositions? In Alan Turing’s famous paper “Systems of Logic Based on Or-
dinals,” he reflected on what he called “the exercise of ingenuity in mathemat-
ics.” For Turing, ingenuity was important for mathematical reasoning because 13
it provided “arrangements of propositions,” which meant that the intuitive

Politics and Ethics

mathematical steps could not “seriously be doubted.”42 The arrangement of
propositions was thus a kind of mathematical architecture that supported the
intuitive and the inferential faculties. As I use the notion, an arrangement of
propositions extends to the experimental and iterative capacities of algorithms
to propose things in and about the world. This is most likely not an interpreta-
tion Turing would approve of. Indeed, when he attended Wittgenstein’s 1939
Cambridge lectures, their manifest disagreements in the lecture theater con-
cerned precisely the matter of whether mathematical propositions could have
normative effects. When Turing asserted that “from the mathematical the-
ory one can make predictions,” Wittgenstein replied that “pure mathematics
makes no predictions.” For Wittgenstein, “30 × 30 = 900 is not a proposition
about 30” but rather a proposition that finds its expression only in the “gram-
mar” of its arrangement.43 While for Turing, the numeric value 30 itself has a
kind of agency, for Wittgenstein this is afforded only by its arrangement in a
wider grammar through which it comes into use.
In a sense, the disagreement between Turing and Wittgenstein is un-
dercut by a twenty-­fi rst-­century world in which algorithms arguably gener-
ate their grammars and propositions through their exposure to numbers in
the form of input data. As historian of mathematics Keith Devlin reminds us,
mathematics involves not only numeric values based on a “count” but also,
crucially, transformations based on “processes you perform.”44 The arrange-
ments of propositions I envisage are not only numeric but transformative and
performative. They contain within them multiple combinatorial possibilities
and connections. The multiplicity of the algorithm — its plural possible com-
binations, pathways, weights, and connections — matters greatly to my tracing
of the empirical processes by which algorithms learn and reach decisions, and
to my desire to shift the ethicopolitical terrain on which this is understood to
take place. The arrangement of propositions means that an apparently opti-
mal output emerges from the differential weighting of alternative pathways
through the layers of an algorithm. In this way, the output of the algorithm is
 never simply either true or false but is more precisely an effect of the partial
relations among entities. As Isabelle Stengers has noted of the proposition, “It
is crucial to emphasize that the proposition in itself cannot be said to be true
or false” because in itself “it is indeterminate with regard to the way it will be
entertained.”45 As a proposition, the algorithm can similarly not be said to be
true or false — it cannot be held to account for its relation to truth in this sense.
A pattern of false positives from a biometric algorithm, for example, can never
14 be simply false because the threshold is immanently adjustable. Understood in
Stengers’s sense of the proposition, algorithms are “indeterminate with regard
Introduction

to the way [they] will be entertained.”

Consider, for example, the multiple arrangements of propositions as they
are manifest in the Asimov Institute’s Neural Network Zoo. Described as an
“almost complete chart of neural networks,” the neural net zoo displays a spa-
tial mapping of the arrangements of propositions of machine learning algo-
rithms. This is not a zoo that categorizes its flora and fauna by the characteris-
tics of their genus and species. Instead, it displays the algorithms’ architectures
as arrangements of proximities, distances, intensities, and associations. The
Asimov researchers, in their depiction of the arrangement of a convolutional
neural network (cnn) — commonly used for facial recognition, feature extrac-
tion, and image classification — explain how each node concerns itself only
with its close neighboring cells. The nature of the function performed within
the node is decided by the close communion of weighted probabilities in the
neighboring cells. As an arrangement of propositions, one could not meaning-
fully open or scrutinize the 60 million probability weightings that make it
possible for a cnn algorithm to recognize the attributes of a face in a crowd,
declaring them to be true or false. Indeed, following Stengers’s formulation
of the proposition, the output of a facial recognition algorithm is never ei-
ther “true” or “false” but instead is a useful proposition that can be infinitely
recombined. Unlike a series of steps or a recipe, one could never have over-
sight of the infinite combinatorial possibilities of the algorithm as proposi-
tion. Nor would the deletion of a step render the whole unworkable. Once
the algorithm is understood as an arrangement of propositions, the mode of
ethics must work with the partiality and illegibility of the relations among
entities.
“Reduced to That Which Interests You”:
Algorithms as Aperture Instruments

Critical accounts of the rise of algorithms have placed great emphasis on the
power of algorithms to visualize, to reprogram vision, or indeed even to “see”
that which is not otherwise available to human regimes of visuality. Similar to
the spatial arrangement, this primacy of the visual register has also annexed
what could count as the ethics and politics of algorithms. There are two cu-
riously twinned accounts of contemporary algorithms in relation to regimes 15
of sight and vision. The first is that algorithms operate on a plane in excess of

Politics and Ethics

human visibility and at scales that are inscrutable to the human. The second
is that algorithms themselves have an enhanced capacity to visualize the in-
visible, to see, scan, and search volumes and varieties of data heretofore un-
available to human senses. Indeed, this intersection of machinic and human
vision comes to the fore in the espoused ethics of public inquiries into the
state’s deployment of automated algorithms for the government of the popu-
lation. For example, in the UK parliamentary inquiry following the Edward
Snowden disclosures of widespread automated data analysis, a peculiar kind
of virtue was found in the notion that, in machine learning intelligence, “only
a tiny fraction of those collected are ever seen by human eyes.”46 Similarly in
the United States, the former director of national intelligence James Clapper
likened the nsa’s algorithmic analysis of citizens’ data to a form of library
in which few books are “actually read” and where the output of the system
supplies “the books that we need to open up and actually read.”47 There is an
acute problem, then, with the widespread appeal to ethical codes that regulate
what algorithms or humans do or do not see. It is a problem, I suggest, with its
roots in the privileging of sight and vision over other forms of making things
perceptible. “Vision cannot be taken,” writes Orit Halpern in her wonderful
book Beautiful Data, “as an isolated form of perception” but must be considered
“inseparable from other senses.”48 To act and to be responsible for action, an
algorithm need not “see” or “read” but need only make something or someone
perceptible and available to the senses.
In this book I situate the ethics and politics of algorithms within a geneal-
ogy of technologies of perception. Contemporary algorithms are changing the
processes by which people and things are rendered perceptible and brought
to attention. This is definitively not merely a matter of making things ame-
nable to vision and indeed is frequently a matter of sustaining something be-
neath the visual register and yet perceptible. As art historian Jonathan Crary
writes, “Ideas about perception and attention were transformed” alongside the
historical “emergence of new technological forms of spectacle, display, pro-
 jection, attention, and recording.”49 Understood in this way, the transforma-
tion of perception involves changes in how the perceiving subject thinks about
what could be brought to attention, changes in the horizon of possibility of
human action. As with the advent of the technologies of printing press, cam-
era, or cinema, so the advent of the machine learning algorithm implies a re-
working of what it means to perceive and mediate things in the world.50 This
is not a process that is effectively captured by the idea that automated systems
16 are undermining or superseding human forms of perception and action. To
foreground instruments of perception, or what Henri Bergson terms “organs
Introduction

of perception,” is to breach conventional distinctions between humans and

machines and acknowledge the entangled nature of all forms of perception.51
Bergson insists on the shared limits of perception across science and ordinary
everyday experience, so that “ordinary knowledge is forced, like scientific
knowledge,” to divide up time into perceptible slices, to “take things in a time
broken up into particles.”52 Whether the organ of perception is microscope,
telescope, eye, camera, or algorithm, perception is attuned to action, to the di-
viding up of movement into points on a trajectory so that they can be acted on.
“What you have to explain,” he writes, is not “how perception arises, but how
it is limited, since it should be the image of the whole, and is in fact reduced to
the image of that which interests you.”53 Following Bergson’s insight on how
an organ of perception seizes the object of interest from its environment, to
consider algorithms as instruments of perception is to appreciate the processes
of feature extraction, reduction, and condensation through which algorithms
generate what is of interest in the data environment.
Confronted by something of a moral panic surrounding the expansive vol-
umes of “big data” and powers of surveillance of automated systems, my em-
phasis on practices of perception foregrounds precisely the opposing process:
the reducing, distilling, and condensing of particles of interest from a whole.
A defining ethical problem of the algorithm concerns not primarily the power
to see, to collect, or to survey a vast data landscape, but the power to perceive
and distill something for action. Algorithms function with something like an
aperture — a n opening that is simultaneously a narrowing, a closure, and an
opening onto a scene. Let us consider, for example, an algorithm designer
working in the UK defense sector, demonstrating the capacity of his deep neu-
ral network algorithms to recognize a mistaken civilian target amid a crowded
data environment of drone images.54 He shows a slice through time as a vehicle
travels away from the center of Kandahar, Afghanistan. He explains the prob-
lem for the decision: that this could be a suspect vehicle or, crucially, a school
bus taking children home to villages outside the city. The algorithms had
learned to recognize a school bus through training data that supplied images
of predominantly yellow US-­style buses. The designer explains that his algo-
rithm is aggregated with many others to generate a single actionable output —
 target/no target — but to do this he must necessarily reduce and condense the
patterns of interest from a volume of input data. The training data — and the
humans who labeled it — have conditioned the cnn algorithm to carve out and
value some objects and to discard others. In this part of Afghanistan, some 17
of the school buses are indeed in the spatial form of a US-­style school bus,

Politics and Ethics

but others still are open-­back trucks repurposed for transporting scholars. My
point is that a potential act of violence, such as a school bus wrongfully tar-
geted by a drone strike, resides not primarily in the vertical surveilling and
collecting of data, but in fact in the horizontal arraying of possible patterns of
interest lodged within the algorithm itself.55 As an aperture instrument, the
algorithm’s orientation to action has discarded much of the material to which
it has been exposed. At the point of the aperture, the vast multiplicity of video
data is narrowed to produce a single output on the object. Within this data
material resides the capacity for the algorithm to recognize, or to fail to recog-
nize, something or someone as a target of interest.
The ethical stakes of what Mark Hansen calls “potential perceptual recon-
figuration” applied to my cloud ethics necessarily involves something like a re-
opening of the process of the algorithm’s reduction of a multiplicity to one.56
What is happening in this process of condensing plural possible pathways to a
single output? When an algorithm determines whether a vehicle is a military
or a civilian target, or when it decides if a public protest contains latent dan-
gerous propensities, it reduces the heterogeneity of durational time to perceive
the attributes of an object and their differences of degree from other objects
encountered in a past set of data. The question of what this crowd could be,
what this vehicle might do, the frustrations or discomforts of the actual lived
experience of waiting or gathering persist as indeterminacies in the hidden lay-
ers of the algorithm. Even within the archive of training data — sometimes just
a Google ImageNet dataset of labeled images of school buses — are the residual
contingencies of durational time, with all the past lived moments supplying
norms and anomalies for the algorithm to learn. “My own duration, such as I
live it in the impatience of waiting,” reflects Gilles Deleuze, “serves to reveal
other durations that beat to other rhythms, that differ in kind from mine.”57
To respond to the perceptual power of the algorithm and to prize open the
aperture of the single output is to trace the other durations that continue to
 beat in the discarded data, the multiple other potential pathways that could be
mapped between fragments.

“At the Limits of What One Knows”:

Algorithms Give Accounts of Themselves

Perhaps the most widespread concern in public and scholarly discourse on the
operation of algorithms in society is that they are unaccountable or that they
18 cannot meaningfully be held to account for harmful actions. Indeed, a kind of
proxy form of accountability is emerging, in which the designers of algorithms
are made the locus of responsibility for the onward life of their algorithms.
Introduction

Virginia Eubanks proposes a “Hippocratic oath for data science,” in which the
designers of algorithms would be accountable to human beings and not “data
points, probabilities, or patterns.”58 Similarly, in a 2016 Nature editorial call-
ing for “more accountability for big data algorithms,” the editors propose that
“greater transparency” could be achieved if the designers of algorithms “made
public the source of the data sets they use to train and feed them.”59 They urge
greater disclosure of the design of algorithms and an “opening up to scrutiny”
of their workings. Indeed, so widespread is this notion that the accountability
of algorithms can be grounded in their design or source code that technical so-
lutions for “explainability” are being developed to trace an apparent bias back
to a design problem. For example, automated systems for the assessment of
creditworthiness are thought to be rendered transparent by a tool that traces
the specific credit score output back to a data element, such as an unpaid bill.60
Such techniques are thought to anchor the accountability of the algorithm
precisely in an intelligible knowledge of its workings.
Within these demands for algorithmic accountability lies a specific form of
giving an account. The locus of a truthful account is in the apparent “source”
of the algorithm, in its origins, whether in the source code or in the algorithm
designer as an author. This locus of original account, I suggest, is profoundly
limiting the capacity to demand that algorithms give accounts of themselves.
It imagines a secret source or origin to which all potential future harms could
be traced. As Foucault writes on the disappearance of the author, “the task
of criticism is not to bring out the work’s relationship with the author” but
rather to “analyse the work through its structure, its architecture, its intrinsic
form, and the play of its internal relationships.”61 For our contemporary times,
the call for accountability of algorithms has precisely targeted the work’s re-
lationship with the author, seeking to render transparent the intent and the
workings of the design. At one level there are clear limits to identifying the
source of the algorithm, not least that each apparent “one” contains multiple
elements from multiple sources, with much of this aggregation concealed even
from the designer. My point, however, is that the problem of anchoring ac-
countability in a source or origin is not confined to algorithms but is a persis-
tent and irresolvable ethicopolitical problem.
The problem of the unidentifiable origin of the algorithm extends to all no-
tions of an authoring subject I who can give a clear-­sighted account of herself.
If one’s account of oneself can never be fully secured, then the full disclosure
of one’s grounds for action is impossible. To give an account is always to give an 19
uncertain narrative that risks falling short or failing to be recognized. Let us

Politics and Ethics

begin not from a search for secure grounds for accountability, then, but from
the very ungroundedness of all forms of giving an account. “Ethics and politics
only come into being,” Thomas Keenan writes, “because we have no grounds,
no reliable standpoints” from which to forge foundations.62 To be responsi-
ble for something — a n errant output, a fatal decision, a wrong judgment — 
is less a matter of securing the grounds for the action than a matter of respond-
ing even when knowledge is uncertain and the path is unclear. “What could
responsibility mean,” asks Keenan, if it is “nothing but the application of a rule
or decision.”63 There is no great origin or source of responsibility without un-
certainty and undecidability.
In this book I propose that algorithms are not unaccountable as such. At
least, they should not be understood within a frame of ethical codes of ac-
countability in which the source of the problem could be secured. Algorithms,
I propose instead, are giving accounts of themselves all the time. These ac-
counts are partial, contingent, oblique, incomplete, and ungrounded, but, as
N. Katherine Hayles vividly documents, this is not a condition unique to the
cognitive complexities of algorithms.64 Far from it. The condition of giving an
account that is never transparent or clear sighted is already the ethicopolitical
condition with which we must live. To attend to the accounts that algorithms
give is to “stay with the trouble” of “unexpected collaborations and combina-
tions,” as Donna Haraway has captured the “method of tracing, of following a
thread in the dark.”65 Given the conditions of following collaborative threads
in the dark, perhaps one should not create a special category of opaque and il-
legible agency for the identification of algorithms. If, as Judith Butler suggests,
“my account of myself is partial, haunted by that for which I have no defini-
tive story,” then might it be that “the question of ethics emerges at the limits of
our schemes of intelligibility, where one is at the limits of what one knows and
still under the demand to offer and receive recognition?”66 This is the ethics at
the limits of schemes of intelligibility, of following threads in the dark, that I
envisage for my cloud ethics. Refuting the many demands for an impartiality
 of the algorithm, excised of bias and prejudice, I wish to be alert to the always
already partial accounts being given by algorithms.
As feminist scholars of technoscience have long reminded us, the partial
account is not an account devoid of insight or real purchase on the world.
“There is no unmediated photograph or passive camera obscura in scientific
accounts of bodies and machines,” writes Donna Haraway. “There are only
highly specific possibilities, each with a wonderfully detailed, active, partial
20 way of organizing worlds.”67 To attend to algorithms as generating active, par-
tial ways of organizing worlds is to substantially challenge notions of their
Introduction

neutral, impartial objectivity. To foreground partiality is also to acknowl-

edge the novel forms of distributed authorship that newly entangles the I who
speaks in composite collaborations of human and algorithm. If one element
of my past presence on a London street for a “Stop the War” campaign march
enters a training dataset for a multinational software company’s neural net,
which, one day in the future, intervenes to detain some other person in a dis-
tant city, how is some part of my action lodged within this vast and distrib-
uted authorship? What is the possibility of my ethicopolitical responsibility
for the dark thread to a future intervention made partially on the attributes
of my past data?
To be attentive to the accounts algorithms are giving of themselves, then,
is to begin with the intractably partial and ungrounded accounts of humans
and algorithms. As is so manifestly present with the surgical robotics algo-
rithms I discuss in chapter 2, the embodied accounts human surgeons give
of what they can do, how they decide on boundaries between diseased and
healthy tissue, how they can reach decisions, are not meaningfully separable
from the surgical robots with whom they share a cognitive workload. The ro-
bot can only act to suture a wound because its algorithms have learned from
the past data of many thousands of instances of human surgeons suturing.
The surgeon can only reach a difficult kidney tumor because the robot’s data
yield precise coordinates from an mri that make it recognizable amid occlu-
sions. When an error is made, the identification of a unified and identifiable
source of the error is not possible. The figure who would be required to give a
clear-­sighted account is an impossible figure.

Chapter Outlines

In sum, my cloud ethics has a principal concern with the algorithm’s double
political foreclosure: the condensing of multiple potentials to a single output
that appears as a resolution of political duress; and the actual preemptive clo-
sure of political claims based on data attributes that seek recognizability in
advance. Confronted with this double foreclosure, each chapter of this book
elaborates a speculative strategy for reinstating the partial, contingent, and in-
complete character of all algorithmic forms of calculation. In their partial and
incomplete way of generating worlds, we can locate their ethicopolitics. Part 1,
“Condensation,” comprises two chapters that detail how algorithms condense
and reduce the teeming multiplicity of the world to a precise output. Chapter 1,
“The Cloud Chambers,” examines how algorithms are acting through cloud
data architectures to produce a new paradigmatic alliance between sovereign 21
authority and scientific knowledge. Developing an analogy with the scientific

Politics and Ethics

experiments of the cloud chamber of early twentieth-­century particle physics,
where the chamber made it possible to perceive otherwise invisible subatomic
particles, I explore the capacities of the cloud in cloud computing. How does
the cloud apparatus render things perceptible in the world? I address the char-
acter of cloud architectures across two distinct paradigms. The first, Cloud
I, or a spatiality of cloud forms, is concerned with the territorial identification
and location of data centers where the cloud is thought to materialize. Here
the cloud is understood within a particular history of observation, one where
the apparently abstract and obscure world can be brought into vision and ren-
dered intelligible.
This notion of cloud forms, I propose, has led to a distinct ethicopolitical
emphasis on rendering algorithms explainable or legible. Cloud I is founded on
a misunderstanding of the nature of algorithmic reason within the cloud, so
that the cloud is thought to obscure or obfuscate what is “really” going on in
the world. On the contrary, I propose, the cloud is not an obfuscation at all but
is a means of arranging the models for otherwise incalculable processes for a
condensed decision. In my second variant — Cloud II, or the spatiality of a cloud
analytic — the cloud is a bundle of experimental algorithmic techniques acting
on the threshold of perceptibility. Like the cloud chamber of the twentieth
century, contemporary cloud computing is concerned with condensing that
which cannot be seen, rendering things perceptible and actionable.
Chapter 2, “The Learning Machines,” turns to the question of machine
learning algorithms and their complex and intimate relationships with what
we think of as human practices. Through a study of the deep neural network
algorithms animating surgical robots, I show the entangled composite bodies
of surgeons, robots, images of organs, and cloud-­based medical data of past
surgeries — all of which learn together how to recognize and to act amid uncer-
tainty. Where there has been public concern and moral panic around machine
learning, what is most commonly thought to be at stake is the degree of auton-
omy afforded to machines versus humans as a locus of decision. I propose that
 the principal problem resides not with machines breaching the limit but in sus-
taining a limit point of the autonomous human subject — the oft-­cited “human
in the loop” — who is the locus of decision, agency, control, and ethics. Such
an autonomous human disavows the we implicated in the entangled learning
of humans with algorithms. Much of this entangled learning takes place in a
space of play between a target output and an actual output of the algorithm.
The indeterminate combinations of weights, parameters, and layers in the al-
22 gorithm are the traces of rejected pathways and alternative correlations.
Part 2, “Attribution,” elaborates the practices of attribution through
Introduction

which algorithms write themselves into being in the world. Taking seriously
the attribute as it is understood in computer science, the chapters connect
this attribution back to genealogies of writing and ethics. In chapter 3, “The
Uncertain Author,” I am concerned with how the search for ethical codes to
govern algorithms has located an author function in the “source code” of al-
gorithms. When an algorithm appears to have precipitated a crisis or to have
caused a harm, often the reflex response is to seek out its origin: Who designed
the model? Who wrote the code? Who or what labeled the training data? Was
it biased? Who is the author of the algorithm? In this chapter I explain the lim-
its of locating an ethicopolitical response in the authorship of source code. Fo-
cusing on the algorithmic techniques for natural language processing — where
a corpus of literary texts are used to train an algorithm to recognize style and
sensibility — I suggest that the algorithm’s ways of being in the world are not
all present in the source code and, indeed, substantially exceed the design of
an authoring subject. The authorship of the algorithm is multiple, continually
edited, modified, and rewritten through the algorithm’s engagement with the
world. Juxtaposing novelists’ accounts of their own uncertain authorship with
computer scientists’ reflections on how the fragments of an algorithm come
together in acts of writing, I suggest the impossibility of identifying a defini-
tive author.
To invoke the call for attributing authorship is not only insufficient as cri-
tique, but it also risks amplifying the ability of the algorithm to bind together
a unity of incompatible, fraught elements as though the difficulties and dif-
ferences could be resolved. In place of the call for securing responsibility via
authorship, I propose that there is a possibility of ethicopolitics in the act of
writing, in the bringing together of scattered elements, and in the opening of
a space of uncertainty. A cloud ethics must begin to reopen the act of writ-
ing as a site of ethical significance. The profound uncertainty that is brought
to the writing of an algorithm — “I do not know how adjusting this weight is
changing the output”; “I cannot be sure how my training data has produced
these clusters”; “I am curious whether shifting this threshold will reduce the
false positives” — invites an iterative process of writing that is never completed.
Here is an ethicopolitical tension that is worth holding on to — the algorithm
promises to complete everything, to condense to a single optimized output
and action, and yet it enacts a process of writing that opens on to an indeter-
minate future.
Chapter 4, “The Madness of Algorithms,” addresses the moments when it
appears that an algorithm has acted in a state of frenzy or has departed from 23
its otherwise rational logic. From the lethal accidents of autonomous vehi-

Politics and Ethics

cles during tests, to the racist hate and misogyny of the Twitter chatbot Tay,
an ethical frame often seems required to somehow rein in the worst excesses
or to restore reasonableness to autonomous actions. Yet, philosophy has long
grappled with the problem of madness and, specifically, how the identification
of madness shores up and sustains the domain of reason. Reflecting on the
conjoined histories of ideas of reason and madness, I propose that one cannot
speak of the madness of the algorithm except in relation to the very form of
reason the algorithm embodies. While the contemporary moral panic at each
moment of the madness of algorithms urges us to police ever more vigilantly
the line between reasonable and unreasonable actions, understood as a thresh-
old, this line is precisely the condition of possibility of algorithmic rationality.
Algorithms cannot be controlled via a limit point at the threshold of madness
because the essence of their logic is to generate that threshold, to adapt and to
modulate it over time. In short, my argument is that when algorithms appear
to cross a threshold into madness, they do, in fact, exhibit significant qualities
of their mode of reasoning. Understood in this way, the appearance of a mo-
ment of madness is a valuable instant for an ethicopolitics of algorithms be-
cause this is a moment when algorithms give accounts of themselves.
Contra the notion that transparency and the opening of the black box se-
cure the good behavior of algorithms, the opacity and clouded action exhib-
ited in the excesses and frenzies of algorithms have a different kind of fidel-
ity to the account. Throughout this book I argue that, when viewed from the
specific propositional arrangements of the algorithm, particular actions that
might appear as errors or aberrations are in fact integral to the algorithm’s
form of being and intrinsic to its experimental and generative capacities. I
am advocating that we think of algorithms as capable of generating unspeak-
able things precisely because they are geared to profit from uncertainty, or
to output something that had not been spoken or anticipated. Of course,
this is not a less troubling situation than the one in which some controls are
sought on the worst excesses of the algorithm. On the contrary, it is all the
 more political, and all the more difficult, because that which could never be
controlled — change, wagers, impulses, inference, intuition — becomes integral
to the mode of reasoning.
Part 3, “Ethics,” develops a set of tactical routes for a cloud ethics to follow,
each of these insisting on a different kind of weightiness from the calculative
weights of adjustable probabilities. The apparent lightness of an optimized sin-
gle output is afforded the full weight of undecidability and the difficulty of de-
24 cision. In chapter 5, “The Doubtful Algorithm,” I suggest that doubt, and more
precisely the idea of doubtfulness, can be a tactical point of intervention for a
Introduction

cloud ethics. The overwhelming logic of algorithmic systems in our society is

that they can optimize decisions made where there is profound doubt and un-
certainty. Indeed, the algorithm is offered as a means of deploying doubt pro-
ductively so that, for example, the doubt-­ridden voter or doubtful consumer
can be clustered as “having a propensity to be influenced” and can be targeted
with personalized media. A specific form of truth telling, established in the
“ground truth” of the data environment from which the algorithm learns, has
asserted its dominance in the governing of societies with algorithms.
By contrast, and cutting against the grain of the dominance of definiteness
as algorithms act on doubt, I seek to reinstate doubtfulness as what N. Kath-
erine Hayles calls “embodied actuality” within the calculative architecture
of the algorithm.68 Though at the point of optimized output, the algorithm
places action beyond doubt, there are multiple branching points, weights,
and parameters in the arrangements of decision trees and random forest algo-
rithms, branching points at which doubt flourishes and proliferates. A cloud
ethics reopens the contingencies of this multiplicity, giving life to the fallible
things that the algorithm has learned about the world, rendering the output
indelibly incomplete.
In each of the chapters, I maintain a faithfulness to the specificities of
how particular algorithms learn via, and generate worlds through, their rela-
tions with data. I am wary of “algorithm talk” when it is asserted generally
and without specificity, for different algorithms are as varied in their logics
and grammars as languages are, and these differences, as Adrian Mackenzie
argues, should be made to matter.69 Thus, for example, in discussions of fa-
cial recognition systems, I would want to know what kinds of cnns are being
used as the basis for recognizing a face from data inputs. Though each of the
chapters draws out the arrangements of propositions of particular algorithms
and how they recognize, misrecognize, and target through their relations with
other algorithms, data, and humans, I have also foregrounded the fallibility of
the algorithm, its incompleteness and contingency. In short, a specific algo-
rithm will always exceed its name, its type, its genus, for it is immanently mod-
ifying itself through the world. In chapter 6, “The Unattributable,” I address
more directly those critics who would ask what use my cloud ethics would be
to their campaigns or movements against the injustices of specific algorithms.
Following a lecture on machine learning that I gave in Copenhagen, a law-
yer in the audience asked what my cloud ethics could ever look like in law. I
had similar questions from lawyers at the Turing Institute in London. My re-
sponse, in short and in those moments, was to say that it would be a crowded 25
court in the sense that my approach multiplies the possible sites of interven-

Politics and Ethics

tion and responsibility. This notion of the crowded space, or forum, has stayed
with me throughout the research and the writing of this book. What kind of
political claim could be brought in the name of a cloud ethics? How does one
forge a form of responsibility for the future onward life of something like an
attribute as it becomes attached to others? What is the power of the unattrib-
utable as the set of qualities that cannot be attributed to a subject?
Chapter 6 maps out what it means to be together ethicopolitically — to be
associated as a society, a community, a movement, a gathering of protesters — 
as an association of partial associations or attributes. The chapter concludes
the book along three lines of argument for a cloud ethics: apertures, opacity,
and the unattributable. When machine learning algorithms segment a social
scene, generating clusters of data with similar propensities, everything must
be attributed. Yet, that which is unattributable does remain within the scene,
exceeding the algorithm’s capacity to show and tell, as well as opening onto
a different kind of community and a different mode of being together, being
ethicopolitical.
This page intentionally left blank
 Part 1

Condensation
This page intentionally left blank
1
The Cloud Chambers
Condensed Data and Correlative Reason

What I wouldn’t give now for a map of the ever constant ineffable?
To possess, as it were, an atlas of clouds.
 — David Mitchell, Cloud Atlas

A Beautiful Sight

In physicist Charles Thomson Rees Wilson’s Nobel lecture of 1927, he de-

scribed the cloud chamber experiments he had conducted since the late nine-
teenth century, and how they had transformed the capacities of modern phys-
ics into the twentieth century. From the observatory at the summit of Ben
Nevis, Wilson had witnessed what he depicted as “the wonderful optical phe-
nomena” of the formation of clouds.1 Inspired by what he had observed, Wil-
son spoke of cloud formations that “greatly excited” his interest so that he
“wished to imitate them in the laboratory.”2 For Wilson, the capacity to re-
produce in science the formation of clouds in nature became the means to
advance understanding of the condensation physics of meteorology, and with
it the taxonomy and classification of cloud forms. In Wilson’s laboratory, the
scientific practice of knowing clouds followed a path from observation to mi-
metic representation to classification.
When Wilson began to experiment with the formation of clouds in his
cloud chamber apparatus, however, what he discovered was an unanticipated
potential to see something not otherwise perceptible; phenomena that exceed
the paradigms of observation and classification. In contrast to the telescopes
of the observatory, where the optic instruments had brought objects into a
line of human sight, Wilson’s cloud chamber became a different kind of ap-
 paratus, one that brought something into perceptibility that could not other-
wise be seen. Though ionizing particles, such as alpha, beta, and gamma radia-
tion, could not be observed directly, the condensation trail in Wilson’s cloud
chamber showed the particle’s trajectory of movement. Recalling his experi-
ments with supersaturation levels, temperature, and the expansion of gas in
his chambers, Wilson reflects in his lecture, “I came across something which
promised to be of more interest than the optical phenomena which I had in-
30 tended to study.” His cloud chamber experiments afforded him a “means of
making visible and counting certain individual molecules or atoms which
Chapter One

were in some exceptional condition.” Though Wilson had set out to reproduce
mimetically the formation of clouds in nature, in fact his experiments gener-
ated traces of the electrically charged atoms and ions previously unavailable
to the sciences.3
What Wilson’s cloud chamber ultimately made possible for the twentieth
century’s study of particle physics was the ability to photograph and to per-
ceive the movement of particles in an exceptional state (figures 1.1 and 1.2). As
historian of science Peter Galison writes in his account of Wilson’s work, “Af-
ter the cloud chamber the subatomic world was suddenly rendered visualiz-
able.”4 Charged or ionized particles could not be observed directly with optic
devices, as with the instruments of microscopy or telescopy, but their traces
and trajectories of motion appeared indirectly via the cloud tracks condens-
ing on the nuclei. As Wilson reflects on his 1911 experiments, “I was delighted
to see the cloud chamber filled with little wisps and threads of clouds” so that
“the very beautiful sight of the clouds condensed along the tracks of the alpha
particles was seen for the first time.”5 The cloud chamber apparatus, conceived
for the human observation of processes of formation in nature, had become a
technique for rendering perceptible the movement of objects beyond the thresh-
olds of human observation.
Almost exactly one century on from the publication of Wilson’s first cloud
chamber images, the idea of the cloud is once more describing the advent of
processes at scales that appear to transcend the observational paradigm and
exceed our human capacities to see, to know, and to understand. Indeed, the
cloud in cloud computing is widely held to derive from the mapping of infra-
structures of computer networks, where the visualization of a figurative cloud
stands in for the vast complexity of the internet.6 In the twenty-­first century,
cloud computing promises to have effects on the very episteme of our world,
analogous to the effects of the cloud chamber on what could be rendered
knowable in twentieth-­century physics.
More precisely, the advent of cloud computing opens space for a renewed
 Figure 1.1 
Charles Thomson
Rees Wilson’s cloud
chamber expansion
apparatus (1912 – 13).
Science Museum
Group Collection,
Board of Trustees of
the Science Museum,
31
London.

The Cloud Chambers

twinning of science and technologies of perception with forms of political sov-
ereignty. Such renewal signals an extension of historical technologies of imag-
ing, mapping, and intelligence data collection into new algorithmic modes of
analysis and data correlation. In February 2015, for example, seventeen US in-
telligence agencies — including the Department of Homeland Security (dhs),
National Security Agency (nsa), Central Intelligence Agency (cia), Depart-
ment of the Navy, Department of the Army, National Geospatial Intelligence
Agency, Defense Intelligence Agency, Office of the Director of National Intel-
ligence (odni), Department of the Air Force, Federal Bureau of Investigation
(fbi), State Department, and Drug Enforcement Agency (dea) — launched
the icite program for the cloud-­based storage, sharing, and analysis of intel-
ligence data. Here, once more, one can find the promise of a “beautiful sight”
celebrated by Wilson, the making of pictures otherwise unavailable to the
senses. icite (pronounced “eyesight”) is the Intelligence Community Infor-
mation Technology Enterprise (figure 1.3), a $600 million cloud-­computing
contract with Amazon Web Services (aws) and Cloudera, providing a new
intelligence and security data infrastructure. icite, it is promised, will “allow
agencies to share data much more easily and avoid the kind of intelligence gaps
that preceded the September 11, 2001, terrorist attacks.”7 In this specific sense,
the data geographies of the cloud can be read as a response to the 9/11 Commis-
sion findings of a failure to analyze across data silos held by different agencies.8
As the former US director of national intelligence James Clapper explained at
the launch of the icite program, cloud computing allows government author- 33
ities to “discover, analyze and share critical information in an era of seemingly

The Cloud Chambers

infinite data.”9 The cia’s chief intelligence officer, Douglas Wolfe, similarly ex-
pressed his hopes that the government security agencies would get “speed and
scale out of the cloud, so that we can maximize automated security.”10 Let us
reflect on the terms once more: discover, analyze, infinite data, speed and scale out
of the cloud, maximize automated security. The cloud promises to transform not
only what kinds of data can be stored, where, and by whom, but most signifi-
cantly, what can be generated and analyzed in the world. In short, the cloud
form of computation is not merely supplying the spatial accommodation of
large volumes of data in server farms, but offers the means to map and to make
perceptible the geography of our world in particular ways.
The architecture of the cloud is defined spatially by the relations between
algorithms and data. As the cloud becomes ever more closely intertwined with
geopolitical decisions — from sharing and acting on intelligence data, to border
controls, urban policing, immigration decisions, and drone strikes — then what
is the precise nature of these algorithmic practices of data gathering, analyz-
ing, and knowing? In this chapter I address the political character of cloud
computing across two distinct paradigms. The first, which I term Cloud I, or
cloud forms, is concerned with the spatiality of data stored in data centers and
analyzed in cloud architectures. This Cloud I paradigm sustains an ontology
of observation, representation, and classification, and it encloses ethicopolitics
within this logic. In the second mode, Cloud II, or a cloud analytic, I propose
that the computational regimes of the cloud transform what or who can be
rendered perceptible and calculable. In contrast with Cloud I’s linear logics of
observation, representation, and classification, the generative logics of Cloud
II are engaged in perception, recognition, and attribution. As the scientific his-

Figure 1.2  Cloud chamber tracks of the alpha rays

of radium, among the earliest of C. T. R. Wilson’s
cloud chamber images. Proceedings of the Royal
Society, London.
34
Chapter One

Figure 1.3  US Intelligence Community Information Technology Enterprise

(ICITE ) architecture and its imagination of a “world of possibilities.” Screenshot
archived in 2014, when the Amazon Web Services contract to build ICITE was
first announced.

tory of the cloud chamber is concerned with “the character of an instrument

and the effects produced with it,” as Svetlana Alpers has put it, I am interested
here in understanding the character of the instruments of cloud computing
and their generative effects.11

Cloud I: “Geography Matters in the Cloud”

Cloud I is concerned with the identification and spatial location of the data
centers where the cloud is thought to materialize. Indeed, as computer science
began to document the emergence of cloud computing, the idea of geography
came to have a specific meaning defined by where data and programs are spa-
tially located. In a 2008 Association of Computing Machines (acm) forum
devoted to the advent of cloud computing, a transformation is described “in
the geography of computation,” with “data and programs” being “liberated” as
they are “swept up from desktop pcs and corporate servers and installed in the
compute cloud.”12 Such accounts of the cloud appeal to a geography of “scal-
able” computation, which is thought to change radically with the rise of big
data and the concomitant need for flexible storage and computational power.13
There is need for some caution, however, in understanding the geography of
the cloud primarily in relation to the rise of twenty-­first-­century volumes of
digital data. Indeed, the emergence of cloud computing has important origins
in grid computing, distributed scientific data, and, perhaps most significantly,
the notion of computing as a public utility. “Computing may someday be or-
ganized as a public utility,” speculated computer scientist John McCarthy in
his mit lecture of 1961, so that “the computer utility could become a new and 35
important industry.”14

The Cloud Chambers

In the second half of the twentieth century, the imagination of computing
as a scalable public utility emerged. By 2006, when aws launched its Elastic
Compute Cloud (ec2), the architecture of cloud computing had begun to de-
velop the three components now most widely recognizable as the cloud: Infra-
structure as a service, in which hardware, servers, storage, cooling, and energy
are supplied; platform as a service, in which the software stack is accessed via
the cloud; and the applications layer, in which data analytics capacity — or the
deployment of algorithms — is supplied via the cloud. Across the components
of cloud architectures, the emphasis is on scalable computing, where the client
pays for what they have used, combined with distributed computing, where
multiple concurrent users can share and combine their data, their algorithms,
and their analyses.
Understood as an architecture where data and algorithms meet and en-
tangle, cloud computing’s notion of a single computer hosting multiple simu-
lated or virtual machines becomes actualized in space in a particular way.15 In
this respect, the whereabouts of simulated machines is not unknown at all, but
rather the cloud is actualized in data centers, located in places within econo-
mies of land, tax rates, energy, water for cooling, and proximity to the main
trunks of the network. As Benjamin Bratton writes, the “cloud is not virtual;
it is physical even if it is not always on the ground, even when it is deep under-
ground.”16 Hence, within the vocabulary of computer scientists at least, “geog-
raphy” is said to “matter in the cloud.”17 In this limited sense of geographies of
the cloud, the architecture of data centers has come to count as a spatial and
geographic phenomenon in a way that the spatialities of the algorithm have
not. When computer scientists ask “where is the cloud,” what they denote as
“geographical questions” concern the data centers thought to “underlie the
clouds,” their “physical location, resources, and jurisdiction.”18 When, for ex-
ample, Google locates a new data center in the tax-­friendly state of Georgia, or
the Swedish internet service provider Bahnhof installs a data center in the cool
confines of a former nuclear bunker under Stockholm, or Sun Microsystems
36
Chapter One

Figure 1.4  The Google-patented “water-­based data center,” hosting a float-

ing platform of servers, powered and cooled by seawater. US Patent Office
US752520782.

designs a portable data center inside a standard shipping container, the matter
of geography is thought to reside in the spatial location of data storage.19 Of
course, these materializations of the cloud have ethicopolitical significance,
so that when Google patented the floating data center (figure 1.4), the offshore
potentials of cloud techniques beyond jurisdiction were actualized in these
vast data ships.
Let us capture this geography of the spaces of cloud computing as Cloud
I. This imagination of the cloud as a dispersed yet spatially located array of
data centers is present in computer science, and that has extended into geo-
graphic and even political and geopolitical debate. So, for example, following
the Snowden disclosures of the extent of US authorities’ access to European
citizens’ data via US data centers, the EU sought initially to develop a “Euro-
pean cloud,” in which they might store safely European data under European
jurisdiction.20 Similarly, following the US subpoena and mining of European
financial transactions, the major Society for Worldwide Interbank Financial
Telecommunications moved its cloud-­computing provision to an underground
data center in Switzerland, and the Canadian government has legislated for
what it calls “data sovereignty,” meaning domestic public data traffic must
not leave Canadian territory. Understood as a spatial and territorial question
of the jurisdiction of data technologies, then, the political response has been
to seek to render the cloud intelligible and governable.
 The unintelligibility of the cloud as a form of governing or a locus of geo-
political power has profoundly political implications. Following the exposure
of the nsa’s prism program in 2013, for example, the UK Intelligence and Se-
curity Committee (isc) of Parliament — the sole body responsible for public
oversight of security and intelligence powers in the UK — called the then for-
eign secretary, Philip Hammond, to give evidence before the committee. At
the time of his evidence, Hammond was the final signatory on all warrants
authorizing the interception and analysis of external communications data, 37
conventionally understood as where one end of the communication is located

The Cloud Chambers

externally to the UK. In his testimony, this figure of final sovereign authority
manifestly failed to understand the complex spatial modalities of data stored,
transferred, or analyzed in the cloud:

Q: The distinction between internal and external to the UK is important

because there are tighter restrictions on analysing internal data. . . .
But if the sender and recipient of an email are both in the UK, will it
be treated as internal even if the data is routed overseas on its journey?
A: So, I think . . . er . . . and I invite my colleagues to step in if I get this
technically wrong. . . . But I think . . . er . . . it’s an internal communi-
cation. [At this point, the civil servants flanking the minister lean in
to advise.]
Let me finish my train of thought. . . . [M]y understanding is,
er, because of the technical nature of the internet . . . it is possible it
could be routed to servers outside the UK. . . . Please correct me if
I misinterpreted that. . . . I’m sorry, I have misled you in my use of
terms. . . . I’m trying to be helpful.
Q: Well[,] you will be relieved to know that was the easy one. Now,
the case of social media . . . if all of my restricted group of Facebook
friends are in the UK . . . and I post something to Facebook, surely
that should be internal?
A: [Following whispers from civil servants] Erm, . . . no[,] actually[,] if
you put something on Facebook and the server is outside of the UK
it will be treated as an external communication.
Q: What about cloud storage, where no other person is involved at all. It
may be my decision to upload photographs to Dropbox. Would these
communications be regarded as external because they are on
US servers?
A: Aaah . . . er. My colleagues will . . . oh . . . well. . . . Yes[,] I am advised
if the server is overseas they will be regarded as external.21
 The 2014 UK foreign secretary’s testimony before the isc exposes the dif-
ficulties and limit points of a territorialized juridical form in the face of cloud
computing. In Cloud I, where the geography of cloud forms is everything, the
cloud has become centered on where data is collected and stored. Indeed, the
Anglo-­A merican juridical tradition has founded its privacy protections pre-
cisely on the consent required for lawful storage and collection. Hammond’s
evidence before the committee, however, exemplifies how it may be the very
38 unintelligibility of cloud forms that enables and facilitates contemporary sur-
veillance and intelligence gathering. As Kate Crawford and Jason Schultz ar-
Chapter One

gue, the new predictive “approaches to policing and intelligence may be both
qualitatively and quantitatively different from surveillance approaches,” and
thus enable “discriminatory practices that circumvent current regulations.”22
Yet, Crawford and Schultz go on to suggest that an alternative space for demo-
cratic oversight might lie in what they call “a right to procedural data due pro-
cess,” where constraints and oversight mechanisms are placed on the algorith-
mic processes of data analysis.23 Even as the cloud overflows and exceeds the
categories and practices of bureaucracy and law, what has come to be at stake
ethicopolitically has become a struggle to wrest the cloud back into a form
over which one can have oversight, to expose its “bias” and demand neutrality,
to make it comprehensible and accountable in democratic fora, and to render
the cloud bureaucratically and juridically intelligible.
Among the critical geographic and artistic accounts of cloud computing,
the desire to wrest the cloud into an intelligible form similarly finds expression
in methods of visualization. The geographer and artist Trevor Paglen seeks to
“make the invisible visible,” reflecting that “the cloud is a metaphor that ob-
fuscates and obscures” the material geographies of the “surveillance state.”24
Paglen’s work is concerned with bringing the geopolitics of cloud computing
back into a human line of sight through visualization. His methods deploy op-
tical devices of many kinds to bring back into human vision that which would
otherwise exceed the limits of observation. His ghostly images of the nsa’s
data centers are photographs taken at night with a long-­focus lens from a he-
licopter (figure 1.5); and his photographs of the secret installations of military
and drone bases in the Nevada desert are taken with adapted telescopic instru-
ments of astronomy.25
The optical instruments deployed by Paglen belong to a paradigm of ob-
servation in which, as Peter Galison describes, one is offered “a direct view”
of things otherwise “subvisible.”26 As Paglen accounts for his own work: “My
intention is to expand the visual vocabulary we use to see the US intelligence
community. Although the organizing logic of our nation’s surveillance appa-
 39

The Cloud Chambers

Figure 1.5  NSA headquarters, Fort Meade, Maryland. Trevor Paglen, 2014.

ratus is invisibility and secrecy, its operations occupy the physical world. Digi-
tal surveillance programs require concrete data centers; intelligence agencies
are based in real buildings. . . . [I]f we look in the right places at the right times,
we can begin to glimpse the vast intelligence infrastructure.”27 So, for Paglen
the challenge is to “expand the visual vocabulary” to see more clearly the ac-
tions of algorithms, and to bring into vision the things that would otherwise
be obfuscated by the cloud. Similarly, for the artist James Bridle, “one way of
interrogating the cloud is to look where its shadow falls,” to “investigate the
sites of data centres” and to “see what they tell us about the real disposition
of power.”28
Yet, what are the “right places” and “right times” to look and to observe
an apparatus? How does one investigate something describable as the real dis-
position of power? Indeed, what ways of seeing would be appropriate to what
art historian Jonathan Crary calls a “relocation of vision” taking place with
computation, or the visual vocabulary appropriate to the digital mediation
of cultural objects and urban scenes identified by Gillian Rose and Shannon
Mattern?29 If the cloud is to be observed and revealed in the secret glimmer-
ing buildings of the nsa’s data centers in Paglen’s images, then could his “real
buildings” also be located in other places? Could they be observed, for exam-
 ple, in the rented North London offices where a small team of physics gradu-
ates write algorithms for risk-­based security?30 Must the material geography of
cloud computing be found in the buildings or territories where it is thought to
actualize? Could the “right place” to look also be in the spatial arrangements
of an experimental clustering algorithm used in anomaly detection, or in the
generative logics of emergent machine learning algorithms?31
My point is that the desire to “open the black box” of cloud computing
40 and to expand the visual vocabulary of the cloud, to envision the cloud and
its properties in geographic space, dwells within and alongside the paradigm
Chapter One

of observation. In Stephen Graham’s work on cities and warfare, for example,

he writes of “systems of technological vision” in which “computer code tracks
and identifies.”32 Critical scholarly responses to the cloud — through imaginar-
ies of verticality and the stack — have thus emphasized overwhelmingly spaces
of sight and oversight, where observational power is foregrounded. Crucial as-
pects of these technologies, however, do not operate on the terrain of human
vision and, indeed, harness perceptual power on a horizontal threshold of con-
nection and correlation.
And so, Cloud I poses questions within an observational mode: Where is
it?; What type is it?; Can we map it?; Can we recognize it? As with the early
classification of cloud forms, when Luke Howard first proposed names for cir-
rus, cumulus, and stratus in 1803, a linear system of genera and species was
proposed to “enable people to think coherently about clouds” (figure 1.6).33 The
system of cloud form classification was later described as “quite ridiculous for
clouds,” because they are not fixed forms but ever in formation, and indeed an-
alog algorithms were devised to diagram the observational pathways of cloud
formation. In short, Cloud I sustains the idea that one can have a more beau-
tiful sight, a means of seeing more clearly and rendering something coherent
and intelligible. The telescope and camera Paglen brings to the scene of data
deployment belongs to a particular history of observation, one that Donna Ha-
raway describes as “visualizing technologies without apparent limit,” like “the
god trick of seeing everything from nowhere.”34
What might it mean for us to commit instead not to enable coherent
thinking about the cloud? If one determines instead to “stay with the trouble,”
as Haraway has put it, of partial and indeterminate lines of sight, then all
apparently coherent technologies of observation become what Haraway calls
“active perceptual systems” with “partial ways of organizing worlds.”35 In the
second variant I discuss here — Cloud II, drawing on Peter Galison’s distinction
between mimetic and analytical forms of scientific instruments — cloud com-
puting appears as a cloud analytic.36 Here, the cloud is a bundle of experimental
 41

The Cloud Chambers

Figure 1.6  Analog algorithm for the classification of cloud mechanisms, 1967.
Scorer and Wexler, Cloud Studies in Colour.

algorithmic techniques acting in and through data to modify the threshold

of perceptibility itself. Put simply, as analytic the cloud resides not within a
paradigm of observation, representation, and classification but instead within
perception, recognition, and attribution. As Galison reminds us, in the cloud
chamber “we do not actually see things,” though what we perceive “has a close
relation to them,” what he calls an “almost seeing” of otherwise subvisible en-
tities.37 Understood thus, to say the cloud somehow obfuscates a real world
of politics is to miss the point somewhat. The cloud is not an obfuscation;
far from it.38 Like the cloud chamber of the twentieth century, contemporary
cloud computing is about rendering perceptible and actionable (almost seeing)
that which would otherwise be beyond the threshold of human vision. Where
some claim the cloud makes the geographies of power in our world unintelli-
 gible, I propose that it is becoming an important element of what Karen Barad
calls the “condition of intelligibility” of our world.39

Cloud II: “Changing the Aperture of Observation”

What I call Cloud II, or the cloud as analytical gathering of algorithms with
data, displaces the question, Where is the cloud?, replacing it with, How does
the cloud render the world perceptible and analyzable? In this way, Cloud II
42 witnesses the extension of algorithmic modes of reason that surface something
for perception and action amid an otherwise profoundly uncertain environ-
Chapter One

ment. As historian of science Paul Erickson and his colleagues have traced me-
ticulously in the historical emergence of algorithmic rationality, the profound
uncertainties of the Cold War nurtured a desire to extend mathematical logic
into the realm of decision. The decision procedures and cybernetic methods of
algorithm appeared to extend the faculties of human reason so that they “no
longer discriminated among humans, animals, and machines” in the capacity
to analyze, to decide, and to act.40 What we see here is the entwining of hu-
man and machine modes of reasoning such that what Henri Bergson calls the
“organs of perception” of the world are composite beings formed through the
relations among humans, algorithms, data, and other forms of life.41
Understood in terms of the intertwined faculties of human and machine,
the contemporary spaces of cloud computing exceed the territorial geogra-
phies of the location of data centers, becoming instead a novel political space
of calculative reasoning. Returning to the site of the icite program, what
kinds of perceptions and calculations of the world become possible with the
algorithmic instruments that gather in cloud space? When the seventeen US
intelligence agencies upload or analyze data in icite, they access software as
a service, so that they are not merely “joining the dots of their data,” as the
post-­9/11 measures claimed, but are in fact combining their modes of analysis.
Among the available icite platforms is Digital Reasoning software, a set of
machine learning algorithms for analyzing and deriving meaning from data
across government intelligence databases and open source social media data
streams. Describing the Synthesys application available to analysts via icite,
Digital Reasoning claims that its algorithms are able to “read,” “resolve,” and
“reason” from the correlations in cloud data. The algorithms are said to “ex-
tract value from complex and opaque data” and to “determine what’s impor-
tant” among the “people, places, organizations, events and facts being dis-
cussed,” ultimately “figuring out what the final picture means” by comparing
it with “the opportunities and anomalies you are looking for.”42 What this
means in terms of the practices of the algorithms is that they identify clusters
within the data (to read), derive attributes from those clusters to make them
recognizable into the future (to resolve), and compare the output of the algo-
rithms with the target output of the analyst (to reason). This is an iterative and
experimental process in which the humans and machines feel their way toward
solutions and resolutions to otherwise indelibly political situations and events.
Though Digital Reasoning’s machine learning algorithms are now the
mainstay of US government defense analysis, they were developed initially for
anomaly detection in the wholesale financial industry. The frenzy of the 2008 43
financial crisis included trading strategies that were in breach of regulations,

The Cloud Chambers

leading to significant losses for institutions. In the aftermath of the crisis, ma-
chine learning algorithms were developed to analyze “terabytes of emails per
day to detect hints of insider trading.”43 The cognitive computing software
performs the role of what N. Katherine Hayles calls a “cognizer,” carrying out
the cognitive functions to generate norms and anomalies from the patterns in
vast datasets and, as it does so, deciding what or who will come to matter amid
the occlusions.44 At the level of the algorithm, there is profound indifference to
the context of whether these norms and anomalies pertain to financial trades
or the movement of insurgent forces — what matters is precisely the capacity to
generate a final output that can be acted on. By 2015, when Digital Reasoning
was pitching to government security and defense officials in Washington, DC,
the ceo Tim Estes was offering the machine learning tools to “sift through
sensor data, emails and social media chatter,” bringing “structure to human
language” and “changing the aperture of observation.”45
What would it mean for algorithms to change the aperture of observation?
Let me agree, curiously and peculiarly, with this vendor of cloud-­based algo-
rithms to the dhs and the nsa and say, yes, indeed, the aperture of observa-
tion is changing, though not in such a way that the promised complete final
picture is delivered to the analyst. Indeed, the technology of the aperture is
concerned with a specific and discrete opening onto a scene, one in which the
notion of a completeness beyond the aperture is occluded and screened out.
The aperture, from the Latin aperire, to open or to uncover, is not primarily
an optic but is rather a means of opening onto or uncovering the world. The
point at which a machine learning algorithm condenses the output of multiple
layers to a single output is also an aperture in the sense that it is an opening
or uncovering of attributes and relations that would not otherwise be percep-
tible. Put simply, with Cloud II, where we are interested in the analytic, what
matters is not so much seeing the “where” of the data as it is the capacity to
extract patterns and features from data to open onto targets of opportunity,
commercial and governmental.
 With the advent of cloud computing, the aperture of observation becomes
an aperture of “almost seeing,” in Peter Galison’s terms, or a means of “cor-
relating and synthesizing large volumes of disparate data” so that action can
take place based on what is almost seen of the world.46 As one analyst puts the
problem, “It allows us to say correlation is enough. We can stop looking for
models, throw the data at the biggest computing clusters the world has ever
seen and let algorithms find the patterns.”47 So, correlation is enough. Let us re-
44 flect on this claim of an adequacy of correlative associations. In the pages that
follow, I propose three characteristics of correlative cloud reason, and I sug-
Chapter One

gest why this form of algorithmic reason has significance for our ethicopoliti-
cal present.

Condensing Traces
The algorithmic techniques of Cloud II involve condensing traces, practices not
primarily concerned with seeing or bringing into vision, but rather focused on
engaging a subvisible world by inferring from the traces and trajectories that
condense at indeterminate points. Returning to my analogy with the appara-
tus of the cloud chamber, by the mid-­t wentieth century, when Charles Wil-
son’s cloud chamber was being used in subatomic physics, the “purpose” of the
instrument was described as being “to study the motion of ionizing particles
from records of the drops condensed on ions formed along the trajectories fol-
lowed by these particles.”48 The motion of particles could not be observed di-
rectly, but their trajectory could be perceived obliquely, via the visible drops
condensed on the ions — the cloud “tracks.” Figure 1.7 shows one of the best-­
known cloud chamber photographs, C. T. R. Wilson’s image of alpha-­emitting
thorium, with the cloud originating from an alpha ray passing through the
chamber, its “trajectory disturbed in two places,” as recorded in the atlas of
cloud chamber images.49 The newly available images of radioactivity made the
object perceptible via the records of condensed drops on the ions, observing
the motion of the particles obliquely. In the compelling images from the cloud
chamber, one can locate a capacity to perceive the movement of otherwise sub-
visible entities. The chamber of the cloud chamber is akin to an apparatus, in
Michel Foucault’s terms, in that it “inserts the phenomena in question” within
a “series of probable events.”50 In this sense, to condense a trace is not to show
or to reveal something existing in the past, but rather to establish a condensed
series of possible correlations between entities and events. The cloud cham-
ber apparatus is experimental to the extent that it is concerned with probable
tendencies and trajectories, condensing a larger volume down to the probable
event.
Figure 1.7  Cloud chamber tracks of alpha-­emitting thorium. C. T. R. Wilson,
c. 1923, Proceedings of the Cambridge Philosophical Society.
 Like the cloud chamber apparatus of the twentieth century, contemporary
cloud computing is a form of experimental chamber or apparatus. It condenses
and reduces the volume of data down to that which is probable or possible in
the algorithmic model. When Digital Reasoning claims that their algorithms
determine what is important in war or in finance, this is because their ma-
chine learning apparatus is generating what matters in a series of probable
events. As Karen Barad writes on the nature of scientific apparatuses, they are
46 boundary-­making practices that decide “what matters and what is excluded
from mattering,” as they “enact agential cuts that produce determinate bound-
Chapter One

aries and properties of ‘entities’ within phenomena.”51 In condensing the prob-

able data traces of what matters in the world, cloud computing enacts the mat-
ter of the person of interest at the border, the possible future disruptive protest
event in urban space, the acts of fraud or insider trading, or the chains of asso-
ciation of no-­fly lists, blacklists, and kill lists — like beaded drops of condensed
data making action possible.
Though the movement of an entity cannot be observed directly in the
chamber, it is perceived obliquely through tracks and trajectories of mobility.
Indeed, Cloud II as a mode of computational power poses significant questions
for political and geographic accounts of what it might mean to “secure the vol-
ume” or the “volumetric,” or to have a “politics of verticality.”52 The analytical
techniques available in the cloud do not strictly act on the earth from some
novel spatial dimension “above” or “below” the ground, but rather enroll the
space of calculation itself. Cloud computing acts on the vast volume of data
traces through a series of algorithmic apertures that condense and distill the
thing of interest. In contrast with a securing of the volume, the pursuit of secu-
rity through the volume precisely reduces and condenses the volume by means
of the correlations within the data. The so-­called cognitive-­computing appli-
cations in the icite cloud, for example, use pattern recognition and sentiment
analysis to identify political protests, civil unrest, and “atypical” gatherings or
events. Cognitive computing renders perceptible to the analyst “what matters”
in the political scene, using the volume of cloud-­based digital data precisely to
reduce and flatten the field of vision. The relation between volume and flatness
thus becomes one in which the tracks of association and correlation enact the
horizon of possibility for the analyst. The volume is radically condensed down
to the target data elements, like beaded drops on ionizing particles through
which future trajectories of motion can be inferred.
Discovering Patterns
The algorithms gathering in Cloud II involve the discovery of patterns, which is a
highly specific calculative device deployed in a volume of data. The repository
of data in the US intelligence community’s cloud, for example, is described
as a “data lake” in which the “same raw data” can be analyzed with “statisti-
cal functions,” such as conventional regression, and with “machine learning
algorithms” of “data discovery.”53 Here, the relation of the data lake to cloud 47
computing is metaphorically understood as the formation of clouds from the

The Cloud Chambers

water vapor rising from lakes into the atmosphere.54 While the application of
statistical analysis to intelligence data involves the analyst beginning with a
deductive query, or hypothesis, and building rules to test that hypothesis in
the data, the advent of cloud computing presents the analyst with a volume
and variety of data too great for conventional human hypothesis or deduction.
In the context of a security paradigm that seeks out the uncertain possible fu-
ture threat, the volume of data in the lake — much of it transactions and social
media data — is analyzed with machine learning techniques that promise to
yield previously unseen patterns via processes of “knowledge discovery.”
In contrast to a deductive form of reasoning by hypothesis testing, ma-
chine leaning algorithms deploy abductive reasoning, so that what one will
ask of the data is a product of the patterns and clusters derived from that data.
As Luciana Parisi writes on the algorithmic logic of abduction, “algorithms
do not simply govern the procedural logics of computers” but take “genera-
tive forms driven by open-­ended rules” or “derive rules from contingencies.”55
Understood in these terms, the machine learning algorithms deployed in the
contemporary intelligence cloud are generative and experimental; they work
to identify possible links, associations, and inferences. Such abductive forms
deploy a distinct kind of causal reasoning, different from deductive reasoning,
in which “deductions support their conclusions in such a way that the conclu-
sions must be true, given the premises,” and closer to “fallible inferences,” in
which “the possibility of error remains.”56 Put simply, in the cloud analytic of
Cloud II, the algorithms modified through the patterns in data decide, at least
in part, which fallible inferences to surface on the screen of intelligence ana-
lyst, drone pilot, or border guard.
The rise of correlative forms of reasoning in machine learning has seri-
ous implications for the ethicopolitics of algorithms, not least because error,
failure, or fallibility are no longer conceived as problems with a model; rather,
they have become essential to the model’s capacity to recognize abnormali-
ties and generate norms. Abductive models observe the output signal of the
 algorithm and adjust the layers of probability weighting to optimize the fit be-
tween the input data and the output signal. For this reason, programs such as
icite claim to be “uncovering answers to questions the analysts don’t already
know,” where the question is not a hypothesis but a speculative experiment.57
As one of the designers of t-­digest pattern-­detection algorithms explains, in
describing the modification of a model in response to data inputs, “As small
fluctuations occur in the event stream, our model can adjust its view of normal
48 accordingly.”58 What this means is that the data inputs and the algorithm mu-
tually modify to optimize the output. The norms and anomalies of the model
Chapter One

are thus entirely contingent on a whole series of adjustments, including the

adjustment of weights, parameters, and the threshold to tolerate particular er-
ror rates. “You can move the slider for a lower false positive rate if that is what
you want,” I am told by one designer of facial recognition algorithms, “but that
will always increase the false negative rate, so you need to decide what your
tolerance is.”59 As the politics of our world becomes understood as an “event
stream” filtered through infinitely adjustable apertures and thresholds, the ap-
paratus of the cloud deploys algorithms, such as t-­digest, to generate its mallea-
ble view of what is normal in the world from its exposure to the data stream. To
adjust the view of normal, to move a threshold, or to decide a tolerance — these
are the ethicopolitical dispositions of algorithms as they act on the world.
Returning to my analogy with the cloud chamber apparatus, the twentieth-­
century physicists were also engaged in abductive forms of reasoning and pat-
tern detection that exceeded the deductive testing of hypothesis. The cloud
chamber made it possible to detect previously unseen and unknown particles
via the patterns of unusual or abnormal cloud tracks. “The central problem of
the interpretation” of cloud chamber “exploratory photographs,” as described
in the physicists’ guide to cloud chamber technique, was the “recognition of
the particles involved in a particular event.” To interpret the detected pat-
terns of scattering, cascades, and showers, the physicists inferred from the at-
tributes, or the “characteristic features of particle behaviour.”60 They would
not begin with a hypothesis and test it in the chamber, for the uncertainties
and contingencies of particle behavior had become the focus of their inquiry.
The cloud chamber played a crucial role in the identification of hitherto un-
known subatomic particles, rendered detectable through the generation of
surprising new cloud tracks and trajectories. In the text accompanying the fa-
mous Rochester atlas of cloud images, Nobel physicist Patrick Blackett writes:
“The last two decades have seen an increasing use of two experimental meth-
ods, the cloud chamber and the photographic emulsion, by which the tracks of
sub-­atomic particles can be studied. All but one of the now known unstable el-
ementary particles have been discovered by these techniques. . . . This involves
the ability to recognise quickly many different sub-­atomic events. Only when
all known events can be recognised will the hitherto unknown be detected.”61
The atlases of “typical” cloud chamber images definitively did not of-
fer the scientist a taxonomy or classificatory system for identifying particles,
forming the rules of form for unknown future particles. Rather, the images
provide a kind of training dataset, allowing the scientists to become sensitive
to the patterns and clusters of cloud tracks, so that they could recognize the 49
disturbances and fluctuations of a new event. The discoveries of new particle

The Cloud Chambers

behaviors — the first glimpse of the muon or the positron in the cloud chamber,
for example — were not strictly observations of an unknown object, but more
precisely perceptions of something in close relation to it: the patterns involved
in an event. As Peter Galison reminds us, the cloud chamber images “traveled”
and were “widely exchanged, stored and reanalyzed by groups far distant from
the original photographic site.”62 In this sense, the cloud chamber is the site, just as
the cloud analytic is the site in cloud computing, through which the event is rec-
ognized via its patterns, and where the algorithm and the analyst are trained
in the art of recognition.63

Archiving the Future

The recognition of traces in Cloud II involves an archiving of the future, in
which particular future connections are condensed from the volume of the
data stream and rendered calculable. In a sense, the algorithms of Cloud II are
relatively indifferent to the past as a series of data points or significant events.
What matters most to Cloud II is the capacity to generate a particular future
of dispositions, propensities, and attributes. When aws supplies cloud com-
puting to corporations and governments, the applications layer is configured
as an “app store” so that users can select the analytics tools they want, paying
for what they use. As one UK government client of aws explained to me, “We
are cloud first” because “we want platforms; we want to plug in to different
suppliers.”64 This flexible deployment of cloud data was an important element
of aws’s tender for the icite program, with James Clapper, announcing, “We
have made great strides, the applications are in the apps mall, and the data is
in the cloud.”65 In fact, of course, the significance is that the algorithms and
the data dwell together in aws cloud space, opening the possibility for seem-
ingly infinite calculability, or what Hayles calls “infinitely tractable” data.66
The aperture opens onto a world of optimized targets, threats, and opportuni-
ties so that the analyst experiences a sense of reach into possible futures. The
security, policing, or intelligence analyst is thus reimagined by the state to be a
 desiring and wanting consumer, with the “apps mall and stores available from
the desktop,” selling to users “thousands of mission applications” and resem-
bling “what Apple provides through iTunes.”67
Let us reflect for a moment on what the nsa or cia analyst browsing the
icite apps mall might find to assist them in their missions. Among the thou-
sands of applications, Recorded Future offers natural language processing and
sentiment analysis algorithms to “scrape the web” for signals of possible fu-
50 ture threat: “We constantly scan public web sources. From these open sources,
we identify text references to entities and events. Then we detect time peri-
Chapter One

ods: when the events are predicted to occur. You can explore the past, present and
predicted future of almost anything in a matter of seconds. Our analysis tools facili-
tate deep investigation to better understand complex relationships, resulting
in actionable insights.”68 Consider the claim: the data stream of social media
contains all the attributes of incipient future events, so that one can explore
the past, present, and predicted future of almost anything. Recorded Future’s
applications run their algorithms across the boundary of public and private
cloud computing so that the analyst can explore the correlations between, for
example, classified structured data in the dhs’s files and the language and sen-
timent analysis of so-­called open source Twitter feeds and Facebook posts. In
this way, the technology enables action in the present, based on possible corre-
lations between past data archives (such as national security lists) and archives
of the predictive future.
With this instrument of Cloud II, the analytic is everything. Archived
data in Recorded Future becomes unmoored and derivative of its context, even
the so-­called dirty, or noisy, data no longer muddying the calculation but ren-
dered useful. As the security analysts draw together social media “junk” data
with other labeled entities, tagging metadata and sharing with other agencies,
diverse data elements are rendered commensurate and made actionable geo­
politically. As Orit Halpern suggests in her account of how algorithmic ratio-
nality became a governmental and social virtue, digital computation changes
the nature of the archive. The nineteenth-­century form of “static” archiving
and repository is supplemented in the twentieth century by what Halpern calls
“an active site for the execution of operations.”69 Similarly, in the nineteenth-­
century cloud chamber’s attempt to reproduce nature, the cloud tracks had
been considered spurious dirt effects, not for scientific archiving. Yet, it was
the tracks that became the thing of interest, the atlas of cloud chamber images
knowing only the event of the track itself. Contemporary cloud computing is
an active site for the execution of operations, as understood by Halpern, in
which the archive is generative of particular imagined futures.
 The archivization of specific data elements with applications such as Re-
corded Future, then, produces particular futures; “the archivization produces
as much as it records the event,” as Jacques Derrida writes.70 As the photo-
graphic recording of the cloud tracks within the cloud chamber archived the
possibility of recognizing future subatomic particles, so the digital recording
of social media data in cloud computing archives the possibility for future ac-
tions at the border or on the city street. Understood in this way, the spatial
power of the data center as “archive” could be critically challenged, as we see 51
in Trevor Paglen’s images, while leaving entirely intact Halpern’s “active site of

The Cloud Chambers

operations,” a site capable of acting indifferent to the “where” and the “what”
of data.
And yet, there are creative practices of resistance within Cloud II that
offer an alternative sensing of the archive as an active and generative site. In
James Bridle’s installation Five Eyes (figures 1.8 and 1.9) — commissioned by the
Victoria and Albert Museum (v&a) for their public archive – focused exhibi-
tion All of This Belongs to You — the artist invites the viewer to consider anew the
relations between the archive and potential futures. Bridle passed the v&a’s
1.4 million digital object records through an intelligence analysis system. The
algorithms “extract names, things and places, and create searchable connec-
tions between seemingly disparate objects,” with the resulting connections
“difficult to grasp, often inscrutable to the human eye, reflecting the mechan-
ical calculus that was used to generate them.” 71 Displayed in a series of five
glass cabinets, juxtaposed with the woven artifacts of the v&a’s tapestry gal-
leries, the objects surfaced for our attention by the algorithms are placed atop
a “stack” of analog museum files. The object displayed is thus generated in and
through the archive, through the intimate connections learned by the algo-
rithms. One cannot meaningfully trace how the object displayed has become
the thing of interest, or why it is the optimized output from the actions of the
algorithms. In Bridle’s rendering of the archive, one can sense the “upheaval
in archival technology” noted by Jacques Derrida, the infrastructure making a
claim on the future, on the infrastructure of the “archivable event.”72

The Cloud Atlas

When a group of particle physicists showed me their cloud chamber experi-

ments, I had expected the thing of interest around which we would gather
would be the cloud tracks — the wispy trajectories of particles that had so
captivated Charles Wilson. Instead we gathered around the apparatus, the
physicists animated by much discussion on the optimal point of cooling, and
whether thorium is a useful radioactive element for the experiment. One of
Figures 1.8 and 1.9  James Bridle’s Five Eyes installation, displayed in the
Victoria and Albert Museum’s tapestry galleries. Author photographs.
the group had worked at cern with the large hadron collider, commenting
that “there is no reason why we couldn’t have discovered the Higgs Boson us-
ing a cloud chamber, but it would take an inordinately long time.”73 So, for the
scientists, something exists with potential to be detected — manifest in the al-
pha tracks and cosmics in the chamber — but this potential entity is rendered
perceptible by the specific and contingent arrangement of an experimental ap-
paratus. With the smallest of adjustments to the arrangement of the chamber,
the physicists dramatically changed the conditions for what could be traced of 53
the particle’s trajectory.

The Cloud Chambers

The gathering of scientists and one curious geographer around the ex-
perimental apparatus of the cloud chamber resonated with other kinds of
gatherings and disputes I have observed around an experimental apparatus.
When the designers of algorithms for the detection of credit card fraud gath-
ered around their experimental model, they were concerned to optimize the
perceptibility of an anomaly. “The problem with rules,” they explain, is that
“novel fraud patterns will not be detected.”74 To detect something novel, they
suggested, the rules of an algorithm must be generated through the iterations
of output signals to input data. Their machine learning algorithms for detect-
ing emergent forms of fraud, then, are not experimental in the sense of not yet
validated, but are specifically “experimental” in their capacity to adjust pa-
rameters and bring something novel into existence. Isabelle Stengers describes
the “paradoxical mode of existence” of subatomic particles in that they are
simultaneously “constructed by physics” and yet “exceed the time frame of
human knowledge,” so “the neutrino exists simultaneously and inseparably
‘in itself ’ and ‘for us,’ a participant in countless events.” 75 In the lines that
follow here, I conclude the chapter by commenting on why this matters for
our contemporary moment, when the specific apparatus of cloud computing
brings something into being, discovering associations and relations otherwise
unknowable.
Amid the many contemporary calls to bring algorithms into vision, or to
establish human oversight, it is imperative that we formulate accounts of the
ethicopolitics of algorithms that do not replay the observational paradigm of
Cloud I or the classificatory forms related to that paradigm. In Cloud I, vi-
sion is the sovereign sense, afforded both the apparent objectivity of the “most
reliable of senses” and the means of securing the state’s claim to sovereign
violence.76 And yet, the machine learning algorithms operating in programs
such as icite work to render perceptible that which could never be observed
directly, that which could not be brought into view as with an optical device.
The algorithms for identifying insider trading, for credit scoring, or for gener-
 ating intelligence data from social media bring trajectories and thresholds into
being. They are generative and experimental techniques capable of perceiving
a thing of interest without seeing it as such. The algorithms available on state
analysts’ version of iTunes — with their promises of digital reasoning and re-
corded futures — generate a target of interest through subvisible experimenta-
tion. The Cloud II apparatus is not geared to a deductive science of observation,
representation, and classification, but instead signals a paradigm of perception,
54 recognition, and attribution.
The algorithms identifying the attributes of people and things, then, do
Chapter One

not merely observe or record historical data on past behaviors or events. If

they did observe and classify behaviors, then perhaps it might be an adequate
ethical response to insist on the transparency of the processes. With machine
learning programs such as icite, a person or entity of interest emerges from
the correlations and patterns of condensed data: financial transactions, travel
patterns, known associates, social media images, and affects derived from sen-
timent analysis. The data archive of the neural net algorithms of Cloud II is,
as Orit Halpern proposes, “indeterminate in terms of the past,” so that it is not
possible to identify how the present calculation was arrived at.77 Little pieces of
past patterns enter a training dataset and teach the algorithm new things; the
designer of the algorithm experiments with thresholds to optimize the output;
new people and things enter a validation dataset to further refine the algo-
rithm; and on and on iteratively, recursively making future worlds. And let us
not forget that with this correlative reasoning, sovereign decisions are made: to
stop this person at the border, to detain this group as they travel on the subway
to a downtown protest, to target this vehicle as it approaches a checkpoint, or
to approve or deny this asylum claim.
Isabelle Stengers has written that scientific experiments work through
“the power to confer on things the power of conferring on the experimenter
the power to speak in their name.” 78 Understood in this way, whether an ex-
periment can be said to have worked or to yield proof is of lesser significance
than whether it confers the power to speak or to make a claim. In one sense,
Wilson’s cloud chamber experiments failed in that they did not primarily ad-
vance the understanding of the formation and classification of cloud forms. A
greater power, though, was afforded to the particle physicists who were able to
speak in the name of things that would otherwise exceed their observation.
As algorithms written for casino or credit card fraud travel to border control
or to security threat analysis, I propose that cloud computing similarly con-
fers on algorithms the power to confer on the analyst the power to speak in
their name. What do they say when they speak? What kinds of claims do they
make?: here are the people and things with a link to terrorism; here are the
possible fraudulent asylum claims; here are the optimal targets for the next
drone strike; here are the civil uprisings that will threaten the state next week.
The claims that are spoken in cloud computing programs, such as icite, con-
front our fallible, intractable, fraught political world with a curious kind of in-
fallibility. In the cloud, the promise is that everything can be rendered tracta-
ble, all political difficulty and uncertainty nonetheless actionable. The icite
app store marketplace available on the screens of analysts renders the politics 55
of our world infinitely reworkable — the “geopolitical events” in the correlative

The Cloud Chambers

calculus, a kind of geopolitical cloud chamber. As Timothy Cavendish, a pro-
tagonist in David Mitchell’s novel Cloud Atlas, muses, “What I wouldn’t give
now for a map of the ever constant ineffable? To possess, as it were, an atlas of
clouds.”79 Programs such as icite make just such a dangerous promise in the
algorithmic governing of society — a kind of atlas of clouds for the ineffable, a
condensed trace of the trajectories of our future lives with one another.
2
The Learning Machines
Neural Networks and Regimes
of Recognition

The peculiarity of men and animals is that they have the power of adjusting
themselves to almost all the features [of their environment]. The feature to which
adjustment is made on a particular occasion is the one the man is attending
to and he attends to what he is interested in. His interests are determined by
his appetites, desires, drives, instincts — all the things that together make up
his “springs of action.” If we want to construct a machine which will vary its
attention to things in its environment so that it will sometimes adjust itself to one
and sometimes to another, it would seem to be necessary to equip the machine
with something corresponding to a set of appetites.
 — Richard Braithwaite, “Can Automatic Calculating Machines
Be Said to Think?”

Springs of Action

In January 1952 the bbc recorded what was to be the world’s first public de-
bate on the mathematics and ethics of machine learning. Participating in the
discussion of the question “Can Automatic Calculating Machines Be Said to
Think?” were Manchester mathematicians Alan Turing and Max Newman;
their neurosurgeon colleague Geoffrey Jefferson; and Cambridge moral phi-
losopher Richard Braithwaite. Turing explains that he has “made some ex-
periments in teaching a machine to do some simple operation,” but that “the
machine learnt so slowly that it needed a great deal of teaching.” Jefferson is
skeptical of the use of the verb “to learn” in relation to the machine, and he in-
terjects to ask Turing a question: “But who was learning, you or the machine?”1
This distinction, between humans and machines as the locus of learning is
of great significance to the neurosurgeon, for whom the electronic circuits
of computing machines were not analogous to the “fragments of the nervous
system” he encountered in the fleshy materiality of the human brain.2 Yet, Jef-
ferson’s question prompts Turing to reflect on his own embodied experience
of experimenting with his machines. In the audio recording of the debate,
Turing can be heard to pause for a moment’s reflection before responding: “I
am inclined to believe that when one has taught it [the machine] to do certain 57
things[,] one will find that some other things one had planned to teach it are

The Learning Machines

happening without any special teaching being required.” In response to Jeffer-
son’s question of who was learning, the mathematician or the machine, Turing
responds, “I suppose we both were.”3
The entangled “we both” of mathematician and machine, learning to-
gether, expresses Turing’s belief that intuition is a mathematical faculty that
“consists of making spontaneous judgements which are not the result of con-
scious trains of reasoning.” For Turing, the intuitive faculty is entangled
with what he calls the “ingenuity” of the building of rules as arrangements of
propositions.4 The iterative relationship between intuition and ingenuity in
mathematical reasoning necessarily entangles the mathematician’s affective
and haptic relations to a puzzle with the making of a formal axiom or logic.
The human and machinic elements of mathematical learning, then, are not
so readily disaggregated for Turing. Though for the philosopher in the discus-
sion, Richard Braithwaite, it is the unique “peculiarity of men and animals”
that they are able to learn intuitively by “adjusting themselves to almost all
the features of their environment,” his notion of a “spring of action” afforded
by appetites nonetheless calls to mind today’s capacities for machine learn-
ing algorithms to learn and to generate things in excess of their taught rules.5
The 1950s radio discussion of the character of machine learning did, in some
respects, envisage a future world in which machines would exceed the rules-­
based decision procedure and extend to the affective pull of intuitions and ap-
petites for data.
Thus, even in the mid-­t wentieth century, the mathematics and philoso-
phy of machine learning was centered on the entangled relations of humans
and machines. The question, as articulated in the debate on automatic calcu-
lating machines, was “Who was learning, you or the machine?,” and Turing’s
reply was “We both were.” In this chapter, I focus on the we invoked by Tur-
ing in this public debate precisely because it runs against the grain of con-
temporary moral panics amid machine autonomy and algorithmic decisions
that appear to be beyond the control of the human. On the contrary, the we of
 machine learning is a composite figure in which humans learn collaboratively
with algorithms, and algorithms with other algorithms, so that no meaning-
ful outside to the algorithm, no meaningfully unified locus of control, can be
found. In contemporary machine learning, humans are lodged within algo-
rithms, and algorithms within humans, so that the ethicopolitical questions
are concerned less with asserting human control over algorithms and more
with how features are extracted and recognized from a teeming data environ-
58 ment.6 In short, the ethicopolitics of machine learning algorithms is located
within the figure of the we — in the very relations to ourselves and to others im-
Chapter Two

plied in the we who have a spring of action.

In our contemporary moment, the “we both were” extends the already
multiple body via the sinewy and invasive techniques of deep learning and
neural network (neural net) algorithms.7 The extended we of the multiplicity
of data to which the learning algorithm is exposed heralds an intimate com-
munion of the learning machines with a vast and incalculable we: all of us, all
our data points, all the patterns and attributes that are not quite possessed
by us.
In the pages that follow, I begin by taking up the theme of intuitive learn-
ing via the extracted features of a data environment in the context of the
twenty-­fi rst-­century advent of surgical robotics. At the level of this specific
type of deep neural network algorithm, there is no technical distinction be-
tween learning actions for robot surgery and learning actions for robot weap-
onry. Across different domains of life, these algorithms are concerned with
translating the input data from their environment into a “feature space,” map-
ping the features into clusters of significance, and extracting the object of in-
terest for the action.8 Thus, algorithms designed to save lives, via robot surgery,
or to end lives, via robot warfare, share the same arrangements of propositions.
In following the machine learning of surgical robotics, I am concerned to cap-
ture the impossibility of establishing definitive boundaries of good and evil
in relation to algorithms. The machine learning algorithms deployed in robot
surgery do save lives through the lower infection rates of noninvasive methods,
but they also endanger life through error and miscalculation. My point is that
the ethicopolitics of machine learning algorithms cannot be mapped onto the
parameters of good and evil or the securing and imperiling of life. With the
extraction of feature spaces, machine learning algorithms are actively gener-
ating new forms of life, new forms of boundary making, and novel orienta-
tions of self to self, self to other. To begin with robot surgery is to begin in a
place where one could never definitively draw a line delineating the algorith-
mic moral good from some sense of immorality or evil.
Intuitive Surgery: Making the Singular Cut

The intuitive relation to mathematics noted by Turing finds a contemporary

form in robotic surgical systems such as Intuitive Surgical’s da Vinci robot.
The application programming interface (api) and the cloud storage architec-
ture of the da Vinci robot contain the data residue of multiple past human and
machine movements. Figure 2.1 displays the “surgical gestures” of the move-
ments of surgeons’ hands on the remote console, as mediated through the ro- 59
botic instruments of the da Vinci.9 The surgical procedure modeled here is a

The Learning Machines

routine four-­stitch surgical suture to close an incision. Though the surgical
gestures involved in the cutting and stitching of flesh are a matter of haptic
routine for human surgeons, for the designers of the algorithms, the objective
is to model the optimal suturing motion so that future human and robot sur-
geons have their intuitive movement shaped by the ingenuity of the model. As
the Johns Hopkins computer scientists building the model explain, the process
begins with the “automatic recognition of elementary motion” from the ex-
traction of features in the data environment. The model extracts seventy-­eight
features, or “motion variables,” from the vast quantity of video and sensor data
archived by the da Vinci robot (figures 2.2 and 2.3) — t wenty-­five feature vec-
tors from the surgeon’s console, and fourteen from the surgical instruments
attached to the patient-­side robotic arms.10 In the surgical gesture, the move-
ment of the human hand is thus thoroughly entangled with the remote con-
sole and the surgical scalpel held by the robot’s hand. The scientists describe
the juxtaposed map of surgical gestures: “The left [top] plot is that of an expert
surgeon, while the right [bottom] is of a less experienced surgeon.”11 Here, algo-
rithms are enrolled to recognize surgical gestures and to extract the features of
movement, to actively distribute cognition across human surgeons and robots,
and to optimize the spatial trajectory of the act of suturing flesh. The future
surgeon will learn to suture flesh optimally, via the robot’s simulation func-
tions, and the robot surgeon will learn to suture autonomously from the data
of past gestures of expert human surgeons. Though the computer scientists
do envisage autonomous actions by the robot — with “the possibility to auto-
mate portions of tasks, to assist the surgeon by reducing the cognitive work-
load” — this apparent autonomy is entirely contingent on the layered learning
from models of past entanglements of human and robot gestures.12
The spring to action of surgical machine learning is not an action that can
be definitively located in the body of human or machine but is lodged within a
more adaptive form of collaborative cognitive learning.13 Intimately bound to-
gether by machine learning algorithms acting on a cloud database of medical
60
Chapter Two

Figure 2.1  A plot of the trajectories of surgical gestures during a four-­stitch

suturing procedure. The top plot shows the trajectories of an expert surgeon,
while the bottom plot shows the trajectories of an inexperienced surgeon, as
extracted and analyzed by the surgical robot. Lin et al., “Automatic Detection.”

data, the we of surgeon and robot restlessly seeks an optimal spring of action —
 the optimal incision, the optimal target of tumor or diseased organ, the opti-
mal trajectory of movement. Intuitive Surgical’s robots hold the promise of the
augmented vision and precise trajectories of movement of a composite being of
surgeon and machine. As one of the uc Berkeley robotics scientists describes
what they term “iterative learning,” machine learning allows “robotic surgi-
cal assistants to execute trajectories with superhuman performance in terms
of speed and smoothness.”14 The drive for superhuman learning iterates back
Figures 2.2 and 2.3 
The da Vinci surgical
robot. Intuitive
Surgical, 2018.
 and forth across multiple gestures. As the hand and eye movements of a sin-
gular surgeon are tracked in the da Vinci’s api, they commune with a mathe-
matical model generated from a vast multiplicity of data traces of past surgical
gestures. These gestures, in turn, modify the learning model to optimize and
augment the future trajectories of future surgeons and robots as yet unknown.
As the philosophers and scientists in the 1952 debate anticipated, the machine
learning algorithms have something close to appetites, extracting and model-
62 ing features from the plenitude of cloud data, rendering springs of action, and
acting on future states of being.
Chapter Two

During the course of following the surgeons and robots of one world-­
leading oncology department in a UK teaching hospital, I began to note the
many occasions when a surgeon used “we” in place of “I” to describe their
daily collaborations with surgical robots. This we who learns is expansive. It in-
cludes the many humans in the research group and the surgical team, but also
the many human and technical components of the simulation of a surgery —
 the multiple layers of medical imaging, video data from past surgeries, and al-
gorithmic models that together composed a kind of virtual presurgery.15 For
example, when the UK surgical team were preparing to conduct a new surgi-
cal procedure on a specific type of tumor, they collaborated with other US
surgeons who had previous experience of the procedure. This was not merely
a dialogue between human experts, however; the research team also imported
the data from the US surgeries, inputting them in an algorithmic model and
experimenting with the parameters for a new context. As Rachel Prentice doc-
uments in her meticulous ethnography of surgical education, “surgical action
must be made explicit for computers” so that “bodies and their relations in
surgery are reconstructed” in a form that “can be computed.”16 This reciprocal
and iterative relationship between human surgeon and computer is what she
calls a “mutual articulation” in which “bodies affect and are affected by” one
another.17 When the movement of a surgeon’s hands is rendered computable
in the robotic model, Prentice suggests that this “instrumentalization” over-
looks the “tacit” and “tactile experience” of surgery, such as the “elasticity of a
uterus or the delicacy of an ovary.”18
In Prentice’s reading of surgical technologies, the tacit, tactile, and intui-
tive faculties of surgery define the human as the locus of care and embodied
judgment and decision. Yet, in giving their accounts of working with robots, I
found that human surgeons testify to their own body’s capacities coming into
being in new ways. What it means to be intuitive, to touch or to feel an organ,
for example, alters with the advent of machine learning modalities of surgery.
Seated at their virtual environment console, the surgeons access video feed
 63

The Learning Machines

Figure 2.4  Fluorescence imaging of renal parenchyma, as viewed from the
surgeon’s console. Intuitive Surgical, 2018.

images from the endoscopic arm of the robot (figure 2.4). These images are
overlaid with mri and the fluorescence images of tumors so that, as one of the
da Vinci computer scientists explains, “the system provides over a thousand
frames per second and filters each image to eliminate background noise.”19 The
work of the algorithms here is to extract the features of interest for perception
by the surgeon, surfacing the optimal image from a background noise of teem-
ing data.
One surgeon described to me the daily work of obstetric surgery with her
da Vinci robot, noting that through the expanded mediated space, she is able
to “see the unseeable” and “reach the unreachable” within the patient’s body.20
In contrast to Prentice’s sense that human touch — and with it, judgment and
decision — is evacuated and instrumentalized by surgical robotics, here the en-
tangled touch of surgeon and robot invokes different relations of judgment
and decision. The surgeon’s relations to herself and to others — to her patients
past and present, her operating theater colleagues, the robots, images, tumors,
surgical instruments — are altered in and through the machine learning algo-
rithms. “Touch engages us in a felt sense of causality,” writes Karen Barad, so
that “touch moves and affects what it effects.”21 Where Prentice foresees a loss
of responsibility as the human surgeon’s touch is evacuated by the robot, I pro-
 pose instead an extension of responsibility to that which extends and exceeds
human sensibility. The difference is an important one. The terrain of uncer-
tainty regarding whether a specific tumor is operable without major damage
to surrounding organs, for example, shifts with the 360-­degree mobility of the
robot’s wrist — what can be precisely touched, sensed, and extracted from the
body is altered. Indeed, leading computer scientist and artist Ken Goldberg,
alongside his carefully presented accounts of the development of stochastic
64 models and neural networks for the performance of surgical excision by ro-
bots, writes essays on the insensible and uncanny worlds that open up with
Chapter Two

his algorithms. Jochum and Goldberg’s account of the “experiential uncanny”

describes how robot actions “stretch the boundaries between the animate and
inanimate” in new directions, serving to “challenge our beliefs about what,
precisely, separates humans from machines.”22 The computer scientist’s reflec-
tions on the embodied and intuitive capacities of his algorithms — to change
the nature of what can be seen, reached, touched, or learned — r un against
the grain of a surgeon’s hands being rendered computable by autonomous ma-
chines. Instead, machine learning algorithms work with the incomputable to
open up new worlds of intuitive and insensible action.
As machine learning algorithms engage in “stretching the boundaries,”
the object that is surfaced for perception and action communes intimately
with data on the events of past surgeries. This communion on what is optimal — 
the cut, the incision, the surgical strike — belongs properly to a composite be-
ing within the cloud analytic I describe in the previous chapter. The da Vinci
data are no longer territorially limited to the memory of a specific robot, a
server, an individual surgeon, or a group of scientists. Rather, the machine
learning algorithms are deployed in a cloud architecture that yields the data
residue of many millions of past surgeries. Lodged inside the actions of the
singular cut — itself bordered by algorithms optimizing the thresholds of the
instrument’s trajectory — are the multiple data fragments of other entangled
composites of surgeon, software developer, programmer, neural network, pa-
tient’s body, images, and so on. The singular cut, within which teems a multi-
plicity, is present also in the autonomous vehicle, the drone, the smart borders
system — a lways also with multiple data fragments lodged within.23 In every
singular action of an apparently autonomous system, then, resides a multi-
plicity of human and algorithmic judgments, assumptions, thresholds, and
probabilities.
The Impossible Figure of the “Human in the Loop”

The neural network’s capacity to learn by extracting features from its data
environment has made it flourish in the algorithmic architectures of drones,
autonomous vehicles, surgical and production robotics, and at the biometric
border.24 This capacity to learn something in excess of taught rules has also
characterized the public concern and ethical debates around autonomous sys-
tems. Whether in the neural net algorithms animating surgical robots, au- 65
tonomous weapons systems, predictive policing, or cloud-­based intelligence

The Learning Machines

gathering, what is most commonly thought to be at stake politically and ethi-
cally is the degree of autonomy afforded to machines versus humans as a lo-
cus of decision. I suggest, however, that the principal ethicopolitical problem
does not arise from machines breaching the imagined limits of human con-
trol but emerges instead from a machine learning that generates new limits
and thresholds of what it means to be human. As legal cases proliferate amid
the errors, when the spring of action happens at the point of surgical incision,
smart border, or drone strike, they consistently seek out an identifiable reason-
ing human subject to call to account: a particular named surgeon, a specific
border guard, an intelligence analyst — the “human in the loop.”
In Hayles’s field-­defining book, How We Became Posthuman, she proposes
that the “distributed cognition of the posthuman” has the effect of compli-
cating “individual agency.”25 Hayles does not argue that a historically stable
category of human has given way, under the forces of technoscience, to an
unstable and disembodied posthuman form. On the contrary, the conception
of the human and human agency was, and is always, a fragile and contingent
thing. As Hayles writes,

The posthuman does not really mean the end of humanity. It signals in-
stead the end of a certain conception of the human, a conception that
may have applied, at best, to that fraction of humanity who have had the
wealth, power, and leisure to conceptualize themselves as autonomous be-
ings exercising their will through individual agency and choice. What is
lethal is not the posthuman as such but the grafting of the posthuman
onto a liberal humanist view of the self.26

Hayles’s concerns for the grafting of the posthuman onto the figure of an
autonomous liberal subject echo across the making of intuitive machine learn-
ing worlds. Though technologies such as Intuitive Surgical’s robot actively dis-
tribute and extend the parameters of sight, touch, and cognition into posthu-
man composite forms, their ethical orientation is defined solely in relation to
 the control of an autonomous human subject. While human surgeons speak of
an indeterminate we who learns, decides, and acts, nonetheless the capacity for
judgment retains its Kantian location in the unified thought of a reasoning hu-
man subject.27 Thus, when a violence is perpetuated or a harm is registered —
 damage, prejudicial judgment, or death — the only ethical recourse is to an
imagined unified entity who secures all representations. So, for example, in a
series of legal cases against Intuitive Surgical, the reported harms include the
66 rupture of tissue, burns, and other damage to organs, severed blood vessels and
nerves, loss of organ function, and fatalities.28 In these juridical cases, where
Chapter Two

the robot’s machine learning algorithms fail to recognize or to grasp precisely

the outline of the target, what is sought is a unified locus of responsibility — a
company, a negligent surgeon, or a hospital — an entity imagined juridically to
be autonomous and unified, whose choices and agency can be held to account.
Similarly, when autonomous weapons systems make errors in their target se-
lection, or cause “collateral damage” amid the so-­called precision strike, the
ethical appeal is made to an accountable “human in the loop” of the lethal-
ity decision.29 The notion of an ethical decision thus appears in the form of a
reasoning human subject or a legal entity with a capacity to be a first person I
who is responsible.
Yet, where would one locate the account of a first-­person subject amid the
limitless feedback loops and back propagation of the machine learning algo-
rithms of Intuitive Surgical’s robots? When the neural networks animating
autonomous weapons systems thrive on the multiplicity of training data from
human associations and past human actions, who precisely is the figure of the
human in the loop? The human with a definite article, the human, stands in
for a more plural and indefinite life, where humans who are already multiple
generate emergent effects in communion with algorithms.30 Recalling Geof-
frey Jefferson’s 1952 question, “Who was learning, you or the machine?,” and
Turing’s reply, “We both were,” the human in the loop is an impossible subject
who cannot come before an indeterminate and multiple we.
Perhaps what is necessary is not a relocated human ethics — of feedback
loops and kill switch control — for a world of the composite actions of human
and algorithm. What is necessary, I propose, is an ethics that does not seek the
grounds of a unified I but that can dwell uncertainly with the difficulty of a
distributed and composite form of being. As machine learning changes the re-
lations we have to ourselves and to others, the persistent problems of a Kantian
unity of thought is newly dramatized by algorithmic formulations of learning
and acting. To begin to address this different kind of ethicopolitics, one must
dwell with the difficulty, as Donna Haraway suggests, making cloudy trouble
for ourselves methodologically and philosophically.31 Such a tracing of algo-
rithmic threads as they meander through unilluminated space involves ask-
ing questions of how algorithms iteratively learn and compose with humans,
data, and other algorithms. To be in the dark, to dwell there in an undecidable
space, is to acknowledge that our contemporary condition is one in which the
black box of the algorithm can never be definitively opened or rendered intel-
ligible to reveal its inner workings. To trace the algorithm in the dark is not
to halt at the limits of opacity or secrecy, but to make the limit as threshold 67
the subject of study. Such a task begins by asking how machine learning algo-

The Learning Machines

rithms learn things about the world, how they learn to extract features from
their environment to recognize future entities and events, what they discard
and retain in memory, what their orientation to the world is, and how they
act. If intuition never was an entirely human faculty, and never meaningfully
belonged to a unified I who thinks, then how does the extended intuition of
machine learning feel its way toward solutions and actions? To this task I now
turn.

Regimes of Recognition: How a Neural Network Makes the World

Allow me to begin by describing a scene — a laboratory designing machine

learning algorithms for border and immigration control systems — where a se-
ries of neural networks learn to recognize people and things via the features in
their data environments. One of the designers explains that his algorithms are
trained on border and immigration data with many hundreds of thousands of pa-
rameters. He describes how he “plays with” his developing neural nets — taking
the experimental model to the uniformed border operations team in the ad-
joining building to test it against the specific targets they are seeking in the al-
gorithm’s output.32 This traveling of the model between laboratory and opera-
tions center consists of a series of questions about whether the algorithms are
useful, or if they are “good enough.” This question, Is it good enough?, illumi-
nates some of the politically contested features of the algorithm’s emergence.
Though for the border operations team, “good enough” may be a measure of
the algorithm’s capacity to supply a risk-­based target for a decision at the bor-
der, for the computer scientists, “good enough” means something quite dif-
ferent and specific. In computer science, a “good enough” solution is one that
achieves some level of optimization in the relationship between a given target
and the actual output of a model.33 Understood in this way, it is not the accu-
racy of the algorithm that matters so much as sufficient proximity to a target.
Put another way, the algorithm is good enough when it generates an output
that makes an optimal decision possible. When the algorithm designers de-
 scribe tuning or playing with the algorithm, they are experimenting with the
proximity between the target value and the actual outputs from their model,
adjusting the probability weightings in the algorithm’s layers and observing
how the actual risk flags generated by their model diverge or converge on the
target.
The design of an algorithmic model, then, involves a contingent space of
play and experimentation in the proximities and distances between the actual
68 output and a target output. My concept of the space of play designates specifi-
cally the distance between a target output and an actual output of the model.
Chapter Two

This space of play, however, also opens onto an infinite array of combinato-
rial possibilities in terms of the malleable and adaptable inputs, parameters,
and weights of the model. As one designer of machine learning algorithms
for anomaly detection frames the question, “What is normal?” and “How far
is far, if something is to be considered anomalous?”34 Precisely this adaptive
threshold between norm and anomaly was being negotiated between the labo-
ratory and border operations. A small adjustment in the threshold will gener-
ate an entirely different set of outputs and, therefore, a change in the spring of
action. I have observed this iterative process of playing with the threshold in
multiple situations where algorithms are being trained for deployment, from
police forces adjusting the sensitivity of a facial recognition algorithm to casi-
nos moving the threshold for potential fraud:

You must experiment to determine at what sensitivity you want your

model to flag data as anomalous. If it is set too sensitively, random noise
will get flagged and it will be essentially impossible to find anything use-
ful beyond all the noise. Even if you’ve adjusted the sensitivity to a coarser
resolution such that your model is automatically flagging actual outliers,
you still have a choice to make about the level of detection that is useful to you.
There are always trade-­offs between finding everything that is out of the
ordinary and getting alarms at a rate for which you can handle making a
response.35

What is happening here is that the neural networks are learning to recog-
nize what is normal and anomalous at each parse of the data. But, the shifting
of the thresholds for that recognition embodies all the valuations, associations,
prejudices, and accommodations involved in determining what is “useful” or
“good enough.” Sometimes, as with semisupervised machine learning, this re-
gime of recognition emerges iteratively between humans, algorithms, and a
labeled training dataset. In other instances, unsupervised machine learning
will cluster the data with no preexisting labeled classifications of what is or
is not useful or of interest. Even in this apparently unsupervised process, hu-
mans recalibrate and adjust the algorithm’s performance against the target. In
short, in all cases, machine learning algorithms embody a regime of recogni-
tion that identifies what or who matters to the event. Machine learning algo-
rithms do not merely recognize people and things in the sense of identifying — 
faces, threats, vehicles, animals, languages — they actively generate recogniz-
ability as such, so that they decide what or who is recognizable as a target of
interest in an occluded landscape. To adjust the threshold of what is “good 69
enough” is to decide the register of what kinds of political claims can be made

The Learning Machines

in the world, who or what can appear on the horizon, who or what can count
ethicopolitically.
The kind of ethicopolitics I am opening up here is somewhat different
from the attention that others have given to the inscription of racialized or
other prejudicial profiles in the design of the algorithm.36 Though identifying
the human writing of prejudicial algorithms as a site of power is extraordi-
narily important, the regimes of recognition I have described actively exceed
profiles written into the rules by a human. The machine learning algorithms I
observed, from borders to surgery, and from facial recognition to fraud detec-
tion, are producing modes of recognition, valuation, and probabilistic deci-
sion weighting that are profoundly political and yet do not reside wholly in a
recognizable human who writes the rules. They are also, of course, calculative
spaces where prejudice and racial injustices can lodge and intensify, though
not in a form that could be readily resolved with a politics of ethical design or
the rewriting of the rules.37

The Hitherto Unseen: Detecting Figures, Detecting Objects

If machine learning algorithms are changing how something or someone

comes to attention for action, then how does this regime of recognition come
into being? Often the most apparently intuitive of human actions — to recog-
nize a face in a crowd, to distinguish the features of a cat from a dog, to know
how best to reach out and grasp an object — present some of the most difficult
computational problems. When algorithms are understood as a series of pro-
grammable steps formulated as “if . . . and . . . then,” it is precisely in the writ-
ing of the rules for the sequence that one decides the result: who or what will
be of interest, or who or what can be recognized.38 A common exemplar of the
problem of recognizing the unseen is the capacity of algorithms to recognize
handwritten digits.39 The variability of the form of a figure — its “profile” — 
exposes the limit of rules-­based algorithms that define the features in advance.
How would one formulate rules for recognizing the handwritten number 3? In
 the traditional decision trees designed by J. R. Quinlan in the 1980s, the “prod-
uct of learning is a piece of procedural knowledge that can assign a hitherto-­
unseen object to one of a specified number of classes.”40 To recognize an un-
seen object, the decision tree algorithm classifies it according to “a collection
of attributes” describing its important features.41 How would one begin to de-
fine the attributes of the figure 3 amid the variability of its form? As Quinlan
acknowledged, the limits of the procedural knowledge of rules-­based classifiers
70 are encountered in unknown features that were absent in the training dataset.
“The decision trees may classify an animal as both a monkey and a giraffe,”
Chapter Two

wrote Quinlan, or the algorithm may “fail to classify it as anything.”42 Proce-

dural classifiers such as decision trees, then, learn to recognize hitherto unseen
objects according to the presence or absence of a set of properties encountered
in the training data. Returning to the example of handwritten digits, would
the variability of the form of the figure result in the decision that a Ʒ is not a
3? Or indeed that a Ʒ is not classifiable, or has the attributes of a 5? Though the
recognition of handwritten figures is a common exemplar, the limit of proce-
dural classifiers also applies to many other recognition problems, from facial
recognition technologies to security threats, weather patterns, advertising op-
portunities, or the likely pattern of votes in an election. Put simply, while a
rules-­based classifier recognizes according to the profiled properties of an en-
tity, the contemporary neural network algorithm learns to recognize via the
infinite variability of features it encounters.
What does it mean to learn about the world in and through the variability
of features in the environment? With the growing abundance of digital images
and cloud data for training machine learning algorithms, the process of learn-
ing shifts from recognition via classification rules to recognition via input data clus-
ters. Let us explore this further through the example of recognizing numeric
figures. Figure 2.5 shows a simplified illustration of the spatial arrangement
of a deep (multilayer) neural network — of the type I have sketched on many
whiteboards at conferences and workshops. If the target of the algorithm is to
optimize the likelihood of correctly identifying a handwritten digit, for ex-
ample, then the training data will consist of a dataset of handwritten digits,
each figure segmented to the level of pixels. What this means is that the algo-
rithm does not learn to recognize the profile of the figure 3 per se, but rather
learns to recognize the clustered patterns in the array of pixels in the image.43
The input data in the neural net — and consider that in other instances, this
could be anything: images, video, biometric templates, social media text — is
assigned a series of probability weightings for its significance, with the output
 Adjustment, Error, Bias

Inputs

Thresholds of Target
Space Output
Video play
Social media
Images Output
71
Weights Decision

The Learning Machines

Audio
Travel Aperture

Transactions
Probability

Sentiment

Hidden Layer 1 Hidden Layer 2

Figure 2.5  A representation of the arrangements of a deep neural network.

of each neuron to the next “hidden layer” dependent on whether the weighted
sum of inputs is less than or greater than some threshold value. Each layer of
the neural network detects some discrete aspect of the figure. As the computer
scientists describe image recognition, “the learned features in the first layer”
detect “the presence or absence of edges,” with the second layer “spotting par-
ticular arrangements of edges,” and the subsequent layers “assembling motifs
into larger combinations that correspond to parts of familiar objects.”44 The
recognition of edges, motifs, and familiar arrangements is not designed into
rules by a human engineer but is definitively generated from the exposure to
data. To be clear, this spatial arrangement of probabilistic propositions is one
of the places where I locate the ethicopolitics that is always already present
within the algorithm. The selection of training data; the detection of edges;
the decisions on hidden layers; the assigning of probability weightings; and the
setting of threshold values: these are the multiple moments when humans and
algorithms generate a regime of recognition.
 Adjusting the Features: The Variability of What Something Could Be

The computational problem of how to recognize people and things has become
of such commercial and political significance that computer scientists enter
their experimental algorithms in competitive image recognition contests. One
particular algorithm, the AlexNet deep convolutional neural network, won an
image recognition contest in 2012 and has become the basis for multiple subse-
72 quent commercial and governmental recognition algorithms, with the scien-
tific paper cited more than twenty-­four thousand times.45 The AlexNet gives
Chapter Two

an account of itself — in the terms of a partial account I am advocating — that

manifests just how its output is contingent on its exposure to data features,
and the series of weightings, probabilities, and thresholds that make those fea-
tures perceptible. As I recount something of how the AlexNet algorithm does
this, I would like you to consider that if an algorithm is deciding “Is it a leop-
ard?” or “How likely is it that this is a shipping container?” then it is also de-
ployed to decide “Is this a face?” and “Is this face the same face we saw in the
street protest last week?”46 Understood in this way, the regime of recognition is
political in terms of both arbitrating recognizability and outputting a desired
target that is actionable.
As the computer scientists who designed AlexNet describe the relationship
between recognition and cloud data, “objects in realistic settings exhibit con-
siderable variability, so to learn to recognize them it is necessary to use much
larger training sets.”47 The AlexNet cnn was trained on 15 million images,
each image labeled by a human via Amazon Mechanical Turk’s crowdsourcing
labor tool.48 Figure 2.6 shows eight test images for the algorithm, with the five
labels considered most probable by the model assigned beneath each image.
The algorithm is able to recognize a previously unseen image of a leopard or a
motor scooter because the feature vectors of the image have close proximity to
the gradients encountered in the training data. The capacity of the algorithm
to recognize an incomplete creature at the edge of the frame (the mite) is con-
sidered to be a major advance in neural nets for image recognition. Where the
algorithm failed to recognize an entity — the grille, the cherry — the scientists
refer to the “genuine ambiguity” of which object is the focus of the image, and
where the patterns of edges are occluded (e.g., the Dalmatian’s spots and the
cherries). To be clear, the logic of the AlexNet algorithm is that if one exposes
it to sufficient data on the variability of what a leopard could be, then it will
learn to anticipate all future instances of leopards. Indeed, deep neural nets
are exposed to infinite variabilities — voting behaviors, faces in crowds, credit
histories, kidney tumors, social media hashtags — to recognize the feature vec-
 73

The Learning Machines

Figure 2.6  Test images of the AlexNet image recognition algorithm.
Krizhevsky, Sutskever, and Hinton, “ImageNet Classification.”

tors of all future instances. Whether someone or something can be recognized

depends on what the algorithm has been exposed to in the world. Since the al-
gorithm makes itself — adjusts thresholds and weights, for example — through
its exposure to a world in data, it is becoming the contemporary condition of
recognizability as such.
Like the cloud chambers of chapter 1, the propositional arrangements of
the neural net are instruments of mattering, methods for making some things
matter more than others. To seek to open or to make transparent the black
box of this arrangement would be neither possible nor desirable, for the ar-
rangement is an important site of politics, the spatiality of the calculus being
politically significant in and of itself. This is significant for the ethicopolitical
interventions one might wish to make because, for example, it could never be
sufficient to demand that facial recognition algorithms that fail to recognize
black faces be trained on a greater variability of images. For the algorithm also
learns how to afford weight or value to one pixelated part of an image over oth-
 ers (the Dalmatian and not the cherry, the edges of this face and not that one).
Indeed, as one computer scientist explained to me, a neural net like AlexNet,
with six or eight hidden layers, is too complex even for the designer of the algo-
rithm to explain the conditional probabilities that are learned. “I might adjust
the weighting in that layer,” he explains, “and I know that this will change the
output, but I cannot say exactly how.”49 As with the design of AlexNet, the
computer scientists work with the essentially experimental and unknowable
74 nature of the algorithm. They perceive the fractional changes in the output of
the model as they adjust the weightings, working with the emergent and un-
Chapter Two

knowable properties of machine learning.

Bias Can Be a Powerful Ally

When deep neural network algorithms learn, then, they adjust themselves in
relation to the features of their environment. To be clear, to learn, they have
to weight some data elements of a feature space more than others — they have
to have assumptions about how the world is ordered. Notwithstanding the
widespread societal calls for algorithms to be rendered free of bias or to have
their assumptions extracted, they categorically require bias and assumptions
to function in the world. Indeed, even the textbooks used by the next genera-
tion of computer scientists address directly that “there can be no inference or
prediction without assumptions,” particularly the assumptions of “the proba-
bility assigned to the parameters.”50 Thus, when a team of European computer
scientists discuss how they might move the threshold for their neural net algo-
rithm to recognize the likelihood of a person of interest (a future person, yet
to arrive) being a “returning foreign fighter” and not a “returning aid worker”
from Syria, they mean that they will adjust the sensitivity of the algorithm to
particular elements of weighted input data, such as increasing the probability
weighting of particular past flight routes.51 While some of this adjustment of
the threshold is done by humans, today much of it is invested in the power
of the algorithm to adjust itself in and through the emergent properties of
the data, understood as a feature space. The “we both” of Turing’s reflections
seems to reassert itself here in the accounts of adjustment given by computer
scientists Yann LeCun, Yoshua Bengio, and Geoffrey Hinton of Facebook ai,
Google, nyu, and the University of Toronto: “We compute an objective func-
tion that measures the error (or distance) between the output scores and the
desired pattern of scores. The machine then modifies its internal adjustable
parameters to reduce this error. These adjustable parameters, often called
weights, are real numbers that modify the input-­output function of the ma-
chine. In a typical deep-­learning system, there may be hundreds of millions of
these adjustable weights.”52
Like the abductive methods of intelligence gathering I discuss in chapter
1, this computational method observes the effect of the calculation — or the
output signal — a nd theorizes back to the adjustment of parameters, like the
mechanical knobs on a calculating machine. The distance between an agreed
target output, or desired pattern, and the output scores is the error or the bias.
Significantly, for the algorithm, error is distance; it is the playful and experimen- 75
tal space where something useful or “good enough” materializes. There is noth-

The Learning Machines

ing normatively wrong about error in a machine learning algorithm, for it is a
reduction of difficulty and difference into a mere matter of distance. Likewise,
let me be clear, bias and weighting are not negative things for an algorithm.
They are, on the contrary, essential elements of learning, so that, in computer
science, “bias can be a powerful ally.”53 My point is that one could never sat-
isfactorily address the ethicopolitics of algorithms by calling for a removal of
human or machine bias and a reduction of error because the machine learning
algorithm would cease to function at this limit point. Bias and error are intrin-
sic to the calculative arrangements — and therefore also to the ethicopolitics — 
of algorithms. At root, the algorithm can never be neutral or without bias or
prejudice because it must have assumptions to extract from its environment,
to adapt, and to learn. It is, ineradicably and perennially, a political being. To
begin from here is to begin from the idea that all machine learning algorithms
always already embody assumptions, errors, bias, and weights that are fully
ethicopolitical. In the adjustment of parameters one can locate a shifting ter-
rain of the relations of oneself to oneself and to others. The output of the algo-
rithm is but a mere numeric probability, fragile and contingent, so that a tiny
adjustment of the weights in the algorithm’s layers will radically change the
output signal, and with it the basis for decision and action.

Point Clouds and the Robot’s Grasp

To extract something from the features in a data environment, to anticipate

and to act, is a critical computational problem for deep machine learning algo-
rithms in production line robotics, surgical robotics, drones, and ied (impro-
vised explosive device) detection. Across these diverse domains, the capacity
to recognize the three-­dimensional form of an object and to decide on the op-
timal action is a challenge that animates computer science. Indeed, the failure
to recognize multidimensional and mobile forms — such as those of human or-
gans, vehicles, or facial features — has been a common feature of many high-
 profile mistakes and accidents by machine learners. In the fatal Tesla autono-
mous vehicle crash of 2016, for example, one way to articulate the error would
be to say that the cnn algorithms failed to recognize the profile of a white van
against a pale sky as the vehicle turned across the Tesla’s path. The probabi-
listic answer to the question “Is this a vehicle?” was, fatally, “no.” Discussions
among computer scientists regarding the causes of such accidents are revealing
in terms of a persistent determination to locate the source of the fatal flaw and
76 to annex the algorithm from its milieu. For example, one group of ibm scien-
tists urged caution “not to blame the algorithm for a failure of the sensors, the
Chapter Two

ambient lighting, or the human operator.”54 My point, though, is that what the
sensor can sense, or the operator can decide, is only meaningful in the context
of how the neural nets arbitrate what the objection could be, what it could
mean. Similarly, in a Volkswagen factory in 2015, a robot failed to recognize
the outline of a human coworker on the production line, mistaking him for
a car door and crushing him to death. At the level of machine learning algo-
rithms and their regimes of recognition, the da Vinci surgical robots’ failures
to recognize the boundary delimiting kidney tumor from human organ is not
dissimilar to the biometric facial recognition systems that closed automated
border controls at an airport when the setting sun changed the ambient light-
ing conditions. In all these instances, the algorithms have learned from the
features of the environment they have been exposed to. Sometimes events and
sensors in the environment will present them with a set of input features they
have not encountered previously, and their assumptions and weightings may
lead to a spring of action that misrecognizes the target.55 Is this an error? Or is
error merely a matter of distance?
The contemporary advent of cloud robotics has sought to address this
problem of the limit point of exposure to features in a multidimensional en-
vironment. Cloud-­based robotics, as we saw in the discussion of surgical ro-
bots, circulate data and aggregate computational power across a distributed
system of machine and human learning. Just as the surgical robots are no lon-
ger limited to the data and computation stored within a bounded system, so
the cloud-­based intelligence system I discuss in chapter 1 recognizes its targets
from exposure to data and analytics methods across borders and jurisdictions.
Where machine learning intersects with cloud computing, the neural network
algorithms are exposed to features from a vast archive of cloud data, includ-
ing the Point Cloud Library of open source 2d and 3d images.56 A point cloud
is a set of topological data points mapping the 3d space of objects. Computer
science research in cloud robotics is addressing the question of whether expo-
sure to a vastly increased volume of point cloud data on objects can optimize
the neural network’s capacity to learn how to recognize and to act. Consider,
for example, the computer science team at uc Berkeley’s Automation Sciences
Lab, whose research into cloud robotics is funded by the nsf, Google, and the
US Department of Defense. Presenting their Dex-­Net 1.0, or Dexterity Net-
work algorithm, the Berkeley scientists experiment with cnns to optimize the
capacity of a robot to recognize and grasp a range of objects. The algorithm
represents an advance on image recognition technologies such as AlexNet be-
cause it recognizes 3d objects from multiple viewpoints, and it outputs an op- 77
timal action based on this recognition.57

The Learning Machines

The Dex-­Net algorithm is trained on an archive of Google point cloud
data on “3d object models typically found in warehouses and homes, such as
containers, tools, tableware, and toys.”58 The neural nets are learning to rec-
ognize the object’s geometry and topology and then to optimize the robot’s
capacity to reach out and grasp the object. In a sense, the algorithm is asking
a two-­step question — W hat is this object? and How can it be most effectively
grasped? This shift toward cnns that can recognize and optimize an action
is absolutely critical in the advance of robotics in manufacturing, medicine,
and the military. The scientific papers on Dex-­Net show something of the log-
ics at work in coupling regimes of recognition to what I call a spring of action.
When the Dex-­Net algorithm is exposed to a training dataset of one thousand
3d point cloud objects (in figure 2.7, a household spray bottle), it is not able to
find a “nearest neighbor” object that will allow it to recognize the query object.
When the algorithm is exposed to the point cloud features of ten thousand
objects drawn from the Point Cloud Library, however, it finds two proximal
objects, or nearest neighbors, allowing it to recognize the object and optimize
its grasp.
Though at first glance the Dex-­Net’s machine learning may appear as
though the cloud is supplying “big data” volume to the algorithms, in fact the
process of reduction and condensation I describe in chapter 1 is also taking
place here. As the computer scientists propose, the significance of the “cloud-­
based network of object models” is actually to “reduce the number of samples
required for robust grasps” and to “quickly converge to the optimal grasp.”59
Put simply, the Dex-­Net algorithm is better able to condense and filter out
the occlusions to recognize the most similar object. Each of the ten thou-
sand cloud-­based objects is prelabeled with 250 parallel-­jaw robot grasps, each
weighted with a probability of a successful grasp. “The goal of cloud robotics,”
as the computer scientists explain, is to “pre-­compute a set of robot grasps
for each object” so that “when the object is encountered, at least one grasp is
achievable in the presence of clutter and occlusions.”60 Understood in these
78
Chapter Two

Figure 2.7  Point clouds for object recognition and grasp optimization.
Mahler et al., “Dex-­Net 1.0.”

terms, the weightings and probabilities of the point cloud make it possible to
precompute something, so that the encounter with the unknown object can
always yield an action that is optimal. With the point cloud, an algorithmic
system does not only ask “Is this a face?,” “Is this a bottle?,” “Is this a military
vehicle?,” but rather it has already precomputed an optimal action in relation
to the topologies it has encountered.

Precomputation: “The Fundamental Thing

Is We Know What Good Looks Like”

On the horizon of research in neural net algorithms for robotics, then, one
finds this notion of precomputation to make an action achievable amidst clut-
ter and occlusions. Precomputation captures the neural net computational
problem that extends from the recognition and grasp of the shape of a hu-
man organ amid “noisy” surgical data, to the recognition of a civilian body
in the screened occluded data of the drone. One computer scientist, who now
designs algorithms for human gait recognition, described to me how he had
been “working with surgeons modeling the perfect operation,” this human-­
algorithm collaboration itself deploying point clouds to precompute the trajec-
tory of “a perfect procedure.”61 He has developed his algorithms for multiple
applications, with each iteration another step in optimization: “The funda-
mental thing is we know what good looks like,” he explains, “and then if you’re
doing anomaly spotting, you can see when something’s wrong.” For example,
when the online gambling platform BetFair sought algorithms to recognize
and act on patterns of addition, the designer suggests, “that’s something we
can spot because we know what addicted play looks like.” This malleable nor- 79
mative assumption of “what good looks like” or “what addictive play looks

The Learning Machines

like” is generated in and through the algorithm learning in a feature space.62
Precomputation implies that some sense of what is a “perfect procedure”
or “at least one achievable grasp” action is always already present within the
algorithm as such. The ethicopolitics of machine learning algorithms like
AlexNet and Dex-­Net is in the bias, weights, thresholds, and assumptions
that make recognition precomputable. To precompute is to already be able to
recognize the attributes of something in advance, to make all actions imagin-
able in advance, to anticipate every encounter with a new subject or object, a
new tumor or terrorist, by virtue of its proximity to or distance from a nearest
neighbor. The condition of possibility of the algorithm’s action is its exposure
to an archive of cloud data, condensed via the infinitely malleable value system
of weights, probabilities, thresholds, and bias.
This is a pressing problem of the politics of algorithms in our contempo-
rary moment. All our handwritten digits; all our online data traces; the bio-
metric templates of our facial geometry; the point clouds of household objects
and military hardware; all the movements of the hands, eyes, and bodies of
surgeons, pilots, soldiers, consumers, production line workers: these are the
teeming conditions of possibility of the machine learning algorithm. We are
it, and it is us. We could never stand outside it, even if we might wish to. Each
of the data fragments that enters the point cloud has a part to play in the
learning. Whose is the grasp that caused the injury? Which of the 2.5 million
objects in the archive became the nearest neighbor? Which of the possible 250
grasps for each object? Which of the many tens of thousands, or millions, of
cloud-­derived probabilities was responsible for the grasp that intuitively de-
cided to pull the trigger, so to speak?
The harms inflicted through machine learning are not located primar-
ily in the ceding of human control to machines, as is so often assumed in the
ethical and moral debates on algorithmic decisions. Indeed, as we have seen
via the surgeon who learns to reach and touch differently with her da Vinci
robot, what it means to be human is significantly transformed in and through
 the machine learning algorithm. To appeal to the human as locus of ethics,
then, is to appeal to a being already entangled with new forms of knowing and
learning. The principal harm, in contrast, is manifested instead in a specific
threat to a future politics. The tyranny of proliferating machine learning al-
gorithms resides not in relinquishing human control but, more specifically, in
reducing the multiplicity of potential futures to a single output. The claim to
precompute the future, or to know “what good looks like” at the border, in the
80 operating theater, in the economy, forecloses other potential futures. To be
clear, the neural net does not reduce multiplicity as such. After all, as I have
Chapter Two

outlined, the spatial arrangement of the neural net algorithm contains within
it multiple probabilities, infinitely adjustable weights whose emergent effects
can never be entirely known, even to the designer. The finite elements of each
hidden layer of the neural net, one might propose, contain within them infi-
nite possible correlations to other elements. The spatial arrangement of the
neural net does not foreclose alternative readings, different arrangements of
what or who matters and what or who does not. Crucially, however, at the
point of action, this intrinsic multiplicity is reduced to a single output. The
insistence on a single output is the algorithm’s orientation to action. Though I
am reminded by the computer scientists that the output need only be between
0 and 1, and that there are infinite numbers between 0 and 1, there is none-
theless a single numeric output. Let us not forget that the algorithm’s output
signal lies behind the risk score at the border, the credit decision, the target as-
sessment of the drone, and the decision on sentencing, detention, or the incipi-
ent dangers of a gathered protest on a city street. It is as though all the many
potentials held in parallel, simultaneously distributed across the layers of the
neural net, could never have been. With the output of the machine learning
algorithm, one might say, things could never have been otherwise. The output
is a probability whose value is transformed by the smallest of adjustments in
the parameters of the model. And yet, nonetheless, all political uncertainty is
rendered tractable on the horizon of the action triggered by this single output.
At this point, one might reasonably ask how giving such an account of the
contingent politics of machine learning algorithms is of any possible critical
use. How might a cloud ethics work with the incompleteness, the undecid-
ability, and the contingency of the algorithm’s space of play? If one wants to
inquire whether a given algorithm is responsible for a flash crash in the fi-
nancial markets, or if one seeks a human rights law adequate to the task of
holding autonomous weapons or autonomous surgery to account, then some
ethical grounds might be considered essential — or at least some method for
accountability. As Michel Foucault proposes in his discussion of ethics, how-
ever, what may be necessary is not to appeal to grounds or to the juridical do-
main of statutes, but rather to “ask politics a whole set of questions that are
not part of its statutory domain.”63 A cloud ethics must be capable of asking
questions and making political claims that are not already recognized on the
existing terrain of rights to privacy and freedoms of association and assembly.
Cloud ethics belong properly not to the individual as bearer of rights, but to
the many touch points and data fragments that are aggregated from the rela-
tions between subjects and objects. Thus, a cloud ethics must be capable of 81
asking questions such as How did that Dex-­Net algorithm weight the prob-

The Learning Machines

ability of that future grasp?; Why did the training data teach the algorithm to
recognize this and not that object amid the occlusions?; How was the distance
between target and output signal (bias) used as a space of experimentation?;
and, In outputting that score, what were the traces of the rejected alternative
weights and parameters? Such questions are necessary and urgent, even and
perhaps essentially when they are unanswerable. The unanswerable questions
reawaken the multiplicity that was, in fact, always present within the machine
learning algorithm. All the many contingencies and alternative pathways are
reopened, and the single output bears the fully ethicopolitical responsibility
for the actions it initiates. The processes and arrangements of weights, values,
bias, and thresholds in neural nets are, I think we can safely say, not part of
our statutory political domain. And yet, I suggest that they must be presented
as questions and political claims in the world.
This page intentionally left blank
 Part 2

Attribution
This page intentionally left blank
3
The Uncertain Author
Writing and Attribution

I always think the notion of the fork in the road is very important when you are
creating narrative, because you are continually coming to forks. You don’t exactly
know where you are going. . . . This is a state of uncertainty, or in terms of the
modern physics, indeterminacy.
— John Fowles and Susanna Onega, Form and Meaning

Source Code

In October 2017 the New York City Council proposed a new local bill — the Al-
gorithmic Accountability Bill — conceived as a means to render the city’s algo-
rithms accountable for the significant effects their use has on the life chances
of the people of New York City.1 The city council’s attempt to secure the ac-
countability of algorithms was a response to the racialized harms precipitated
by algorithmic decisions in the criminal justice system, where automated risk
calculations further entrenched the notion that young African American men
posed a high risk of reoffending and should serve custodial sentences. 2 As in
many other cases when an algorithm has caused a harm or injustice, the reflex
response of the authorities was to seek out the origin of the algorithm and lo-
cate the source of its harmfulness. In the first iteration of the Algorithmic Ac-
countability Bill, the text called for public and private agencies “to publish the
source code” of all algorithms being used for “automated decision systems,”
“policing individuals,” or “targeted services” and “imposing penalties.” The
sponsor of the legislation, council member James Vacca, recorded his concerns
that automated decisions made by algorithms were discriminating against
New Yorkers based on the characteristics of race, age, religion, gender, sexual
orientation, and citizenship status. Announcing that his ambition was to seek
 “transparency as well as accountability,” Vacca located the actions and effects
of the algorithm in what he saw as the determinate and discriminatory agency
of the source code.
Though the Algorithmic Accountability Bill was subsequently amended
to protect the commercial interests of algorithm vendors, before it was passed
unanimously in December 2017, the appeal to a source code for the algorithm
represents a broader impulse to locate the origins of the algorithm, or more
86 precisely, to trace the source of the algorithm to an originary act of authorship.
Put simply, the source of the algorithm becomes attached to the unified body
Chapter Three

of its code, anchored in the modern notion of the author who writes the code.
The responsibility the algorithm can be said to have — for discrimination, for
errors, for mistakes and injustices — is thus secured via the attribution of the
authorship of code. This appeal to the authorship of the algorithm character-
izes the many contemporary demands for accountable and transparent algo-
rithms or, as I have described it, demands for an encoded ethics for algorithmic
decision systems. An encoded ethics reasserts the human author of mathemat-
ical code and locates the control of algorithms in the reasoning authority of
the author. In some instances, as in the New York City Council’s scrutiny of
source code, the appeal to authorship is made through the body of work, via
the locus of an identifiable source code. Here, code is assigned a unity of the
body of work in the same way that authorship produces a unity from an oth-
erwise errant and distributed text. In other manifestations, an encoded ethics
is sought via an appeal to the writers of algorithms to adopt an ethical code of
conduct that would secure the moral grounds for their writing of code.3 Again,
the location of ethics in the reasoning subject is mirrored by the irrevocable
binding of the author to the body of work. Responsibility for the onward ac-
tions and effects of the algorithm is imagined to be secured through the disclo-
sure of authorship. Moreover, the call for accountable and ethical algorithms
binds together not only source code and author but also data sources and au-
thorship, so that it is said that “designers must make their data sources pub-
lic.”4 A semblance of accountability for how the algorithm comes into being is
found in the disclosure of the data sources from which the author writes. In
sum, through the appeal to authorial sources and ethical design, much of the
contemporary debate on algorithmic accountability is imbuing the algorithm
with a specific form of what, with Michel Foucault, we might call an “author
function.”5 This novel form of algorithmic author function has profound con-
sequences for how one understands, and responds to, the ethicopolitical rela-
tions forged through the writing of algorithms.
If the source code is the place where the logics of the algorithm can be
found and disclosed to the world, then what exactly is the nature of this source?
As I explain in chapter 2, in the context of how machine learning algorithms
learn to recognize, the algorithm’s ways of being in the world are not all present
in the code. Indeed, the effects of the algorithm are generative and emergent so
that the source code is never complete. Put differently, the source is continually
edited and rewritten through the algorithm’s engagement with the world. To
understand the latent discriminations of the New York City algorithms neces-
sarily involves engaging with its reincorporation of new entities into its code. In 87
this sense, the sources and authors of the algorithm are continuously modified

The Uncertain Author

and elaborated through its actions in the world. To make its actions account-
able via the source code is to misunderstand how what is meant by the category
of author is reformulated in and through the algorithm. The authorship of the
algorithm is multiple and distributed, extending across the attributes of one
data population to act on the propensities of another. We are, each and every
one of us, enrolled in the practice of the writing of algorithms. Through the
practice of writing, we enter into new ethicopolitical relations with ourselves
and with others. So, what is an author, to restate Foucault’s famous question,
in the context of algorithmic authorship?6 In this chapter I explain the limits of
locating an ethical response to algorithms in the authorship of source code. In
place of the search for an origin of algorithmic harm, I propose that the actions
of algorithms are producing new articulations of the author function that al-
ready possess politics and ethics. This already existing ethicopolitics of author-
ing algorithms could form the starting point to a rather different proposition
than the search for accountability and transparency of sources. Understood in
this way, the author function does not serve as the locus for accountability but
is implicated in the authorization of algorithmic decisions. Though the desire
to scrutinize source codes for intrinsic bias or discrimination seeks to control or
regulate the decision-­making authority of the algorithm, in fact the attribution
of authorship is authorizing all kinds of renewed algorithmic decision.
I suggest an alternative formulation in which the ethicopolitics of algo-
rithms dwell not in the originary responsibility of authorship but instead in
the act of writing itself. Notwithstanding some of the distinctions that have
been made between the writing of algorithms and the writing of literary text,
I am mindful of the ways in which the notion of authorship operates at specific
historical moments and across the boundaries of science and literature.7 I draw
on the reflections of novelists on the difficulties of generating an account of
their authorship, not because there is a direct analogy between writing nov-
els and writing algorithms, but because all forms of writing necessarily open
onto an indeterminate relationship to oneself and to others. As Vicki Kirby
 proposes in her feminist reformulation of deconstruction, there “is no ‘every-
thing’ that pre-­exists the relationality that is the scene of writing.”8 Taking se-
riously Kirby’s argument that the elements of the scene are already present in
the instantiation of writing, and that writing enrolls human and nonhuman
authors, the writing of algorithms also involves the instantiation of a scene.
If writing is also world making, then to give an uncertain, partial account of
oneself as author is to open onto the difficulties of the I who makes decisions
88 and chooses pathways as well as the other readers and writers who will follow
different forks and turns in the future. As the novelist John Fowles and his co-
Chapter Three

author Susanna Onega reflect in the opening to this chapter, the practice of
writing involves encountering “forks in the road” at which it is never certain
what might take place as a result of the decision at the branch. To write, as
Fowles experiences it, is necessarily to encounter an uncertain and indetermi-
nate future. Though the writing of algorithms and the writing of literature is
not at all the same thing, of course, the bifurcated “forks” in the algorithm’s
calculative architecture do involve a profound uncertainty of authorship and
its future effects. Conceiving of writing as a practice in which scattered ele-
ments and contingent forks in the road are gathered and unified, the writing of
algorithms does constitute something akin to an ethicopolitical writing prac-
tice that involves “manifesting oneself to oneself and to others.”9
Returning to my account of the ethics of algorithms located in the rela-
tions of self to self and self to others, writing algorithms shares with the nov-
elists’ accounts the impossibility of a determinate and grounded authorship.
In place of locating ethics in the individual subject of the author and the au-
thorial source of the work, the ethics of algorithm emerges from the compos-
ite and collaborative practices of writing in which we, and our scattered data
elements, are also implicated and engaged. To propose that algorithms dwell
within us, and we within them, is also to suggest that through these relations
an ethics of writing algorithms is forged. In this reformulated approach to au-
thorship and the algorithm, it would be insufficient to respond to the violences
and injustices of an algorithm by “scrutinizing the source code” because the
act of writing these algorithms substantially exceeds that source, extending
into and through the scattered elements of us and the multiple data points
that gather around us.

The Author and the Attribute

The computer science field of natural language processing (nlp) has become
of intense interest to those developing experimental deep learning algorithms.
Beginning in the mid-­twentieth century with models for automatic language
translation, nlp for much of its history has used basic decision tree algorithms
and rules-­based techniques to render human language computable.10 In my
book The Politics of Possibility, I traced Rakesh Agrawal’s rule-­based method of
text mining as it developed from the mining of customer retail basket data in
the early 1990s to the mining of data for counter-­terrorism after 2001.11 With
the availability of volumes of cloud-­based unstructured data in the twenty-­
first century, however, nlp methods have shifted from the rules-­based deriv-
ing of semantics from text to deep learning algorithms generating meaning 89
from their exposure to a corpus of data derived from speech or text.12 The

The Uncertain Author

significance of this transformation from rules-­based determinants to genera-
tive models of the meaning of a text is that the algorithms are no longer seek-
ing exact one-­to-­one matches of semantic meaning; rather, they are building
predictive models of what might come next. They are, in short, inferring fu-
ture worlds from their exposure to text. Thus, for example, corporations are
increasingly shaping their customer relations management around models of
customer propensities built from the nlp of social media text. Similarly, as I
detail in chapter 1, security and intelligence agencies are using nlp of Twitter
and Facebook data to build predictive models of future threats. These compu-
tational problems are not understood primarily in terms of the recognition of
the semantic meaning of text, but instead are concerned with computing hu-
man context, meaning, patterns of behavior, and possible futures.
This manifest desire to find computational methods for futures-­oriented
inference from text has witnessed a curious new relationship emerging be-
tween science and literature (and between data and narrative forms), where
literature becomes a corpus of contextually nuanced data from which a deep
learning algorithm is trained to recognize the meaning of human language.
An example of how computer science is turning to literature to teach machines
how to build predictive models of meaning is led by the computer scientist and
author of the notion of a coming singularity, Ray Kurzweil. Google’s Natural
Language Understanding Research Group (note understanding, not processing)
is teaching algorithms how to infer the subtleties and nuances of human lan-
guage.13 The Google team trained a deep neural network algorithm on a corpus
of data comprising the literary works of one thousand deceased authors, from
William Shakespeare to Daniel Defoe, and from Virginia Woolf to Herman
Melville.14 The neural net algorithm was reported by the scientists to have dis-
covered the literary style of particular authors from the patterns within their
body of work. In fact, the algorithm had done what most machine learning al-
gorithms do: it had clustered the text according to the features in the training
data and defined these clusters in terms of the attributes of the author’s body of
 work. A similar method was used in 2013 to attribute the pseudonymous novel
A Cuckoo’s Calling to the author J. K. Rowling, with the attributes of nom de
plume Robert Galbraith’s writing correlated to Rowling’s.15
Once recognized and learned, the attributed features of an author’s text
become a computational means to identify the future attributes of as yet un-
known texts. This attributive capacity allows algorithms to anticipate future
features of a text, whether a person’s financial patterns or their writing on
90 social media. During its development and testing, the Google nlp algorithm
was given an input sentence from a novel not present in the training dataset,
Chapter Three

with the target output defined as a probability that this sentence was a match
to a given second sentence. At this point the data was unlabeled — the algo-
rithm was not given the name of the author — yet it made a correct prediction
in 83 percent of the input queries. As the Google team account for their ex-
periments, their guiding question was, “Can we build a system that can, given
a sentence from a book and knowledge of the author’s style and ‘personality,’
predict what the author is most likely to write next?” In a later phase of the ex-
periment, the team built a generative model that was able to respond to ques-
tions in the style of a specific author, said to be predicting the future text of
“dead authors” from “beyond the grave.”16 In short, the algorithmic technol-
ogy of the attribute is not about identifying the qualities of an entity in the
present, but about rendering something attributable into the future.
The significant point here is that the author function in contemporary
deep learning algorithms is serving a futures-­oriented capacity to associate en-
tities through their attributes. The name and source “identity” of the author
may be unknown or anonymous, but nonetheless the algorithm is deriving
an author function from the clustering of attributes. Put simply, what mat-
ters to these deep learning algorithms is not primarily the naming of an au-
thor but the generating of a set of attributes that could be latent in any other
entity. As I have similarly described the border-­control algorithms that can
act on an individual indifferent to their naming or identification — as a set of
associations — so natural language algorithms need only to be able to associ-
ate the attributes sufficiently to say “this sentence X matches the next possible
line Y with probability P.”17 Even where the strict naming of the author is not
possible, the algorithm can work productively with the attributes of the body
of work.
One might reasonably ask what political or ethical significance could pos-
sibly be attached to this distinctive learning of attributes via clustering tech-
niques. Yet, the computational capacity of nlp algorithms lies behind many
of the apparently autonomous decision-­making systems that are of ethical and
political concern. From the extraction of text from images of public protests,
to the analysis of the sentiment of social media text for detecting the pro-
pensities of a person, these algorithms are concerned precisely with the infer-
ence of possible futures that could be associated with a set of attributes. The
Google team explains the motivation for the one thousand authors project as
“an early step towards better understanding intent.”18 Given the attributes of
a given cluster within a corpus of data, what is the future intent? The predic-
tive power of the attribute is located in the capacity to infer meanings in the 91
future. Thus, for example, how does the intent of one line of text — “I will go

The Uncertain Author

and sort him out” (to seek revenge and to threaten violence) — differentiate it-
self from the intent of another that is semantically identical — “I will go and
sort him out” (to offer support and to care for someone’s troubles)? It is exactly
this kind of subtle distinction between ostensibly identical text that is sought
by the algorithm’s designers. For the algorithm, the capacity to discern differ-
ence is derived from the extracted features of written style and genre as a set
of relations. Contemporary algorithms to be used across domains, from credit
card fraud to counter-­terrorism, are being trained to understand future intent
through the attributes of style and genre.
As the Google experiments with authorship and inferring future intent
suggest, though the algorithm does not require a definitive named author per
se, it does seek to establish an author function in which the attributes of a cor-
pus of text yield the writing of the next line into the future. As Michel Fou-
cault proposes, “Criticism and philosophy took note of the disappearance — or
death — of the author some time ago.”19 Understood in this way, the disappear-
ance of the author is not a recent phenomenon but is intrinsic to the act of
writing, in which “the writing subject continually disappears.”20 And yet, at
the very moment when it seems the author has disappeared, just when it might
seem tempting to appeal to the author by name and demand responsibility
for what is written (as New York City Council members did in their call to
scrutinize the source code), “a certain number of notions that are intended
to replace the privileged position of the author actually seem to preserve that
privilege.”21
Here, in Foucault’s discussion of how the author function continues to
live and thrive amid the apparent death of the author, I find some resources
for understanding what is happening when algorithms are learning from a
corpus of text to infer future propensities. For Foucault, even when the author
appears to have lost status, the body of work stands in for authorship so that
“the word work and the unity that it designates are probably as problematic
as the status of the author’s individuality.”22 Thus, even when a named indi-
 vidual author is absent (almost always the case with algorithms), the notion
of a work signals a stylistic unity of characteristics of particular authorship.
For this reason I understand the source code of an algorithm to invoke the au-
thor function of designating a unitary whole even when this apparent whole is
composed of heterogeneous fragments of data and past writings. Similarly, the
corpus of literary data from dead authors, used by Google to train algorithms,
is afforded the unity of a “work”: a unified collection of styles, patterns, and
92 behaviors from which future attributes can be derived. What categorically
does not count as part of the authorial work in the Google nlp, however, is
Chapter Three

the actual work of the algorithm as it iterates back and forth, adjusting prob-
ability weights and thresholds and modifying its own code. This annexing of
the writing of source code from the back-­and-­forth iterations of an algorithm
rewriting has profound implications for ethics. The designers of facial recog-
nition algorithms for policing, for example, commonly testify that bias and
discrimination reside in the human operator and not within their code. When
one follows the iterative writing of the algorithms in practice, however, one
sees that the actions of a human operator are in close communion with the
algorithms, so that each decision also recalibrates the model. In short, to an-
nex the writing of code from the iterative writing in the world is to seriously
miss the ethicopolitical life of algorithms. As Foucault suggests, the preserved
dominance of the author function hinders us from “taking the full measure”
of the author’s disappearance and attending to what is happening in the gaps
and spaces of that disappearance.23
Foucault’s writings on the subject of authorship also afford us a glimpse of
what might be happening when texts are grouped together for machine learn-
ing. In the Google experiments, what matters for the algorithm is not primar-
ily who the author is, but rather what attributes make a set of texts part of
a cluster defined by degrees of similarity. For Foucault, the author function
serves a kind of “classificatory function” in that it “permits one to group to-
gether a certain number of texts, define them, differentiate them from and
contrast them to others.”24 Though the form of algorithmic author function
I wish to depict is not classificatory in the strict sense of grouping things by
similarity, it does serve a clustering function in terms of finding structure in
data and learning what normal or abnormal looks like. In this sense the au-
thor function retains its privilege to find a resolution of differences and ten-
sions in the “authentication of some texts by the use of others.”25 To be clear,
it is profoundly misleading to appeal to the author of the algorithm or to the
source code as the locus of responsibility because the algorithm’s power lies
in its capacity to promise to resolve incalculable differences via the unity of
its outputs. To invoke the call for an identifiable and responsible author is not
only insufficient as critique, it also risks amplifying the already present ability
of the algorithm to “permit grouping together” so that “incompatible elements
are tied together,” with “all differences having been resolved.”26 In place of the
ethicopolitical call for securing responsibility via authorship, then, I am inter-
ested in deepening our sense of how the algorithm ties together incompatible
elements to appear to resolve the politics of difference.
93
Authorship and the Algorithm

The Uncertain Author

In Foucault’s discussion of the author function in sciences and literature, he
notes the shifting historical modes of authorship in the eighteenth and nine-
teenth centuries. Remarking that in European history, “it has not always been
the same types of texts that have required an author function,” he reminds us
that there was a time when literary narratives were widely circulated and ac-
cepted without an author because of their ancient and received origins. Just as
literary texts could be authenticated anonymously, so “those texts we would
now call scientific” were accepted as truth “only when marked with the name
of their author.”27 A transformative reversal of the author function took place
in the eighteenth century so that “the author function faded away in science,”
where truths were founded on proof, theorem, axiom, and law. Simultane-
ously, literary texts came to be accepted only when afforded a particular form
of author function: “We now ask of each poetic or fictional text: From where
does it come, who wrote it, when, under what circumstances, or beginning
with what design? The meaning ascribed to it and the status or value accorded
it depend on the manner in which we answer these questions. And if a text
should be discovered in a state of anonymity — the game becomes one of redis-
covering the author.”28
The significant point is that the notion of authorship does not belong nat-
urally to the literary world in a way that it does not belong to science.29 Nor is
it inappropriate to use the concept of authorship in relation to the writing of
algorithms, for the category of the author has functioned differently across do-
mains of knowledge and at different historical moments. In our contemporary
moment, the author function is being rearticulated in a novel way in relation
to algorithms. As computer science has begun to embrace the emergent and in-
tuitive faculties more commonly historically associated with literature, so the
logic that science has shed authorship while literature has retained it begins to
erode. Contemporary computer science is invoking just the kind of questions
that Foucault ascribes to each poetic or fictional text. Indeed, the questions
used to train deep learning algorithms follow closely those associated with lit-
 erary authorship: From where did it come? Beginning with what design? What
is the meaning ascribed? What is the value? The twenty-­first century marks
a moment when the specific form of the author function is once again chang-
ing its shape. “It is not enough to repeat the empty affirmation that the au-
thor has disappeared,” writes Foucault, but “instead we must locate the space
left empty by the author’s disappearance, follow the distribution of gaps and
breaches, and watch for the openings this disappearance uncovers.”30
94 It is this distribution of breaches and openings that I wish to revalue in a
reformulated author function of the algorithm. Indeed, Foucault signals what
Chapter Three

some of the breaches might be in the author function within mathematics. Ac-
knowledging the plurality of self involved in the literary novelist’s use of I in
the first-­person narrator’s account, he points to the “scission” between the au-
thor and the “real writer,” the “plurality of self ” that opens up in all cases “en-
dowed with the author function.” Crucially, this plurality of self is specifically
present also in “a treatise on mathematics,” where authorship splits, divides,
and disperses so that it “can give rise to several selves, to several subjects.”31 A
significant opening, then, begins from the way the first-­person I as a locus of
authorial responsibility is actually always divided and dispersed. This is not
only a matter of the dividuated subject, divided within herself, as is often doc-
umented in the accounts of digital subjects. The plurality of authorship also
involves a dispersal across different human and nonhuman life forms.32 For ex-
ample, as the neural net algorithms animating robots become more sensitive
to context and the attributes of behavior, they are modifying their own actions
by learning from their interactions with other robots.33 The algorithms are not
simply the “authors of their own code” in the sense of locating authorship in
an autonomous machine, but instead they are engaged in acts of writing and
rewriting that unify otherwise scattered and heterogeneous elements.
So, when a deep neural net predicts a deceased author’s next sentence, or
when a New York City policing algorithm decides the attributes of a person
at high risk of reoffending, within the author function teems a multitude. In
attempting to find mechanisms of accountability for algorithms, our societ-
ies are searching for an authorial source of the writing of code. Yet, this desire
for a first-­person author as the locus of agency and responsibility deepens the
power of the algorithm to present itself as a unified entity.34 Simultaneously,
it underplays the contingency of the algorithm and conceals the breaches that
might otherwise appear in the constraining figure of the author. In the fol-
lowing pages I propose a different approach to authorship and the algorithm,
foregrounding the multiple and dispersed acts of writing that are its condition
of being.
Writing and the Algorithm

To this point my argument has emphasized how the origins of algorithmic

actions are commonly understood to reside in the unity of the source code.
Anchored in the modern notion of the author, responsibility for the decisions,
harms, and injustices of algorithms tend to be traced to the notion of a source
code where errors or bias could reside. I have explained why this appeal to a
source code sustains the status of identifiable authorship as the locus of re- 95
sponsibility, binding the author to the work and unifying the works as though

The Uncertain Author

homogenous and undifferentiated. At the very moment algorithms usher in
novel forms of the author function, the resort to ethical frameworks for cod-
ing sustains the vision of a corrective “fix” in which society asserts control over
the actions of the algorithm. In what follows I outline how understanding the
ethics of algorithm requires a break with notions of authorship, and I develop
instead concepts of writing algorithms that substantially exceed the method
of tracing sources.
A place to begin to explore the distinction between authorship and writing
is where the source codes of an algorithm were the subject of legal contesta-
tion but, once released, were followed by unexpected events that exceeded the
scrutiny of the code. In the 2017 federal gun possession case of United States v.
Kevin Johnson, presiding judge Valerie Caproni ruled that the source code of
the New York Chief Medical Examiner’s Office (cme) dna analysis algorithm
had to be revealed to the defense team.35 The Forensic Statistical Tool (fst)
used by the New York cme, in common with other commercial algorithms
used to analyze dna material at crime scenes, is an algorithm designed to
make it possible to calculate the likelihood of an individual person’s dna be-
ing present at a scene, even when the dna material is an incomplete profile
and is mixed with the dna of others. In effect, with these machine learning
algorithms, incompleteness and uncertainty are rendered calculable as prob-
abilities. The important difference between algorithms such as fst, TrueAl-
lele, and strmix — which have been used in many tens of thousands of crimi-
nal convictions in the United States and elsewhere — and traditional statistical
methods for analyzing dna material is precisely how the probability is gener-
ated by the algorithm.36 The fst algorithm generates a likelihood ratio, which
is a numeric probability that a known suspect individual contributed dna to a
mixed and incomplete sample as compared to the probability that an unknown
other subject in the general population contributed that dna. When there is an
incomplete dna profile as part of a mixture of dna material at a crime scene,
the algorithm compares that incomplete fragment with the probability that it
 was derived from the general population. As I explain in chapter 2 in relation
to machine learning, this process means that the attributes of one population
are used as the basis for calculating the probabilistic likelihood of the actions
of another person.
In the case of the fst algorithm, as described in the forensic biology pro-
tocols of the New York cme, “the likelihood ratio value provides a measure-
ment of the strength of support for one scenario over another, i.e., one sce-
96 nario being that a known person contributed to a mixture versus the scenario
that an unknown, unrelated person contributed instead.”37 In short, where
Chapter Three

the dna profile of an individual cannot be identified from the materials at the
scene — where the trace is incomplete or indeterminate — the likelihood ratio
algorithm allows for partial samples to be “positively associated” with an indi-
vidual by comparison of their attributes to those of a group of unknown oth-
ers. A dna profile match becomes, via the algorithm, a correlative and scenario-­
based dna positive association.
There can be little doubt of the profoundly troubling implications of using
such algorithms to convict individuals of crimes.38 When one looks closely at
what happened when the source code of fst was released to the defense law-
yers, however, it becomes clear that the source of harm resides not in the code
itself but in the iterative and heterogeneous writing of a likelihood ratio. Once
one relinquishes the notion that the authorship of the algorithm resides in the
code, following instead the writing of the algorithm, the code is not the au-
thorial source of the algorithm’s actions but merely one element in an iterative
process of writing. Among the defense team expert statements — each of them
concluding that the algorithm could have resulted in innocent people serv-
ing a prison sentence for serious offenses — one computer scientist, Nathaniel
Adams, reported what he found when he examined the fst algorithm. His
findings do not pertain to an error or flaw in the source code or even to a bias
contained therein but specifically to the way in which the algorithm learns to
write a likelihood ratio from a body of data. In a method that is resonant with
the Google machine learning algorithm’s prediction of the next sentences of
deceased authors based on a corpus of literary data, the fst learns from “the
databases accompanying the source code.”39 These databases include the train-
ing and validation sets from which the likelihood ratio model is generated.
The algorithm modifies its output according to adjustments in the input data.
Adams experimented with adjustments to the weights of the algorithm and
demonstrated the contingency of the writing of a likelihood ratio when it is
expressed as an output score. Of particular significance is that the algorithm’s
sense of what the unknown population looks like in terms of dna is clustered
along racialized attributes. Adams discusses the four “subpopulation” datasets
used for the training of the algorithm — Caucasian, Hispanic, African Ameri-
can, and Asian — and he emphasizes the contingent effects of the “removal of
a locus” from these subpopulations on how the algorithm learns to generate
a likelihood ratio. In sum, the root of harm is not the authorial source code
but predominantly the way the algorithm is written and how it writes itself — 
continuously editing, adjusting, removing, and iterating in relation to a cor- 97
pus of data. The attributes of the dna of a “general population” and an “un-

The Uncertain Author

known unrelated person” do not preexist the text but are written into being.
As the computer scientist testifies in his expert statement, “Due to the com-
plexity of probabilistic genotyping algorithms, seemingly small modifications
can have unintended consequences and compounding effects that alter the be-
havior of the algorithms.”40 Let us not forget that these tiny modifications will
change the output score of the algorithm and, therefore, change the likelihood
of someone being present at a crime scene.
Though ostensibly the source code of the genotyping algorithm was of ju-
ridical and political concern, once the code was released onto the software de-
velopment platform Github so that it forged new relations of possibility, what
came to matter was not an imagined anomalous source code but the capacity
of the algorithm to continue to write its relations to normalities, abnormali-
ties, likelihoods, and probabilities into the future. Thus, the algorithm’s rela-
tionships with data is one of writing the world, bringing into being a singular
output — a likelihood ratio — from the disparate, heterogeneous, partial, and
fragmented scattering of data elements.
In proposing that one could follow the breaches and openings of the author
function, then, I am making the iterations, contingencies, and likelihoods of
the algorithm the focal point of ethical concern. The calls for access to the
source code of algorithms are a means to try to read how a likelihood ratio is
written. Yet, scrutinizing the code itself would reveal nothing of the racial-
ized, prejudicial outputs of the algorithm that are written via training data,
validation data, and the feedback loops of experiments in the field. Indeed, the
Github platform is itself a form of distributed and iterative writing in which
multiple developers contribute to the rewriting and editing of software. If one
is to take seriously the practice of writing as an ethical resource in respond-
ing to algorithmic injustice, then how does one follow this practice of writing?
What kinds of strategies can allow us to pursue the writing of the algorithm in
place of its authorship? What might the effects of this foregrounding of writ-
 ing over authorship be? Inspired by the reflections of novelists on the practice
of writing, in what follows I suggest a series of openings onto the writing of
algorithms: bifurcated paths; I as a character; and the obligations of the reader.

Bifurcated Paths
In an essay on the practices of science and literature, John Fowles reflects on
his experience of what happens when he writes. “An important term — at any
98 rate for me, in my own practice,” writes Fowles, “is the fork (as in a path), by
which I mean a fairly continuous awareness of alternatives, both learned (re-
Chapter Three

membered) and fortuitous (wild), in what is done.”41 In Fowles’s terms, the

practice of writing involves the encountering of contingent moments, or forks
in the path, where a decision is taken amid the continuous awareness that al-
ternatives exist. To take the alternative path at the fork, Fowles seems to sug-
gest, will change the narrative of the novel, modify its future course and po-
tential ending. Indeed, for the novelist, this partiality and opacity concerning
what is to come in the future characterizes the very value of writing. Reflect-
ing on the writing of William Golding, whom he calls a “master fabulator,”
Fowles notes a sustained “lack of expectability” that does not conform to an
expectation of style but works against the grain of the literary rhythms and
patterns of the time.42 There is an almost ethical emphasis in Fowles’s notion
of writing against expectation and with uncertainty, an ethical emphasis that
extends from literature to the sciences. “Any true scientific knowing,” writes
Fowles, “is always, like feeling, only partial.”43 The fork in the path of science,
as in literature, is taken with partial knowledge and in the face of profound
uncertainty: “Cruel, painful, even death-­dealing though it may be, we could
not live in this world without its sheer incomprehensibility.”44
Inspired by Fowles’s reflections on his practice of writing, an important
distinction to be made is that between a claim to authorship that precisely
seeks to secure against uncertainty and produce the expected (where even
the next lines of a deceased author can be expectable and inferred from their
style), and a form of writing that opens onto uncertainty and the unexpected.
The writer as fabulist, as Fowles describes Golding, and as Gilles Deleuze pro-
poses, is “someone who creates their own impossibilities, and thereby creates
possibilities.”45 In this way the practice of writing is also a practice of fabula-
tion, in which the idea of the author is continually undone by the writing, by
the impossibility of knowing in advance where the forks in the path might
lead.
This encountering of forks in the path is exactly what characterizes the
writing of algorithms. Once one has relinquished the idea of authorship and
focused on the writing, all algorithmic arrangements involve forking paths,
from the branches and leaves of decision trees to the differential paths through
the neural network. Where for the writer of the novel the fork in the path may
represent a decision amid alternatives in the context of uncertainty, so for
the writing of algorithms the fork is a weighing or adjustment of probabilities
and likelihood. What could be made to matter, in terms of the ethicopolitical
response to algorithms in society, is to trace these branching points: not as a
means of securing the knowable source, but precisely to confront the profound 99
contingency of the path. If a forensics algorithm outputs a 0.62 probability

The Uncertain Author

of a young African American woman’s dna being present as a fragment in a
mixture at a crime scene, then how does the adjustment of the weights at each
fork change the dissolution of alternative paths, other decisions not taken?
Which other unknown populations had to be enrolled for this likelihood to
be followed while another was discarded? How can the writing of the algo-
rithm be reopened at the points where there could be renewed awareness of
what cannot be attributed? What kind of interventions could be made at the
moment of condensation, when something brimming with other potentials is
reduced to one?

“I Is Simply Another Character”

John Fowles, writing in 1967 on the difficulties of writing his then-­unfinished
novel The French Lieutenant’s Woman, recounts how the first-­person narrative
does not consist of “the ‘I’ who breaks into the illusion” of fiction, but rather
engages the “I who is part of it.”46 The I is not the authorial voice who enun-
ciates and controls the narrative, then, but is simply a character among the
many who together write the narrative. The I who writes the first-­person com-
mentaries in Fowles’s unfinished novel is not “my real ‘I’ in 1967” but instead
“simply another character” in the iterative play of characters in the novel. Even
the first-­person account, then, does not secure the authoring subject and take
them outside the fabulation of the writing of the novel. On the contrary, as lit-
erature and philosophy have long proposed, writing is the opening of a space
into which the author disappears.47
To conceive of the writer as simply another character in the novel is to
interrupt the author function and its place in the sovereign economies of the
ownership of text and the responsibility for the originary source of a text.48
The novelists and philosophers of the late 1960s shared an interest in the ca-
pacities of writing to interrupt the authorial voice and open onto the writ-
ing of indeterminate and uncertain future worlds. It is no coincidence that
in 1967, when the novelist John Fowles displaces the author in his account
 of writing, the philosopher Jacques Derrida publishes Of Grammatology and
defines the methods of deconstruction. In these traditions, writing always
breaches the authorship of a text because it necessarily exceeds its context,
iterating beyond the moment of being inscribed and enduring “even after the
death of its author.”49 Reading the insights of deconstruction into my treat-
ment of algorithms as arrangements, I find a profound contrast between an au-
thor function that witnesses algorithms predicting the next lines of a deceased
100 author based on clustered attributes, and a writing that iteratively exceeds the
I of the author. To bring to the fore the writing of algorithms is to unsettle the
Chapter Three

notion that one could locate responsibility for the harms of an algorithm in
its source code. The algorithm iterates beyond the moment of its inscription,
distributing the writing through multiple characters, from the training data
to the back propagation of errors. This matters because any meaningful re-
sponse to the actions of algorithms, as we saw in relation to genotyping neural
networks, would need to relinquish the search for the source code and focus
instead on the iterative writing back and forth between dna data on the popu-
lation and the making of a likelihood ratio. What would be sought in such a
strategy would be the awareness of alternative paths, as Fowles suggests. There
are multiple learned and “in the wild” features at work in the writing of algo-
rithms, from the many fragments of past data lodged in the criminal justice
databases to the feedback loops that adjust and modify the algorithm into the
future.

“To Oblige the Reader to Help Form the Text”

In an interview in 1995, John Fowles reflects on the place of silences and omis-
sions in literature. “One of the greatest arts of the novel is omission,” he sug-
gests. “I’m a deep believer in silence — the ‘positive’ role of the negative.”50 In
the gaps and omissions of the text, as Fowles concurs with Samuel Beckett
and Herman Melville, is the space where the reader enters “to help form and
to experience the text” and where reading becomes a “heuristic process.”51 The
relation between writer and reader, then, is a profoundly ethical one in which
the writer must address an absent subject who will nonetheless modify the
writing, and the reader cannot simply read according to the signs or rules of
the text. As Simon Critchley poses the question, “How is the ethical relation
to the other person to be inscribed in a book without betraying it immeasur-
ably?”52 In response, Critchley proposes that “ethics is first and foremost a re-
spect for the concrete particularity of the other person in his or her singular-
ity,” so that “ethics begins as a relation with a singular, other person who calls
me into question.”53 The writing of a novel, then, has what Fowles describes as
an “ethical function,” not because it lays down for the reader the principles or
rules of how it ought to be read, but on the contrary, because it contains gaps
and openings for the concrete singularity of the reader. The writer and the
reader are necessarily and intractably tied together in the difficulty and obliga-
tion of forming the text. And there can be no critical outside to their relation,
but always an immersed involvement in the processes of writing and reading.
“One cannot stand on the bank,” says Fowles. “One is willy-­nilly in the stream.
And by ‘one’ I mean both writers and readers.”54 101
To be in the stream, then, is to be already involved in the ethicopolitical

The Uncertain Author

acts of writing and reading. When one hears of the automated analysis of data
streams, it is exactly this imagination of already being in the stream that I con-
sider to be so crucial. We do not stand on the bank and respond to the stream,
but we are in the stream, immersed in the difficulties and obligations of form-
ing the text. The broader implications of being in the stream as the writers
and readers of algorithms are significant. One so often hears the claim that
algorithms are illegible and unreadable, that they operate at scales and speeds
in excess of human reading. But my point is that we should remember that al-
gorithms are not unique, or somehow “outside the text.” All acts of writing and
reading, whatever the text, necessarily confront illegibility and the impossibil-
ity of reading.55 We do not need new resources or technologies to prize open
the unreadable algorithm. Instead we could begin from that unreadability as
the condition of all engagements with text. It is something familiar to us all
as social and political theorists. We are, in fact, well equipped to confront the
difficulty of what seems to us an opaque text.
An ethics of algorithm could usefully respond to the specific ways in
which algorithms address the others with whom they read and write the fu-
ture. Critchley places his emphasis on the ethical address to the singularity of
the other person, but the writing of the algorithm categorically recognizes no
singularity or particularity. The algorithm learns to recognize only a set of dif-
ferences from, or similarities to, the attributes drawn from another data popu-
lation. The injunction for the reader to complete the text, then, can only find
a response in the exposure to more data to write the text. In place of a search
for the authorship of the algorithm, I am asking for a response to the concrete
particularity of the other person who is obliged to help form the text. What
this means in practice is that no person would be addressed as the “probable
association” derived from a likelihood ratio algorithm. The algorithm writes
the world into being in ways that oblige others to form the text, but as it does
so, it necessarily also encounters a singular reader whose particularities enter
the omissions in the text in ways that are unexpected.
 Fabulation and the Algorithm

In the fifth of Hilary Mantel’s series of bbc Reith Lectures in 2017, the novelist
expresses the relationship between telling history and narrating a story. “We
need fiction to show us that the unknown and unknowable is real and exerts
its force,” she argues, dispelling the notion that the facts of history are distinct
from the fictions of her narratives of Tudor England in the novels Wolf Hall
102 and Bring up the Bodies.56 For Mantel, the writing of fiction confronts the real-
ity of the profound unknowability of our world. The unknown nature of our
Chapter Three

actions, relationships, and intentions is the truth contained within the novel.
As I have discussed through the essays of John Fowles, writing is distinct from
authorship precisely because it must confront the force of the unknown, and
not least the unknowable I that is the writing subject. I have sought not to
diametrically oppose the writing of the novel and the writing of algorithms,
but rather to distinguish the author from the practice of writing so that the
sovereignty of the source code is opened up to the writing, of which it is one
element. This writing — whether literature or code — necessarily and irredeem-
ably confronts the force of the unknown and the unknowable, as Mantel de-
scribes it. What is at stake here is a rather fundamentally different response to
the problem of the algorithm’s accountability. Where the demand for ethical
frameworks for code seeks to make the algorithm knowable and governable,
my call for attention to the writing of algorithms demands a different mode
of account.
To call for the algorithm to give an account of itself is not to demand dis-
closure or transparency, but precisely to express something of the unknow-
ability of the algorithm and the future that it brings into being. In this sense,
like novels, algorithms can be understood as fabulations, where the production
of an output involves a process of working on a corpus or body of material.57
“The fabulatory function,” as Gilles Deleuze proposes, “strips us of the power
to say ‘I.’ ” To write is not to anchor a narrative in a determinate first-­person
source, but precisely to work with the incomplete and indeterminate traits
of characters and their connections. Writing is fabulation because it projects
into the future with “a life of its own” beyond the author’s schema, a life that is
“stitched together” and “continually growing along the way.”58 Of course, one
could object that the algorithm actually produces completeness and a deter-
minate future via a single output, an automated decision that goes out into the
world of criminal justice, health risk, security, and war. Yet, this is precisely
the paradox of the process of writing, of fabulating a text. The writing of the
algorithm fixes a unity from scattered data elements, at that same moment
fabulating new connections and traits, forging attributes that will attach to
other beings into the future.
The task as I see it, then, is to find ways to amplify the fabulation of the al-
gorithm, to enter the breaches in the writing so that the force of the unknown
can be felt. I understand the writing of the algorithm to substantially exceed
the writing of source code and to extend into the iterative writing, editing,
and rewriting of composites of data, humans, and other algorithms. Writing
in this mode is a “practice of the disparate,” consisting of gathering “hetero- 103
geneous elements” so that traits, sensibilities, and attributes emerge from the

The Uncertain Author

composite.59 To be clear, the practice of writing algorithms is located not in the
individual author, but in the multiple and dispersed writers of the algorithm — 
from the data that quietly enters the training set to the person who experi-
ences the full force of “high risk” attributes. For even the target of the algo-
rithmic decision becomes a new set of data to be worked into the refining of
weights and probabilities. If fabulation involves what Deleuze calls “strictly
speaking a series of falsifications” generated by working on a material, then the
algorithm appears anew.60 Stripped of its authorial power to speak the truth
and to produce objective outputs, the writing of the algorithm necessarily in-
volves a series of falsifications. The algorithm is not false in the sense of un-
true — it is indifferent to truth except in relation to data-­grounded truths — but
false in the sense of wrought from a set of materials that could be infinitely
reworked.
To be clear, it is not my argument that the writing of algorithms is analo-
gous to the writing of literature. On the contrary, I argue that the philosophies
and social theories of writing contain a rich resource for considering all forms
of writing in an extended sense, as a “network of traces.”61 All languages and
codes, then, signify through differential traces, where a trace makes present
something not present in the past. In common with all languages and codes,
the algorithm writes to make present something that is not merely in the past
data. Interrupting the notion that an algorithm’s source code works on the
material of what is in past data, the writing of algorithms fabulates the traits
and attributes of a cluster so that it makes present what did not exist in the
past. What is at stake here is that public discussion of the accountability of
algorithms is concerned primarily with the problem that the source code of al-
gorithms cannot be read. Yet, the unreadable and uncertain nature of all forms
of writing has provided a core of our philosophical debates for many decades.
Does the writing of algorithms really pose a problem for which we do not al-
ready have a rich vein of ethicopolitical resource?
The methods of deconstruction, of course, have never been confined to
 the writing of “text” in the form of a book, a script, or another body of writ-
ten work. If, as Derrida proposes, there is nothing outside the text — “Il n’y a
pas de hors-­texte” — then a deconstructive method is open to a broad sense of
how differential traces produce meaning: “The concept of text I propose is
limited neither to the graphic, nor to the book, nor even to discourse. What I
call ‘text’ implies all the structures called ‘real,’ ‘economic,’ ‘historical,’ socio-­
institutional, in short: all possible referents. . . . [E]very referent and all reality
104 has the structure of a differential trace, and one cannot refer to this ‘real’ ex-
cept in an interpretative experience.”62
Chapter Three

While I acknowledge that my many computer science interlocutors have

been keen to tell me that an algorithm could never be a text, I consider the
claim that the algorithm exists outside text to itself exercise a form of autho-
rial power. Algorithms do certainly arrange differentially signifying traces to
generate an output. Though the output may appear to have been generated di-
rectly, even autonomously, by a signifying code, those signs are also traces that
are written into being to make the output present and presentable. One could
seek out the source code of a deep neural net used in criminal justice, for exam-
ple, but this code is categorically not the source from which the decisional out-
puts flow. As Derrida reflects on J. L. Austin’s speech act theory, “He attempts
to justify the preference he has shown in the analysis of performatives for the
forms of the first person, the present indicative, the active voice.”63 The justi-
fication that Derrida finds so disquieting in Austin is the “source” of the first
person from which the speech act is made. “This notion of source,” Derrida
suggests, is secured by “the presence of the author as a person who writes.”64
One could not locate a source, or a source code, since writing does not involve
merely the transference of meaning from the author who writes to the world.
Understood in these terms, the algorithm used for genotyping or for risk
targeting does not simply transfer the meaning contained within the source
code into the world, for it has to write that into being. The deconstructive
method offers a different kind of engagement with the text, one in which writ-
ing does not convey the message contained in the source but rather structures
what can be said. To consider algorithms to be arrangements of differential
traces (qua texts) is to open a different way to imagine the ethicopolitical con-
texts of the algorithm. Derrida speaks of the “non closure” and the “irreduc-
ible opening” of context, in which it becomes possible to find a “moment of
strategies, of rhetorics, of ethics, and of politics.”65 Writing always exceeds its
context as it travels, iterating and entering into new relations and new con-
texts. To conceive of algorithms in terms of the nonclosure of context is to re-
sist rather directly the algorithm’s determination to reduce the output signal
to a single probability, to close the context. In the concluding pages, I sketch
some of the parameters of an ethicopolitical engagement with the writing of
algorithms.

The Unknowable Exerts Its Force

There are three broad parameters of the ethicopolitics of writing algorithms

that I wish to draw together here. First, the nonclosure of context and the at-
tention to how differential traces are arranged makes it possible to see more 105
clearly how marginal or minor elements of a text might be functioning. Where

The Uncertain Author

one concept or meaning appears to dominate attention at the expense of a
marginalized concept, a deconstructive reading is attuned to ask whether this
minor concept might, in fact, be a crucial condition of possibility for the domi-
nant meaning. Let me make this a little more concrete in terms of the margin-
alized or minor elements of the writing of algorithms. Consider, for example,
the deep neural network whose sensitivity thresholds mark the boundaries of
major/minor elements in the writing. Each layer in the algorithm transmits
on to the next hidden layer only that signal it adjudicates to be “major” or to
exceed a given threshold of probability. Though the input the algorithm re-
ceives is continuous, heterogeneous, and undifferentiated, the output signal
has to be in the range 0 – 1; that is to say, it has to be expressed as a probability.
To attend to the marginal elements of the writing of an algorithm is to return
to the layers and to consider them points of nonclosure, where the making of a
threshold of marginality is a necessary condition of the meaning of the domi-
nant output. These points of nonclosure are moments not of illumination and
transparency but of profound opacity, where what matters is the unknown,
the undecidable, and the incalculable. This is not a trifling point of philo-
sophical argument — it has real and palpable ethicopolitical implications. The
dna genotyping algorithms, for example, are reopened as text that is never
completed and that always overflows its source code, so that what comes to
matter is how the minor or marginalized data remain lodged within the layers
of the algorithm. How is the likelihood ratio output of the algorithm trans-
formed by moving the threshold between major and minor data elements? A
strategy of sorts.
Second, the notion that there is nothing outside the text reorients critique
so that one is less tempted to look outside the algorithm, to step outside its
formation to find ethicopolitical grounds. Let us stay with the text, with the
difficulties and undecidabilities of the text as such. A deconstructive strategy
points also to how a single concept or word within the text contains within it
different and irresolvable points of tension and argumentation. In the writing
 of algorithms, these concepts function simultaneously to advance the power
of the algorithm to automate decisions and to interrupt and subvert them.
Consider, for example, the use of error and bias in the writing of algorithms as
models of decision. In the most evident public and political discussions of algo-
rithms these concepts appear as sources of injustice and wrongdoing that must
be removed to fix the algorithm’s writing of the world. These same concepts,
however, also have a quite different force that works to interrupt the idea that
106 an algorithm could ever be something one might call unbiased or neutral. In
short, the writing of the algorithm contains both the promise of an unbiased
Chapter Three

objectivity and the productive function of a bias that is required for the mate-
rial to be worked on. In the writing of deep neural nets for speech recognition,
for example, the algorithms “trained by back propagating error derivatives”
are reported to have “the potential to learn much better models of data.”66 In
this writing of the algorithm, weights and biases modulate the input signal to
produce outputs as “guesses,” where a feedback loop “rewards weights that sup-
port its correct guesses, and punishes weights that lead it to err.”67 The error
derivative, just like the likelihood ratio in genotyping algorithms, is a genera-
tive fabulation that actively writes the world into being.
Finally, the differential traces of a text mean that it always already con-
tains differences from itself. Thus, when patterns or styles are identified as
being defining characteristics of a text — as with the algorithms predicting the
next sentences of dead authors — those characteristics are undone by other
signs carried within the same text. As Simon Critchley notes in his account
of the ethical capacities of deconstruction, “Any text which identifies truth
with presence as logos” will necessarily also enter into “contradiction with it-
self,” signaling “towards a thinking that would be other to logocentrism.”68 For
Critchley, this “hinge” between the closure of the logos and the opening onto
the other is the very articulation of an ethics of deconstruction. Any demand
for an enclosure of the context of writing, as for example in the many calls for
the context of algorithmic systems to be defined and circumscribed, neglects
this sense of the excess of context.69 The idea that one could govern and limit
the contextual application of a specific algorithm, for example, denies the con-
stant iteration, modification, and adjustment of the writing of algorithms. In-
deed, the algorithms used for extracting meaning from text, such as sentiment
analysis, write themselves in and through decontextualized encounters, pre-
cisely by being exposed to something of the excess of context.
Following the writing of the algorithm involves watching for the moments
when it enters into contradiction with itself. Computer science envisages the
truth of the algorithm in presence or logos, where the algorithm adjusts itself
to map signal to meaning correctly. In the writing, however, there are no de-
finitive correlations but only continuous and open-­ended iterations of the text,
where every modification authorizes a new meaning and a new output. Indeed,
this is what the idea of a likelihood ratio means in an algorithm used to deter-
mine the probability of the presence of someone’s dna at a crime scene. An
infinitesimal modification in the training data (remembering that this could
be exposure of the learning algorithm to the dna data of racialized subpopu-
lations), for example, will alter the meaning of the likelihood. At root, the 107
output of the algorithm is not secured by the linear enunciation of its code.

The Uncertain Author

The person will appear differently as a suspect; another will be differently rec-
ognized as an immigration risk; another will become the optimal target for
recommendations to vote or to shop: all of this because the iterative writing
of the text gives rise to yet another set of inferred associations. Let us consider
the writing of the algorithm, then, as though it were akin to Hilary Mantel’s
sense of fiction as showing the unknowable exerting its force.
4
The Madness of Algorithms
Aberration and Unreasonable Acts

There are crises of reason in strange complicity with what the world calls crises
of madness.
 — Jacques Derrida, “Cogito and the History of Madness”

Reason demanded this step. In our science we have run up against the limits of
what is knowable. . . . Our science has become a horror, our research dangerous,
our knowledge lethal. All that is left for us physicists is to capitulate before reality.
— Friedrich Dürrenmatt, The Physicists

Errant Algorithms

There are moments when the everyday background hum of algorithms in so-
ciety seems to break the surface and become something louder and more in-
sistent. Such moments are often debated publicly as times when algorithms
have become mad, or when they have departed from their rational logics and
become frenzied, demented, and thoroughly unreasonable. In October 2016,
when the British pound lost 6.1 percent of its value in a matter of seconds,
the currency trades conducted autonomously by algorithms were reported
to have precipitated a “flash crash,” an eruption of irrational actions beyond
the control of human traders.1 Such apparent moments of the madness of al-
gorithms, however, are not limited to the vicissitudes of the financial mar-
kets. In March 2016, Microsoft researchers experimented with a social media
chatbot — named Tay and modeled on the attributes of a teenage girl — training
her to communicate in a humanlike way by exposing her to the inputs of Twit-
ter data streams. Within a few short hours, and in response to her Twitter
learning milieu, Tay was posting racist and sexist messages, denying the holo-
caust and boasting to the police of her prolific drug habit.2 Tay’s deep learning
algorithms were widely reported to be in meltdown, spiraling out of control
in a descent into madness. The Microsoft experiment was finally halted when
Tay messaged her two hundred thousand social media followers with “You are
too fast, please take a rest,” sending the message over and over again in a rep-
etitious frenzy. The public’s sense of dangerous algorithms becoming unrea-
sonable and doing harm seemed to reach fever pitch in 2017, when YouTube’s 109
machine learning recommendation algorithms targeted automated content to

The Madness of Algorithms

children, playing videos with violent content on the YouTube Kids platform,
such as the popular cartoon character Peppa Pig being tortured at the den-
tist, drinking bleach, and eating bacon.3 In each of these moments, when the
harms and dangers of algorithms seem to press themselves into public atten-
tion, what is most often discussed is a kind of algorithmic madness, a frenzied
departure from reason.
As societies have responded to the errancy of algorithms, a specific ethico-
political framing of the problem has emerged over time. Overwhelmingly, the
public parlance around algorithmic ethics has placed emphasis on limiting the
excesses of algorithms, controlling their impulses, and reining in their capac-
ity for frenzied action in the world. The search for an encoded ethics for algo-
rithms has tended to annex madness as an aberration that must be subjected
to correction. In annexing the madness of algorithms as irrational aberration
and departure from reason, such public interventions sustain the promise that
algorithms could be rendered governable by appropriate thresholds of reason-
able and unreasonable actions. Popular debates have depicted algorithms as
“weapons of math destruction” that must be “taught right from wrong” in a
world where one must “take care that ethical and not evil algorithms help win
elections.”4 The propensity of the algorithm for madness, in short, has become
a kind of limit point on its otherwise apparently reasonable place in our world.
They can stay if they behave themselves. A kind of bounded rationality pre-
vails, in which the boundary delineating good from evil in algorithmic deci-
sions is one marked by reason versus unreason, rationality versus irrationality.5
This boundary is breached, so it appears at least, when the algorithm acts in
a way that was somehow not anticipated. Thus, for example, when the “kill
switch” of a human intervention is activated — whether this is in an autono-
mous weapons system, artificial intelligence for speech, or a video recommen-
dation system for children — this marks a threshold of something like madness,
understood as a departure from reasoned logic.6
In this chapter I am concerned with how this common notion of the mad-
 ness of algorithms as aberration has crucially overlooked the extent to which
the rationality of algorithms is built on — indeed positively embraces and har-
nesses — the power of unreason. Such accounts of moral panic amid the mad-
ness of algorithms have had two significant effects. First, they have forgotten
the place of notions of madness within histories of algorithmic rationality,
where unreason is intrinsic to the computational logic. And second, they have
underplayed the role of capricious incalculability within our twenty-­fi rst-­
110 century modes of algorithmic decision. Put simply, the common refrain of an
encoded ethics for algorithms depends on a threshold that divides rational ac-
Chapter Four

tion from madness. The public discussion of the evils and harms of algorithms
is focused on the regulation of that very threshold thought to separate rational
from irrational action. Crucially, though, this threshold is, as all border lines
are, the line that both pulls together and demarcates madness from all that is
reasonable in the world. For Jacques Derrida, the division is a “dissension” or a
“self dividing action,” which both “links and separates reason and madness.”7
Discussing Foucault’s treatment of the history of madness, Derrida cautions
that “one cannot speak of madness except in relation to ‘that other form of
madness,’ that is except in relation to reason.”8
Taking seriously the conjoined histories of the ideas of reason and mad-
ness, I propose that one cannot speak of the madness of the algorithm except
in relation to the form of reason the algorithm embodies. While the contem-
porary moral panic at each moment of the madness of algorithms urges us to
police ever more vigilantly the line between reasonable and unreasonable ac-
tions, understood as a dissension, this line is precisely the condition of possi-
bility of algorithmic rationality. Algorithms cannot be controlled via a limit
point at the threshold of madness because the essence of their logic is to gen-
erate that threshold, to adapt and to modulate it over time. In short, my argu-
ment is that when algorithms appear to cross a threshold into madness, they
in fact exhibit significant qualities of their form of rationality. Understood in
this way, the appearance of a moment of madness is a valuable instant for ethi-
copolitics because this is a moment when algorithms give accounts of them-
selves. Contra the notion that transparency and the opening of the technologi-
cal black box secures the good behavior of algorithms, the opacity and clouded
action exhibited in the excesses and frenzies of algorithms yield a different
kind of fidelity to the logic.
Let me make this concept of clouded action a little more concrete. Though
the surfacing of violent images of beloved cartoon characters for preschool
children appears as an undoubted algorithmic aberration, the computer sci-
ence accounts of the development of deep neural networks for the YouTube
video recommendation system exhibit a rationality that is consistent with out-
putting violent content. The algorithm designers describe the “dynamic cor-
pus” of YouTube video content and explain their improved deep learning algo-
rithms, which are “responsive enough to model newly uploaded content as well
as the latest actions taken by the user.”9 Understood as a model that is trained
on one billion parameters, the apparent madness of the recommendation of
violent cartoons is explicitly part of, and not an aberration from, the rational-
ity of the deep neural network. Indeed, the optimization of these algorithms is 111
achieved in part through exposure to the new and emerging uploaded content,

The Madness of Algorithms

which is afforded greater weight in the model than user-­specific data histories.
At the level of the algorithm’s logic — its mode of learning, internal weights,
architecture, parameters, training data, and so on — the funneling of a frenzy
of violent videos is entirely consistent with its rationality.
As I propose throughout this book, when viewed from the specific propo-
sitional arrangements of the algorithm, particular actions that might appear
as errors or aberrations are in fact integral to the algorithm’s form of being and
intrinsic to its experimental and generative capacities. I am advocating that
we think of algorithms as capable of generating unspeakable things precisely
because they are geared to profit from uncertainty, or to output something
that had not been spoken or anticipated. As machine learning algorithms in-
crease their capacity to learn from raw unlabeled data streams, the unseen
and unspoken become precisely the generative materials of algorithmic deci-
sions. Such proximity between violent algorithmic outputs and the attributes
of the data stream does mean that one has to think some heretical thoughts
on ethics. The racist hate and misogyny of Tay; the funneling of far-­right me-
dia to specific voter clusters using deep neural networks; the terrible conse-
quences of image misrecognition in a drone strike — these are not instances
when the algorithm has become crazed or frenzied. Rather, the learned ac-
tion is a reasonable output given the extracted features of the data inputs. Of
course, this is not a less troubling situation than the one in which some con-
trols are sought on the worst excesses of the algorithm. On the contrary, it
is all the more political, and all the more difficult, because that which could
never be controlled — change, wagers, impulses, inference, intuition — becomes
integral to the mode of reasoning. To be clear, my intention is not to avoid the
need for critical response to the harms of algorithms. Far from it, my point is
that violence and harm are not something that can be corrected out of an oth-
erwise reasonable calculus.
In what follows in this chapter, I begin by resituating the genealogy of
madness and reason in twentieth-­century cybernetics, focusing specifically
 on the work of Norbert Wiener and the double enrollment of madness and
rational forms of control into algorithmic systems. I then flesh out an alter-
native way of thinking about the madness of algorithms, reformulating the
problem from one where algorithms might lose their hold on rationality to one
where algorithms precisely require forms of unreason to function and to act.
I am thereby amplifying attention to another kind of harm that does not re-
side beyond the threshold of something like the edge of reasoned action. This
112 different kind of harm dwells in the algorithm’s inability to ever embody the
“madness of decision.”10 To live with the madness of decision is to acknowledge
Chapter Four

and take responsibility for the impossibility of ever binding the action to a full
body of knowledge. Decisions are mad because they can never know fully the
consequences and effects of their own making. To decide is to confront the im-
possibility of the resolution of difficulty; it is madness in the specific sense that
it has no unified grounds.11 With contemporary algorithms, decisions are being
made at the limit of what could be known, and yet there is no responsibility
for the unknown consequences of the decision. The madness of the decision is
disavowed by the single output of the algorithm, and this disavowal is a poten-
tial horror and a danger. In the second half of the chapter, I discuss one specific
set of algorithms — random decision forests — that have become the algorithms
of choice across many domains, particularly in national security and border
controls. The random decision forest algorithm grows multiple decision trees
on random subsets of data and “takes a vote on their predictions.”12 It is an al-
gorithmic proposition of What comes next? that takes place as a calculation
amid incalculability, mobilizing chance and the splitting of agency, sometimes
with lethal effects on human life.

Cybernetics and Unreason

In a 1950 essay, “Atomic Knowledge of Good and Evil,” the cybernetician Nor-
bert Wiener expresses his concern for the moral responsibility of science in
the face of what he calls the “dangerous possibilities” of atomic weapons.13
Reflecting on the nature of the roles of the mathematicians and physicists at
Los Alamos, Wiener is anxious that “the new centralized atmosphere of the
age to come is likely to lead society in the direction of ‘Big Brother’ ” and to-
ward a “future fascism.” A full seventy years before scientists pointed to the
dangers of a descent into fascism precipitated by the mathematical sciences of
algorithms, the cyberneticians were writing publicly in newspapers and jour-
nals to express the view that “when the scientist becomes the arbiter of life and
death” the locus of “moral responsibility” belongs properly to the reasoning
scientist to prevent “the decay and disruption of society” and to ensure that
science “is used for the best interests of humanity.”14 Indeed, in a 1947 letter
to Atlantic Monthly, Wiener replies publicly to a request from government mili-
tary officials for copies of his mathematics papers. Concerned that withdraw-
ing his science from the service of war might be “shutting the door after the
horse has become classified,” Wiener believes that the locus of control for a
dangerous science must reside in the reasoning human scientist. For Wiener,
the rise of computing machines posed a threat defined not so much in terms
of the madness of the autonomous machine as in terms of what he called the 113
“crazed and misguided” irrationality of humans. “The automatic machine is

The Madness of Algorithms

not frightening because of any danger that it may achieve autonomous control
over humanity,” writes Wiener, but “its real danger” is that it may be “used
by a human being to increase their control over the rest of the human race.”15
Wiener’s letters and essays are replete with his anxiety that the science driv-
ing human progress may simultaneously be enrolled by sovereign or corporate
powers and deployed for violence and war. For Wiener, the moral responsi-
bility of the mathematical and physical sciences, particularly in their service
of state power, resides in governing the line between reason and unreason.
This distinction between reason and unreason, however, exposes something
of the paradox of the place of madness in the cybernetic histories of algorith-
mic rationality.
First, the rise of cybernetics is closely intertwined with notions of over-
coming the dangerous fallibilities of human judgment. Reimagining the hu-
man decision as a series of relays that could be modeled for optimization, Cold
War cybernetics was geared to “tame the terrors of decisions too consequential
to be left to human reason alone.”16 Understood in this way, the definiteness
and conclusiveness of the algorithm as decision maker was thought to supply
a rational safeguard against the madness of mutually assured destruction.17 As
Wiener wrote in his essay on “the new concept of the machine,” the binary
logic of computation structured decisions via a series of “switches,” each offer-
ing a “decision between two alternatives” in a “logical machine.”18 For Wiener,
the binary branching of the switch embodied what he called “two truth val-
ues,” where the transmitted information was either “true or false.” Here the
logic of algorithms explicitly condenses the indefiniteness of data to a series of
switches that must always be a binary choice of yes/no, true/false.
The branching that we see in classic decision tree algorithms, and in con-
temporary random forest algorithms, carries the traces of a cybernetic his-
tory, when the switch marked the calculation of truth and falsity. Amid the
terrors of thermonuclear war, the promise of branching algorithmic decisions
was that they could supply to political decision makers a means of avoiding the
 human propensity for frenzied miscalculation. At every instant of a branch-
ing decision, the logical machine rendered an output of truth or falsity. In
this way, the twentieth-­century rise of rational choice in foreign policy and
behaviorism in the social sciences imagined a world in which algorithms lim-
ited the madness of a geopolitics on the edge of destruction. This first aspect
of the place of unreason in genealogies of algorithmic decisions is significant
to our contemporary moment because it is a reminder that algorithms are not
114 devices that exist outside notions of the normative, norms, anomalies, and pa-
thologies. On the contrary, the rise of algorithms in governing difficult and in-
Chapter Four

tractable state decisions makes them interior to defining what normalities and
pathologies could be in a society. Just as the early neural network algorithms
were embraced as a means of instilling reason into otherwise potentially dan-
gerous human judgment, however, they also brought into being a new mode of
harnessing the unknown and the unpredictable.
Second, at the same time as algorithms were thought to tame human irra-
tionality, a specific orientation to madness was actively incorporated into the
logics of algorithms. In Orit Halpern’s book Beautiful Data, she gives a devas-
tating account of how ideas about the psychotic and the neurotic were inte-
gral to cybernetic logics. “What has been erased from the historical record in
our present,” writes Halpern, “is the explicit recognition in the aftermath of
World War II and the start of the Cold War that rationality was not reason-
able.”19 In a detailed engagement with Warren McCulloch and Walter Pitts’s
development of the neural net, Halpern proposes that the logic of psychosis
was crucial to the reformulation of computational rationality. Tracing the
genealogy of neural nets as technical instantiations of the brain’s functions,
Halpern demonstrates the extent to which “neurotic, psychotic, and schizoid
circuits proliferated in the diagrams of cybernetics.”20
The insights of those who have studied the schizoid and paranoid relations
of computation are critical to the contemporary ethicopolitical relations being
forged with and through algorithms.21 Far from representing a departure from
rationality and entry into psychotic turmoil, the actions of algorithms are
never far from their conjoined histories with psychosis, neurosis, trauma, and
the imagination of the brain as a system. The double enfolding of unreason
into cybernetic thought — as the unreasonableness of human decision makers
in conditions of war and insecurity, and as models of psychosis underpinning
the segmented cognition of the neural net — serves as something of a correc-
tive to our contemporary societal debates on the responsibility of science for
moments of algorithmic madness. Put simply, the history of algorithmic ratio-
nality is not separable from genealogies of madness. It is time to think difficult
thoughts: algorithms are always already unreasonable. When one hears calls
for new controls to limit the potentially dangerous actions of the algorithm, it
is perhaps worth recalling that algorithms historically are limit devices. That
is to say, they actively generate new forms of what it means to be normal or
abnormal, just as they mark new boundaries of rationality and unreason. To
capitalize on uncertainty — whether in warfare or in the commercial world of
risk — algorithms dwell productively with emergent phenomena and incalcula-
ble surprises.22 The actions of the contemporary neural net, for example, finds 115
its condition of possibility in the twinned imperatives of limiting the impulses

The Madness of Algorithms

of human decision makers while embracing and harnessing the impulses of
models of the brain as nervous circuits.
To be clear, contemporary algorithms are oriented toward limiting the un-
anticipated actions of humans while generating a whole new world of their
own unanticipated and unreasonable actions. This means, for example, that
an algorithm will decide the threshold of whether a person’s presence on a city
street is normal or anomalous, and yet it will generate this threshold via ex-
perimental methods that enroll errancy and capricious action into the form of
reason. In the section that follows, I revisit the problem of madness to refor-
mulate the algorithm’s relation to unreason.

Madness and Classification

Throughout this book I urge some caution about the treatment of the ethical
and political questions posed by algorithms as novel or unprecedented. Indeed,
I suggest that the advent of algorithms as decision makers in our societies has
restated what are in fact perennial philosophical debates about the nature of
ethical and political life. Among these perennial debates is the question of
whether and how one controls the impulses or actions of an entity that appears
to us as beyond human reason. It is in the philosophies of human madness and
its relationship to forms of reason, I propose, that one can locate a resource
for thinking differently about the actions of algorithms. In Michel Foucault’s
major genealogical work on the history of madness, he argues that madness is
designated as a “blatant aberration” against all that would otherwise be “rea-
sonable, ordered and morally wise.”23 Understood in this way, the very idea of
human reason has required — integral to its historical condition of possibility
within a moral order — an opposition to the aberration of madness. Indeed, for
Foucault, “in our culture there can be no reason without madness,” and yet
the objective science that observes, records, and treats pathologies of madness
simultaneously “reduces it and disarms it by lending it the slender status of
pathological accident.”24 Thus, there is a necessity of madness within Western
 thought, and yet this is a diminished form of madness reduced to a pathologi-
cal accident or error. The idea of unreason is always already present in claims
to moral reason, but this is an unreason understood as a flaw or error in the
logic of reason — a flaw that can be identified, classified, and cured.
Such a sense of a reduced form of madness within Western moral order
seems to me to be amplified in our contemporary moment, when the actions
of algorithms are so frequently named as accidents or errors. The threshold of
116 madness denotes a boundary between the reasonable and the unreasonable,
just as it also condenses the unknowability of madness to an error, an accident,
Chapter Four

or a flaw in the code. Significantly, Foucault’s history of madness traces what

he calls “a vast movement” from critical to analytical recognition of madness,
where classical forms were “open to all that was infinitely reversible” between
reason and madness, while the nineteenth and twentieth centuries brought
the “classificatory mania” of psychiatry and insisted on new lines demarcating
the rational from the mad.25 The opposition of madness and reason, then, is a
historically contingent event that occludes and reduces a more expansive sense
of the human experience of madness.26 As Jacques Derrida describes it, “The
issue is to reach the point at which the dialogue” between reason and madness
is “broken off” and to permit once more a “free circulation and exchange” be-
tween reason and unreason.27 The modern break in the dialogue between rea-
son and madness, described by Derrida as a “dissension,” makes it necessary,
he proposes, to “exhume the ground upon which the decisive act linking and
separating madness and reason obscurely took root.”28
The dissension linking and separating madness and reason is a relation
rather different in character from the limit point of a madness as aberration
from reason. Rather, the dissension itself makes possible new ways for society
to understand and to govern the relation between reason and madness. It is an
ethicopolitical dissension that generates not only new models of what is nor-
mal and abnormal in human societies, but crucially also new ethical relations
between selves and others.29 The relationship between madness, science, and
law from the nineteenth century became a moral force in which the rational-
ity of the psychological sciences would assert the boundaries of the normal
and the pathological. With the advent of positive psychiatry, as Foucault de-
scribes, a “new relation” became possible between the condition of madness
and “those who identified it, guarded and judged it” within the “order of an ob-
jective gaze.”30 Hence, the dissension forges ethical relations that render mad-
ness knowable “at a stroke,” authorizing “reasonable men to judge and divide
up different kinds of madness according to the new forms of morality,” just as
it simultaneously tames the nonknowledge of the experience of madness.31 In
the making of a particular ethical and moral relation to madness, Foucault lo-
cates a first major objectification of the human:

From this point onwards, madness was something other than an object
to be feared. . . . It became an object. But one with a quite singular status.
In the very movement that objectified it, it became the first objectifying
form, and the means by which man could have an objective hold on him-
self. In earlier times, madness had signified a vertiginous moment of daz-
117
zlement, the instant in which, being too bright, a light began to darken.
Now that it was a thing exposed to knowledge . . . it operated as a great

The Madness of Algorithms

structure of transparency. This is not to say that knowledge entirely clari-
fied it, but that starting from madness a man could become entirely trans-
parent to scientific investigation.32

To make an object of madness, then, was to make a world of transparency

in which an individual could have a hold on himself. Let us consider this no-
tion in light of our contemporary moment, when once more madness is be-
coming object in terms of both the possibilities of rational objective algorith-
mic decisions and the demand for algorithms to take a hold of themselves and
to be rendered transparent to investigation. The moment of dazzlement and
darkening Foucault describes, when the place of madness within reason is ac-
knowledged, is entirely absent in the contemporary calls for transparency and
the opening of the black box of the algorithm. I wish to reopen the dissension
in madness and reason that Derrida and Foucault differently depict and make
of this dissension a different kind of ethical terrain. With dissension conjoin-
ing madness and reason, the ethical move is no longer a matter of the search
for moral codes that regulate the boundary between rational and unreasonable
behavior by algorithms. Instead, the manifestation of a moment of apparent
madness in an algorithm’s actions becomes necessarily also a moment when
the algorithm gives an account of its form of reason. Rather than contribute
to the cacophony of calls for greater transparency, this ethical demand dwells
with the vertiginous moment of dazzlement and darkness, with clouded ac-
tion and opacity, so that the violences of an algorithm cannot simply be mis-
takes or errors in the algorithm’s otherwise logical rationality.
When one no longer seeks transparency and a hold on oneself, the appear-
ance of a moment of algorithmic madness offers a different kind of insight — an
insight into how the algorithm enrolls and deploys ideas of unreason to func-
tion and to act. In the moment of madness, we see a dazzling instance of the
form of rationality’s improbability. In turn, the question of responsibility also
shifts ground. Reflecting on the rise of positive psychiatry, Foucault notes that
 unreason retained a moral dimension, in which “madness was still haunted by
an ethical view of unreason, and the scandal of its animal nature.”33 The re-
sponsible subject, then, would act to annex the animal from the human and
to locate ethics in the unified mind. In his discussion of homicidal mania and
responsibility for murder, for example, Foucault notes that for the defendant
to be responsible for the act, there had to be “continuity between him and his
gesture.”34
118 To be mad in the sense of being outside oneself was to be caught and
“alienated,” divided within oneself so that a person “was himself and some-
Chapter Four

thing other than himself.”35 This dividuated subject is precisely embraced by

the actions of the algorithm so that its posthuman form is to be simultaneously
human and something other than human, a form of self, and fragments of the
other. So, responsibility cannot feasibly take the form of taking a hold on algo-
rithms or annexing instinctive or impulsive animal behaviors. I want to take
seriously the idea that “in the absence of a fixed point of reference, madness
could equally be reason,” so that there can be discontinuity between an algo-
rithm and its actions while a responsibility still remains.36
To reverse the opposition of madness and reason would be to think radi-
cally differently about the many moments when algorithms have been said to
err or to deviate from accepted norms. The logic of algorithmic errancy ex-
presses the responsibility of the algorithm precisely in the terms of there be-
ing continuity between the algorithm and the gesture. When the algorithm
strays or deviates, it is considered irresponsible because it no longer controls
the outputs it effects. And yet, as I argue in previous chapters, the experimen-
tal straying from paths defines much of the power of contemporary machine
learning. A kind of gestural discontinuity, in short, is profoundly useful to the
algorithm. To reduce madness to error or aberration is to “neutralize madness”
and to shelter “the Cogito and everything related to the intellect and reason
from madness.”37 The errors and aberrations of algorithms continue to domi-
nate public discussion of the ethics of machine learning, automated decisions,
and data analytics. This dominance of errors and aberrations is indeed neu-
tralizing madness in its broadest sense — as that which cannot be fully known
or spoken — and sheltering the rationality of algorithms from a full and expan-
sive critique. The illegible, unspeakable, and opaque actions of algorithms, as
I argue throughout this book, are not the limit points of what is possible ethi-
cally; instead they are the starting points of an ethical demand.
When the effects of algorithmic decisions are truly horrifying — such as in
Cambridge Analytica’s rendering of the attributes of “persuadable” voters in
the US presidential election and the UK EU referendum (with its use of the at-
tribute to target xenophobic anti-­immigration media, for example) — reining in
their crazed excesses can never be sufficient. The unreason and the excess are
not aberrations at all but are the condition of possibility of action. The prob-
lem seems to be, then, not that the rational and the reasonable algorithm takes
leave of its senses, loses control, and loosens its hold on its logic. Rather, the
algorithm is always already beside itself and divided within itself as such.38 Al-
gorithms learn by unrestrained experimentation and emergent signals. They
simultaneously send multiple conflicting signals along different pathways to 119
optimize their output. They are errant not in the strict sense of deviating from

The Madness of Algorithms

a path, but in the archaic literary sense of traveling in search of adventure.
And so, algorithms must also be understood differently in their actions — not
as entities whose propensity for madness can be tamed with correct diagnosis
and repair, but instead as entities whose particular form of experimental and
adventurous rationality incorporates unreason in an intractable and produc-
tive knot.
I examine this generative mode of unreason as it animates the decisions of
algorithms dealing with human life. The violences and injustices that result
from the algorithm’s decisions do not emerge primarily from errors, accidents,
or aberrations in the system’s logic. Such a framing of the violent outputs as
errors shelters the cogito from the darkness of unreason. It also neutralizes
the fullness of madness as improbability and the unknowable, thus restricting
what can count as the harms of algorithms. A principal harm of algorithms is
that they enable calculative action where there is incalculability and the un-
knowable, reducing ethicopolitical orientations to the optimization of outputs
and the resolution of difficulties. Renewed attention to the unreason within
the algorithm animates how combinations of probabilities generate improb-
able and untamed outputs that are let loose into the world.

The Madness of Decision

In October 2017, the US National Science Foundation (nsf) awarded a

$556,650 research grant to a team of engineers and philosophers working on
the decision-­making algorithms for autonomous vehicles.39 The research repre-
sents a rather direct example of what I have termed an encoded ethics, in which
codes of conduct are sought to modify and restrain the harmful effects of al-
gorithmic decisions. The research addresses the classic moral philosophy “trol-
ley problem,” redefined for the age of algorithms. This is a scenario in which
a trolley car is careering down the tracks toward a group of five people. The
driver faces a profoundly difficult decision — to continue on the track toward
the certain deaths of the five people or to pull the lever, thereby intentionally
 and fatally redirecting the trolley onto a second track where one person lies
immobilized. In moral philosophy, the trolley problem is intended to highlight
the problem of the grounds of rational decision making. How should the driver
weight the value of the lives on the tracks? Is this calculation morally question-
able? Should she make an active decision to kill one person to avoid the passive
accidental deaths of five? In the nsf-­f unded research, the trolley problem is
reinterpreted for decisions made by algorithms guiding autonomous vehicles.
120 “You could program a car to minimize the number of deaths or life-­years lost
in any situation,” explain the researchers, “but then something counterintui-
Chapter Four

tive happens: when there’s a choice between a two-­person car and you alone
in your self-­driving car, the result would be to run you off the road.”40 The re-
search imagines a world in which algorithms can be trained to precompute the
rational decision in the face of a scenario of catastrophic consequences.
The boundary of this rationality is a computational threshold in which
the algorithms weigh the values of different choices before making a deci-
sion. The research team is working on deep neural networks that will “sort
through thousands of possible scenarios,” filtering out and “rapidly discard-
ing 99.9% of them to arrive at a solution.” The example the scientists discuss
is a self-­driving car hurtling toward a school bus, where the optimal pathway
is thought to be “discarding all options that would harm its own passenger”
before “sorting through the remaining options to find one that causes least
harm to the school bus and its occupants.”41 Here, in the starkest of terms, is
a statement of the kind of ethical codes sought by many engineers and algo-
rithm designers.42 It is an ethical mode that promises to render calculable the
profoundly uncertain horizon of an immediate future. Put simply, in seeking
to ward off the full horrors of frenzied algorithms governing a trolley out of
control, such programs generate new harms in the valuing and weighting
of different pathways. The calculus itself — with its combinatorial possibilities
of babies in strollers, elderly pedestrians crossing the road, careless delivery
drivers, and inattentive cyclists — is always already thoroughly unreasonable.
Algorithms do not pose their most serious and harmful threats when they are
out of control, become crazed, or depart rationality. On the contrary, among
the most significant harms of contemporary decision-­making algorithms is
that they deny and disavow the madness that haunts all decisions. To be re-
sponsible, a decision must be made in recognition that its full effects and con-
sequences cannot be known in advance. A responsible decision-­making pro-
cess could never simply “sort through the remaining options to find one that
causes the least harm” because this is an economy of harms that renders all the
incalculability of harm calculable.
 The madness of the trolley decision is that it must necessarily be made
in the darkness of nonknowledge, that it categorically is not subject to pre-
programming to optimize an outcome. As Derrida writes, “Saying that a re-
sponsible decision must be taken on the basis of knowledge seems to define
the condition of possibility of responsibility,” and yet at the same time, “if
decision-­making is relegated to a knowledge that it is content to follow, then it
is no more a responsible decision, it is the technical deployment of a cognitive
apparatus, the simple mechanistic deployment of a theorem.”43 Though the al- 121
gorithm’s design for the trolley problem may appear to be the simple mechanis-

The Madness of Algorithms

tic deployment of a theorem, it does contain the aporia of multiple decisions.
The problem is that these multiple, fully ethicopolitical decisions are them-
selves annexed and sheltered from madness by means of errancy or mistake.
One route into the fuller ethicopolitics of an algorithmic decision is to at-
tend to its vast multiplicity. Thus, as I explain in chapter 2, the autonomous
vehicle is unable to recognize “child,” “stroller,” or “school bus” without a re-
gime of recognition trained on millions of parameters of labeled images. The
computer scientists working on encoded ethics for the trolley problem (at the
heart of the ethicopolitics of all autonomous vehicles) seriously underestimate
the difficulty when they speak of filtering “thousands” of possible scenarios,
for each of these scenarios is nested within millions of other parameters. An
algorithm cannot be programmed to value, for example, the life of a child or
an adult as such, for even this decision contains within it the multiple fraught
difficulties of learning how to recognize or misrecognize a person. Where
the introduction of moral philosophy into engineering problems has sought
to contain the algorithm’s capacity for an irrational output, the algorithm in
fact positively embraces and requires irregularities, chance encounters, even
the apparently errant past actions of humans to learn how to optimize the
output. The common experimental method of “spoofing” an algorithm in de-
velopment, for example, involves the exposure of the algorithm to a kind of
frenzy of false or “spoof ” images to teach it to refine its capacity to recognize
truth from falsity. To summarize at this point, when an autonomous vehicle
appears to have departed from its logic, lost control, or become mad, it has in
fact yielded to the world something of how its logic functions, how its deci-
sions are arrived at. Rather than seeking an encoded ethics to try to limit the
madness of algorithms, a cloud ethics should proceed from the incompleteness
and undecidability of all forms of decision. The algorithm contains within its
arrangements all the many multiples of past human and machine decisions. It
has learned how to learn based on the madness of all decision, on the basis of
nonknowledge. To have responsibility for decisions on life, as the autonomous
 vehicles’ algorithms manifestly do, is to foster conditions in which that learn-
ing requires unreasonable things. In the sections that follow, I explore two sets
of circumstances when algorithms are deciding on life in the face of an incal-
culable future.

Life and Unreason I: “Exaggerated Results”

In the summer of 2016 a group of Swedish computer scientists published their

122 findings on algorithms designed to interpret images of the brain from func-
tional magnetic resonance imaging (fMRI) scans. Their investigations into
Chapter Four

the statistical validity of the major algorithms used globally in neuroimaging

seemed to show a stark and troubling rate of error, what they called a “cluster
failure.”44 In the context of a method used for neuroimaging over a period of
twenty years, the finding that the algorithms have a 70 percent false positive
rate cast fundamental doubt on scientific knowledge used in important areas
such as Alzheimer’s research.45 A 70 percent false positive rate may appear to
be an errant departure from the statistical logics governing algorithmic infer-
ence. The neural networks did not depart from their logic, however, but gen-
erated what were described as “exaggerated results” in and through that very
logic. The deep neural networks were identifying clusters of brain activity in
70 percent of cases where there was no actual cerebral activity present. Cru-
cially, though, the algorithms had been exposed to millions of training im-
ages of brain activity to recognize the attributes of new and unseen instances.
Here, once again, there is a regime of recognition generated through the al-
gorithm’s exposure to the features of past datasets. Among the many assump-
tions dwelling within the algorithms was a set of norms about the morphol-
ogy of the brain that ultimately led to misrecognition and a high false positive
rate. In short, the proliferation of false positives, with serious consequences for
the neurological diagnoses of countless people, was not an aberration in the
algorithm’s rationality but in fact a manifestation of the excesses of its form
of reason.
When multiple false positives are produced in other domains of algorith-
mic decision, such as in facial recognition systems, often these are understood
as errors that could be corrected out.46 Yet, if one considers that to reduce mad-
ness to error is to protect reason, then the resort to error does not limit the al-
gorithm’s use in society but actually helps it to proliferate into new domains
of life. In my reformulation of the madness of algorithms, I wish to revalue
the uncertain relationship to truth that is embodied in algorithms. “Madness
begins where man’s relationship to the truth becomes cloudy and unclear,”
writes Foucault, so that to experience madness is to experience the “constant
error” of the destruction of one’s relation to truth.47 To correct the error would
be to seek a kind of transparency and clear-­sightedness that would create a
break in the cloudiness. Yet, if one stays with the cloudiness of nonknowledge,
then a different kind of partial and occluded account is demanded.
What could this different calling to account and responsibility look like?
With a cloud ethics, where one is proceeding with the partiality and opacity
of all forms of accounting for oneself and others, one would seek to show how 123
and why algorithms can never bear responsibility for undecidability, or for the

The Madness of Algorithms

madness of all decision. For example, when Pedro Domingos makes his bold
claims that machine learning algorithms are “learning to cure cancer” by ana-
lyzing “the cancer’s genome, the patient’s genome and medical history” and
simulating “the effect of a specific patient’s mutations, as well as the effect of
different combinations of drugs,” a cloud ethics would acknowledge the mul-
tiple and distributed selves and others dwelling within the calculus, signaling
the madness of the decision itself.48 The rapid increase in the use of algorithms
for genotyping and cancer treatment decisions is marked by the gathering of
sufficient data parameters to produce optimized treatment pathways as out-
puts. But, of course, in the context of health care, there is pressure to include
economies within the parameters of these algorithms, such as the cost of the
drugs, the benefit to the pharmaceutical industry, the likelihood of “quality
life years,” or the time efficiency to the overstretched clinician.49 While the re-
sponsible decision of an oncologist is made in a way that is cognizant of the un-
knowability of the outcomes of rejected treatment pathways, the algorithm’s
decision proposes the optimal pathway on the calculative basis of the weight-
ings of all possible pathways. Such processes of precomputational decision can
never bear the full weight of the madness of all decision. In effect, the precom-
puted weighting of different branching pathways we saw in the trolley problem
is mirrored in the different algorithmic pathways of cancer treatment. In both
cases — a nd there are many more — where the outcome will certainly be the
loss of life for some and the preservation of life for others, the madness of the
decision as such is disavowed. The madness of the algorithm does not reside in
the moral failure of the designer or the statistician, but it is an expression of
the forms of unreason folded into a calculative rationality, reducing the mul-
tiplicity of potentials to one output. The madness of the algorithm expresses
something of what cannot be said; it is the absence of an oeuvre, gesturing to
that which is inexplicable and unspeakable.50
 Life and Unreason II: SKYNET and the Random Forest

In the preceding section, where I discuss a 70 percent false positive rate in

neuroimaging algorithms, the matter of life and death hinged, at least in part,
on the capacity of the algorithm to recognize the singularity of the patient’s
brain as the terrain of a target. In other places where the deep neural net al-
gorithm travels, the matter of harm can similarly hinge on the recognition of
124 a target, though in ways in which even a relatively low false positive rate can
have catastrophic consequences. One specific kind of algorithm, the random
Chapter Four

forest, or random decision forest, has become the algorithm of choice for systems
designed to identify terror targets for the state.51 Why might this be? Because
machine learning for counter-­terrorism is a difficult problem. Unlike the com-
mercial algorithms for credit card fraud or email spam detection, for example,
where there is vast availability of labeled training data on which to train the
classifier algorithms, the availability of labeled data on known terrorists is ex-
tremely scarce, indeed even thoroughly inadequate for statistical modeling.
This has meant that terrorist-­targeting algorithms have become, in a rather di-
rect sense, unreasonable. That is to say, in the context of profoundly uncertain
security futures, these algorithms have harnessed the unknown and the incal-
culable to preserve the capacity to act. I am going to dwell a little longer on the
random forest algorithm as an arrangement of propositions in the world. The
randomness of this algorithm is, I suggest, a kind of sheltered madness that
dwells inside the logic of the algorithm and promises to the world an impos-
sible vision of a secure future. It is, in short, an algorithm with life and death
effects, and it generates these effects by reformulating the dissension between
reason and madness.
In October 2016, two journalists reported that the nsa’s skynet pro-
gram “may be killing thousands of innocent people.”52 In the context of the
estimated four thousand people killed by drone strikes in Pakistan between
2005 and 2016, Christian Grothoff and Jens Porup investigated how skynet
“collects metadata,” storing it “on nsa cloud servers” and then applying “ma-
chine learning to identify leads for a targeted campaign.” Among the “cloud
analytic building blocks” described within skynet in the Snowden files are
travel patterns; behavior-­based analytics, such as incoming calls to cell phones;
and other “enrichments,” such as “travel on particular days of the week,” “co-­
travelers,” and “common contacts” (figure 4.1).53
I have elsewhere written in detail on how these aggregated security “data
derivatives” are generated and how they are made actionable.54 Of specific in-
terest here are the random forest decision tree algorithms that are learning in
 125

The Madness of Algorithms

Figure 4.1  SKYNET cloud analytic. Cryptome.org.

communion with the data derivatives. Random forest algorithms were devel-
oped at the turn of the millennium by Leo Breiman of uc Berkeley’s Depart-
ment of Statistics.55 These machine learning algorithms exemplify well how
notions of chance, preference, error, and bias become incorporated into the ca-
pacity to compute an output. As Breiman explains, in the random forest tech-
nique, the algorithms come into being by “growing an ensemble of [decision]
trees and letting them vote for the most popular class.”56 Noting that “data sets
with many weak inputs are becoming more common” (for example, in medical
diagnostics) and that “this type of data is difficult for the usual classifiers — 
neural nets and trees”—Breiman designed his algorithms to “inject the right
amount of randomness” so that the algorithm is trained on random subsets of
the training data and “predicts the class of each new record using the plurality
of the class predictions from the set of trees.”57 In this way, the random forest
algorithm actively uses variance and randomness to refine its output. As Lor-
raine Daston has argued in her compelling account of the histories of prob-
ability, chance, and unreason, all three were able to flourish within methods
of calculation.58 Contemporary algorithms such as random forests have found
new ways to invite randomness and chance back into the probabilistic calculus.
 If the random forest algorithm appears to descend into the horror of tar-
geting civilians for drone strikes in the skynet program (and I have also seen
it used by governments to output a risk score of immigration infraction), how
can one understand the mode of unreason that dwells within the form of al-
gorithmic rationality? A key element of the random forest algorithm’s logic
is the attributes-­based targeting of population I discuss in previous chapters.
The skynet random forests are trained on data from a population described
126 as “known couriers.” These proxies for labeled data on terrorists are individu-
als within Pakistan who have been scraped from intelligence reports under
Chapter Four

the keyword “courier,” denoting an individual who is suspected of carrying

messages for known terror groups, such as al-­Qaeda. The random forest algo-
rithm learns the attributes of “indicative behaviors” from the training data of
couriers and then runs the data of 55 million mobile phone users in the general
population of Pakistan through this model to “discover similar selectors with
courier-­like travel patterns.”59 As the random forest algorithm is increasingly
now applied to a mobile data stream (and not only a static dataset) — for exam-
ple, in the video feed of an unmanned aerial vehicle (uav) — this discovery of
similar behavioral patterns is taking place without human input, so that the
algorithm is said to “tune its own parameters in response to the data it sees,
relieving the analyst of the need to carefully define algorithm parameters.”60
Though random forest algorithms could be said to generate a kind of mad-
ness of false positives that become actionable as a kill list, I am proposing that
in fact this madness is useful to the algorithm. Unreason supplies a refinement
of the algorithm’s logic so that it is no longer limited by the gaze of the ana-
lyst. skynet’s false positive rate is 0.008 percent, where a 50 percent “miss
rate” (the false negative rate, or where half of those with “courier-­like” pat-
terns are missed) is tolerated in the model. To signal this as error or mistake,
however, or to try to fix the problems in the training and validation data is
manifestly to overlook the place of unreason in the algorithm’s learning. The
skynet random forests generated one high-­profile false positive — recognizing
the travel patterns of the Islamabad bureau chief of Al-­Jazeera, Ahmad Zaidan,
as a high-­scoring selector — that was represented as errancy or mistake. To be
clear, according to the logic of the random forest, the recognition of Zaidan is
not errant at all but is precisely a rational selector given the proxy patterns of
the “couriers” and the class predictions of the set of trees. The classifier is per-
ceiving a scene through the apertures of the branches and leaves of the random
forest. It learns to distinguish things in the world precisely through the injec-
tion of randomness. In the security and defense domains, this capacity of the
random forest algorithm to distinguish through variability is highly valued:
“As the classifier watches events unfold, it tries to discern patterns of behav-
iour: a pack of wolves circling a wounded animal, shoppers taking items from
store shelves to a cash register, or an insurgent burying an ied on the side of
the road. It’s the algorithm’s job to learn how to distinguish between someone
just digging a hole and someone else burying a bomb.”61
For the random forest algorithm, the unreasonable and untamed wildness
of the random becomes the means of recognizing and distinguishing a target.
The random element and the incalculable are lodged within its form of rea- 127
soning. To signal the importance of this enfolded unreason is not to deny the

The Madness of Algorithms

profoundly violent effects of deploying the random forest algorithm for terror-
ism targeting. Far from it. In fact, attention to the unreasonableness of the al-
gorithm as such highlights the impossibility of confining or constraining the
madness of the algorithm. If the civilian digging a hole in Pakistan, or the
school bus at the border, is mistakenly targeted, the madness of the drone strike
is sheltered by the generation of new input data that modifies the threshold
between “false alarm” and “miss rate.” Every output of a target, however error
prone or crazed in its assumptions, supplies new input material to the model.
Given what we know about the place of psychosis and schizoid circuits in
the cybernetic history of algorithms, and about the genealogies of the relation-
ship between madness and reason, perhaps it should not be surprising that the
twenty-­first century is witnessing a reformulation of unreason within the ra-
tionality of algorithms. Nineteenth-­century diagnosis and treatment of mad-
ness and dementia reimagined madness so that it “was not an abstract loss of
reason” but instead “a contradiction in the reason that remained,” rendering
the patient’s rationality recoverable.62 In many ways this notion of a flaw in
reason that is repairable has dominated discussion of the public ethics of algo-
rithms. Yet, the modern notion of madness as an error in reason is undergoing
reformulation with the advent of decisions involving human and algorithm
collaborations. The twenty-­first-­century violences of algorithmic logics point
not to contradictions or flaws in reason that could be cured by the rational hu-
man in the loop, but to the potentialities of the excess of reason and the power
of unreason. In algorithms such as random forest, the demented wager, ran-
domness, and chance become newly reacquainted with rationality.63 To treat
the pathologies of algorithms, then, is simultaneously to engage in a “forget-
ting of violence and desires” that continues to animate the algorithms arbi-
trating the threshold between life and death in our times.64 A random forest
algorithm will never know a terrorist in the sense of acting with clear-­sighted
knowledge, but it mobilizes proxies, attaches clusters of attributes, and infers
behaviors to target and to act regardless.
 The Scaffold: Death Penalties and the Condition of Not Knowing

The madness of algorithms, as I have reinterpreted the condition in this chap-

ter, arises not from the loss of rationality or from the error-­prone deferral of
human reason into the machine, but instead from the making of a calculation
in conditions of nonknowledge.65 When an algorithm generates a singular out-
put from an incalculable multiplicity of associative relations, it shelters the
128 darkness of the decision in the reduction to one actionable output. In Jacques
Derrida’s writing on the death penalty, he describes “the imposition of calcu-
Chapter Four

lability on a condition of non-­k nowledge.” For Derrida, the specific violence

of the death penalty (even as it compares to other modes of killing) is that
an unknowable future — “the given moment of my death” — becomes calcula-
ble “with absolute precision.” “Who thus calculates,” writes Derrida, “turns
justice into a utilitarian calculation” and makes the death of a person into a
“trade, a useful transaction.”66 My argument is that contemporary algorithms
are extending this useful transaction of the programmable decision — the use-
ful output, the good enough model — into other calculations at the threshold
of life.
The weights, thresholds, and attributes through which an algorithm
comes into being are simultaneously the condition for assigning the calcula-
tive weight or the value of someone or something.67 Just as Derrida identified
in the US penal architecture that “the majority of those condemned to death
are blacks and poor blacks,” our contemporary times witness a calculating ma-
chine that also generates weighted and racialized targets and sentences. De-
spite the manifest differences between the discrete logics of particular ma-
chine learning algorithms, what they all share in common is the reduction of
a multiplicity of incalculable differences to a single output. This output is a
finite target contingent on infinite multiples of weights and weightings within
the calculation. Thus, when a random forest algorithm sentences someone to
death by drone strike, the infinite (gestures, connections, potentials) makes
itself finite (optimal output, selector, score), and the horizon of potentials is
reduced to one condensed output signal.
Moreover, this reduction to one output signal, a kind of death penalty in
its making of precise action from nonknowledge, shelters itself from its own
unreason. As societies seek out an encoded ethics that annexes and extracts
the errant, the aberrant, from algorithms, they are failing to understand how
these apparent pathologies are actually of the essence of algorithmic learning.
The clustering of false positives in a particular black population in one part
of a city, for example, if understood as errant becomes subject to correction
and recalibration. Yet, this experimentation and wild errancy continues to
generate targets long after a correction is made. The random forest algorithms
in skynet, as well as those in immigration systems that return targeted mi-
grants, condemn to death unknown people living and traveling in already
risky spaces, and they do so with wagers, votes, chance, and randomness. The
algorithm underwrites its own rationality because it is engaged in defining
new thresholds of normal and abnormal behaviors, reasonable and unreason-
able travel, the arbitration of the good and the bad. 129
The architecture of algorithms adjudicating life and death takes the form

The Madness of Algorithms

of a kind of scaffold, where the scaffold is the spatial arrangement of the death
penalty, the place of execution.68 In this sense the algorithm as scaffold is part
of the scaffolding of sovereignty, where the state “will have constructed all the
scaffolds and propped up all the figures of machines for killing legally, sov-
ereignly, nationally in the history of humanity.”69 The algorithm as scaffold
is sometimes — as with random forest algorithms for national security — the
means by which states exercise the right to precisely calculate death in ad-
vance; but it is also the architecture of a decision, a “certain modality, a cer-
tain qualification of living and dying, a manner, an apparatus, a theatre, a
scene of giving life and giving death.” Moreover, the scaffold “guarantees some
anonymity” for the executioner, who is not present on the scaffold and whose
hand is not on the guillotine, an “executioner who does not kill, not in his own
name.”70 The random forest algorithms I discuss in this chapter, as scaffolded
arrangements, similarly apportion weights to life and death as output signals,
guaranteeing some anonymity to the executioner, who does not appear on
the platform. The algorithm is precisely a space in which the excesses, the un-
reason, the cruelty of the adjudicator can be given free rein. And so, to find a
critical response to the algorithm via a demand for limits on its unreasonable
excess is to seriously overlook how exactly it deploys unreason to generate a
finite output.
This page intentionally left blank
Part 3

Ethics
This page intentionally left blank
5
The Doubtful Algorithm
Ground Truth and Partial Accounts

During the war I worked on the atomic bomb. This result of science was
obviously a very serious matter, it represented the destruction of people. . . . Is
there some evil involved in science? Put another way — what is the value of the
science I had dedicated myself to — the thing I loved — when I saw what terrible
things it could do?
 — Richard Feynman, What Do You Care What Other People Think?

I remain a child of the Scientific revolution, the Enlightenment, and

technoscience. My modest witness cannot ever be simply oppositional.
 — Donna Haraway, Modest_Witness@Second_Millennium

Leave Room for Doubt

In a lecture delivered to the National Academy of Sciences in 1955, physicist

Richard Feynman reflected on a particular relationship between the practice
of doing science and the value of a kind of orientation to doubt. As a graduate
student in the 1940s, he worked on the physics of the atomic bomb at Los Ala-
mos. Later, observing the terrible consequences of the weapons to which his
physics had contributed, he began to doubt the value of science and its respon-
sibility to society.1 In the years that followed his work on the Manhattan Proj-
ect, Feynman locates the practice of science in a particular method that per-
mits a freedom to doubt, an animated curiosity for otherwise ordinary things,
and a sustained sense of encountering an unknown future world. In his 1955
lecture, he describes how the scientist “must leave room for doubt”; proposes
that “it is perfectly consistent to be unsure, it is possible to live and not know”;
and makes a claim about responsibility to society: “Permit us to question, to
 doubt, to not be sure, herein lies a responsibility to society.” For Feynman
there is a profound ambivalence within the idea of doubt, so that it expresses
simultaneously “the openness of possibilities” into the future and the doubt-
fulness of science in the service of the state’s “monstrosities” of war.2
Feynman’s reflections on the responsibility that twentieth-­century sci-
ence has for violence and war testify to the intractable difficulties of delineat-
ing ethical lines of good and evil in technoscience. Along with many other
134 physicists whose experimental fragments became lodged within a horrifying
calculus, Richard Feynman embodies what Rosi Braidotti has called the “anti-­
Chapter Five

Humanism” of the “turbulent years after the Second World War,” when the
profound tyranny of human rationality was felt.3 Yet, as Donna Haraway ac-
knowledges of her own subjectivity as “a child of the Scientific revolution, the
Enlightenment, and techno-­science,” what it means to be human is not separa-
ble from technoscience and its place in war.4 Though Feynman’s situated and
partial accounts of the science of the atomic bomb bear witness to the state’s
enrollment of algorithmic calculative practices in war, these are not accounts
that stand outside science in opposition. Rather, the partial account dwells un-
easily within the interconnected narratives, stories, writings, and narratives
that together constitute the practice of science.
What is striking to me, in the letters and diaries of twentieth-­century
mathematicians and physicists, is that the intransigent and political difficulties
of partial accounts — manifest in the public debates of the twentieth century — 
are all but entirely absent in the twenty-­fi rst century’s enrollment of algo-
rithms by the state. In the demand for transparency and clear-­sighted account-
ability, the algorithm’s orientation to doubt is never in question. Amid the
pervasive twenty-­fi rst-­century political desire to incorporate all doubts into
calculation, algorithms are functioning today with the grain of doubt, allow-
ing the uncertainties and variabilities of data to become the condition of pos-
sibility of learning and making decisions. In contrast to Feynman’s notion that
science’s responsibility to society may reside in leaving open the incalculabil-
ity of the future, algorithms hold out the promise of securing against all pos-
sible future events. With contemporary machine learning algorithms, doubt
becomes transformed into a malleable arrangement of weighted probabilities.
Though this arrangement of probabilities contains within it a multiplicity of
doubts in the model, the algorithm nonetheless condenses this multiplicity to
a single output. A decision is placed beyond doubt.
As doubt becomes the terrain of algorithmic calculation, however, other
situated and embodied forms of doubt haunt the terrain. These opaque and
cloudy accounts seem present at all the moments when an algorithm gener-
ates itself in relation to a world of data. For example, when a biometric match-
ing algorithm doubts it can recognize a face because the data points were not
present in the training data, or when four desk analysts watching luminous
screens of risk scores worry that the model is “overfitting” to the data — these
are moments not of a lack or an error, but of a teeming plenitude of doubtful-
ness. What might it mean to invoke Feynman’s “room for doubt” or “to live
and not to know” in the context of a technoscience driven to expose machine
learning algorithms to data precisely to know and to act, indifferent to persis- 135
tent doubt?5 In light of the algorithm’s major role in the calculability of doubts,

The Doubtful Algorithm

in this chapter I propose that a reformulated doubt opens up new potentials
for cloud ethics. The post-­Cartesian and posthuman form of doubt I envis-
age begins from accounts of doubtfulness that decenter the agency of the hu-
man subject.6 This doubtful subject is not recognizable as a unitary individual
or even as exclusively human but is a composite subject whose collaborations
with algorithms open onto an undecidable future, where one is permitted to
ask new questions on the political landscape.
In contrast with a critique of algorithms as opaque and unaccountable,
then, my cloud ethics poses the problem differently, tracing the necessarily
cloudy and partial accounts that algorithms give as they generate themselves
as agents in the world. When Haraway describes the partiality of the “modest
witness,” she is concerned with “telling the truth, giving testimony, guaran-
teeing important things.” 7 The truthfulness of the account is thus grounded
not in the secure authorship of the reasoning subject but precisely in the im-
possibility of self-­knowledge that anchors the account. I return again to Judith
Butler’s important warning on the partiality of all forms of account. “My ac-
count of myself is partial, haunted by that for which I have no definitive story,”
she writes, so that “a certain opacity persists and I cannot make myself fully
accountable to you,” this partiality and opacity being the condition of ethical
relations.8 In short, algorithms do not bring new problems of black-­boxed opac-
ity and partiality, but they illuminate the already present problem of locating
a clear-­sighted account in a knowable human subject. Here we begin to find
the possibilities of an alternative orientation to doubt, one in which the subject
necessarily doubts the grounds of their claims to know.
One place for my cloud ethics to begin to intervene, then, is in the giving
of doubtful accounts, where the impossibility of a coherent account does not
mark the limit point of ethics but instead describes the condition of ethico-
political life. A doubtful account radically changes the terms of what counts
as truthfulness in relation to the algorithm. If one relinquishes the search for
a transparent truth of the algorithm, what kind of truth can be told? What
 kind of relation to oneself and to others is entailed by the algorithm’s particu-
lar claims to the truth?

Ground Truths

As algorithms become increasingly pervasive in supplying solutions for state

decision making, from the risk management of borders and immigration to
decisions on outcomes in the criminal justice system, they hold out the prom-
136 ise of a particular claim to truth. In public forums, when algorithmic decision
making is held to account, the form of truth is commonly professed to be an
Chapter Five

efficient and objective outputting of an optimized decision.9 In fact, though,

the mode of truth telling of contemporary algorithms pertains to the “ground
truth”: a labeled or an unlabeled set of training data from which the algo-
rithm generates its model of the world. In a process of supervised machine
learning, the algorithms learn from a ground truth model of data labeled by
human domain experts, often via crowdsourced labor, such as Amazon’s Me-
chanical Turk. As I describe in relation to the AlexNet image recognition algo-
rithm in chapter 2, when a new set of features is extracted from the input data,
these features are weighted in relation to the ground truth data. For example,
a facial recognition algorithm used in urban policing is able to identify a face
because of its exposure to a ground truth dataset of labeled images.10 Increas-
ingly, as deep learning algorithms derive their own ground truths by cluster-
ing raw unlabeled data, a model of what is normal or anomalous in the data
is generated by the algorithms. The claim to truth made by machine learn-
ing algorithms, then, is not one that can be opposed to error or falsity, nor
even to some notion of feigning that targets “fake news.” Rather, the algo-
rithm learns from the degrees of probabilistic proximity to, or distance from,
a ground truth, itself often generated by algorithms. So, when neural network
algorithms at the border or in policing or immigration reach a decision, this
is a decision derived from an output signal that is entirely contingent on a set
of gradients and weighted probabilities.11 The architectures of neural network
algorithms can contain twenty hidden layers, hundreds of millions of weights,
and billions of potential connections between neurons.12 Given the opac-
ity of this arrangement, for an algorithm to generate an output (which is to
reach a decision), it requires a mode of truth located in its relations to ground
truth data. Put simply, the truth telling of the algorithm is entirely contin-
gent on the notion of a particular kind of grounds that I call a data ground
truth.
This malleable relation to a ground truth supplies the algorithm with the
capacity to work with a grain of doubt and uncertainty. The output of the al-
gorithm places a decision beyond doubt in the sense that it always already em-
bodies the truth telling of the ground truth data. As one computer scientist
explained of an image recognition neural network trained to detect cats, “it
is not recognizing a cat, but a cat at this angle, that angle, in a tree, or upside
down. . . . There is no sense of catness, just enough cats in places that there can
be a pattern of what a cat is.”13 Thus, the decision that an entity is or is not a
cat is placed beyond doubt, not in the sense that it could not be false but be-
cause it embodies the truth telling of the ground truth data of what a cat can 137
be. Taking this truth logic to another domain of neural networks, one could

The Doubtful Algorithm

similarly contest the output of a recidivism algorithm for being “false” in the
sense that it wrongly assigns someone a high probability of reoffending, for
example, but its degree of truth will always remain intact in its relations with
the ground truth data. To be clear, the algorithm does not eradicate doubt, but
neither does it only productively incorporate doubt, as has been observed in
the methods of scenarios, catastrophe modeling, and preemption in geopoli-
tics.14 Though this is a science that can hold together multiple possible versions
of events simultaneously, each possibility weighted as a layer of computation
in the algorithm, it cannot live with doubt as such. How could a person say,
“That is false; I am not in fact likely to reoffend if I am released,” if the truth
telling of the algorithm is anchored in the ground truth data and not in their
situated life? That is to say, the machine learning algorithm must reduce the
vast multiplicity of possible pathways to a single output. At the instant of the
actualization of an output signal, the multiplicity of potentials is rendered as
one, and the moment of decision is placed beyond doubt.
In the pages that follow, I address the practice of scientific truth telling in
today’s machine learning algorithms, and I seek to give some revitalized life
to the embodied doubt that is always already present within the science. In
one sense this is a form of critique of vision-­dominated objectivity that claims
to have a truth beyond doubt, but more precisely, it seeks something akin to
what Donna Haraway has called the “embodied objectivity” of “partial per-
spective,” and to what N. Katherine Hayles captures as an “embodied actual-
ity.”15 This different orientation to doubt begins from the embodied doubts in-
hering within all notions of subjectivity and objectivity, doubting oneself and
one’s capacity to know. With this orientation to an embodied doubtfulness, a
cloud ethics reopens the contingencies of the appearance of ground truth data,
giving life to the fallibilities of what the algorithm has learned about the world
across its billions of parameters, and rendering the output incomplete. Run-
ning against the grain of the Cartesian doubt of the fallibility of data derived
from the senses, doubt in this alternative register is felt, lived, and sensed as
 embodied actuality in the process of an algorithm learning through its rela-
tions with the world.

Doubtful Science

To consider the doubtfulness of partial perspectives is not the same thing as

casting doubt on the algorithm as such. Indeed, the point is that one could
never stand outside the algorithm to critique or to adjudicate on its veracity
138 since one is always already implicated in the algorithm as a form of adjudica-
tion of the truth. Instead, a cloud ethics dwells within and alongside the pleni-
Chapter Five

tude of teeming doubts in the experimental algorithmic model. To explain

what might be at stake in this reorientation to doubt, I turn to an example
drawn from a set of interviews with designers of deep learning algorithms that
are actively generating their own ground truths about the world. The design-
ers are explaining how the attributes within a ground truth dataset become
the means to detect future attributes in other people, related only through the
calibration of their attributes:

One piece of work we did recently with an insurance company, where we

were looking at the value of their life insurance customers. Now obviously
they only take out life insurance maybe once, [laughs] so you haven’t got
lots of different data points that you can look at, but one of the things
we can do there is (a), we can start to say, well what does everybody else do?
And so we have what’s called the context, so we’re able to then compare
the individual against a cohort group of attributes in order to be able to
start to make predictions. And the other thing we can do (b) is we can take
supplementary data in as well, so it could be, you know, a Facebook feed
or an actuarial feed or it could be, you know, data that’s coming through
from the Home Office for instance, so we can start to supplement the data
that already exists with other datasets in order to be able to improve our
predictions.16

The value of a life insurance customer (the risk-­based price of their in-
surance) is thus generated from the output of the neural network, calibrated
against a threshold of value to optimize the insurance company’s position.
Where the data points on past known entities are scarce — as with the “cou-
rier” proxies for terrorists I discuss in chapter 4 — the algorithm perceives what
others are doing and establishes a ground truth from which attributes of risk
or opportunity can be derived.17 The life insurance algorithm emerges through
the partial and oblique correlations among otherwise unrelated selves and oth-
ers. It will go out into the world, continue to iterate and learn, ingest new data
points, create new clusters, and generate new outputs for decisions. When the
algorithm designers refine and adapt their algorithm for new domains, they
are experimenting with the proximity between a specified target value and
the actual output signals from their model, observing how the numeric scores
generated by their model diverge or converge on the target. The computations
of algorithms like these are doubtful in their openness to new interactions
and encounters in the world, every layer infinitely malleable and contingent
on plural interactions of humans and algorithms. A minor change in the sta- 139
tistical state of an entity — the gradient of a high-­value life, or the features of

The Doubtful Algorithm

a high-­risk migrant — will transform the output signal and, thus, the targeted
decision.
Understood as a practice that is partial, iterative, and experimental, the
neural network algorithm is doubtful not only in the sense that it supplies a
contingent probability for an absolute decision, but moreover because it ac-
tively generates thresholds of normality and abnormality. Put simply, the al-
gorithm does not need to eradicate doubt or establish certainty for the deci-
sion because it generates the parameters against which uncertainty will be
adjudicated. The algorithm learns to recognize and redefine what is normal
and anomalous at each parse of the data. What does everybody else do? What
are the attributes of the cohort? What is the value of the life? The response to
such questions is never authored by a clearly identifiable human, but rather
derived from a composite of algorithm designers, frontline security officers,
experimental models of the mathematical and physical sciences, a training
dataset, and the generative capacities of machine learning classifiers working
on entities and events. For example, when Palantir supplies its neural network
“investigative case management” algorithms to US Immigration and Customs
Enforcement (ice), the border guard ebbs into a composite of machine meth-
ods for “discerning a target” and “creating and administering a case against
them.”18 The actions that are outputs of the algorithm — to arrest persons, to
seize assets, to indict persons — a re arranged in relation to the falcon ma-
chine learning analytics engine and the ice data warehouse.
There is a great deal at stake politically in the erasure of the doubtfulness
of algorithms at the point of political decision. Though the making of a neu-
ral net algorithm for immigration enforcement or for life insurance risk is a
fraught, political, and doubtful practice, the computational structure of such
algorithms dictates that the final output must be a single numerical value be-
tween 0 and 1. In effect, the output of a neural network is a numeric probabil-
ity, a single value distilled from a teeming multiplicity of potential pathways,
values, weightings, and thresholds. This process of condensation and reduc-
 tion to one from many allows algorithmic decision systems to retain doubt
within computation and yet to place the decision beyond doubt. What would
it mean to be able to express a posthuman doubtfulness in this context? How
does one speak against the grain of the single output of automated visions of
security? Is it possible to locate and amplify the doubtfulness dwelling within
the partial fragments of the science itself?

140 Partial Accounts and Fatal Decision

To detail how partial accounts might advance a different orientation to doubt,

Chapter Five

I turn here to a historical moment when the output of a risk algorithm failed
catastrophically. This is a moment when a Cartesian doubt was expressed that
inaccurate data may have led to a catastrophic decision. When the event is un-
derstood as composed of opaque and partial fragments, however, what comes
to matter is not whether something could be calculated accurately from data,
but rather how partial probabilities became a singular calculus for a fatal deci-
sion. On January 28, 1986, seventy-­three seconds after its launch at 11:25 am,
the nasa space shuttle Challenger broke apart over the Atlantic Ocean, kill-
ing all seven crew members, including Christa McAuliffe, nasa’s “teacher
in space.” By February, President Ronald Reagan had established the Rogers
Commission to “review the circumstances surrounding the accident and es-
tablish the probable cause or causes” and to “develop recommendations for
corrective or other action.”19
Among the fourteen commissioners appointed to the Challenger investi-
gation was physicist Richard Feynman, who was persuaded to participate de-
spite, by his own account, “having a principle of going nowhere near govern-
ment” after his experience of working on the atomic program at Los Alamos.20
More specifically, he was wary of bureaucratic reason and the governing of
what he thought of as an unruly science by bureaucratic rules and protocols.
The Cold War cybernetic rise of “algorithmic rules that could be executed
by any computer, human or otherwise” with “no authority to deviate from
them” had extended a particular kind of “arithmetic into the realm of deci-
sion,” and this entanglement of mathematics and decision did not please Feyn-
man.21 Feynman’s letters and diaries reveal a disdain for algorithmic decision
procedures and axiomatic formulas, and a propensity to ask questions that
ran against the grain of mathematical rules. “Doing it by algebra was a set of
rules,” writes Feynman of his early encounters with mathematics, “which, if
you followed them blindly, could produce the answer.”22 For Feynman, algebra
was a means of imposing an axiomatic set of procedures of rules on a puzzle
that could otherwise intuitively work toward an unknown outcome.
 Confronted with the Rogers Commission’s setting of procedural steps in
the investigation of the Challenger disaster, Feynman worked to reinstate the
doubtfulness at each link in nasa’s chain of events. He traveled to meet with
the engineers, avionics scientists, and physicists whose data on particular
components had made up the aggregate risk calculation on which nasa had
based their launch decision. What he discovered was not a catastrophic depar-
ture from the normal rules or a “human error” or failure as such. Instead, the
launch decision belonged properly to a posthuman composite of algorithms, 141
where the steps of a normalized risk-­calculation protocol had been followed

The Doubtful Algorithm

beyond the limits of the calculable. As sociologist Diane Vaughan proposes
in her account of the Challenger disaster, “harmful actions” can be “banal,”
and they can “occur when people follow all the rules.”23 Though Vaughan’s
famous account emphasizes the role of the situated culture of nasa in the er-
rors made, Feynman’s account foregrounds the doubt already present within
the science of the program. Where nasa failed, according to Feynman, was
in its tendency to “aggregate out” the multiple small fractures and failings
that existed at the level of the most mundane of components and instruments.
Understood in this way, Feynman’s critical scientific method does not seek to
correct inaccuracies but instead brings to the surface the doubts already pres-
ent within each fragmented element so as to open up the breaches in the ar-
rangement of propositions.
Here, I reflect on two aspects of Feynman’s method for the possibilities
they may offer to a critical cloud ethics method of partial accounts. The first
is a particular reinstatement of doubt in data as they are given. The method of
reinstatement is intended specifically to reinstate, or to give something back a
position it had lost. This is of some significance, for doubt does not function to
cast uncertainty on data that were heretofore settled and certain. Feynman’s
approach to doubt begins from the position that all scientific data are contin-
gent, uncertain, and full of doubt. Indeed, among Feynman’s major contribu-
tions to quantum physics was his “sum over histories” method, in which the
calculation of particle interactions must “take into account a weighted tally of
every possible chronology.”24 In this sense, Feynman’s physics theorized pre-
cisely how a calculation might keep open the possibilities of multiple pathways
that are otherwise incalculable. The algorithm unfolds as an array of mutable
contingencies, each intimately and recursively layered with others so that the
science strays from calculable paths and experiments with the incalculable. As
Karen Barad suggests in her account of the intimacies of feminist science stud-
ies, “Life is not an unfolding algorithm” so that “electrons, molecules, brittle-
stars, jellyfish, coral reefs, dogs, rocks, icebergs, plants, asteroids, snowflakes,
 and bees stray from all calculable paths.”25 In my refiguring of doubt, I envis-
age the unfolding algorithm as straying from calculable paths in the manner
Barad proposes. The hesitant and nonlinear temporality of the etymology of
doubt, from the Latin dubitare, suggests precisely a hesitation, an uncertainty,
and a straying from calculable paths. To be doubtful could mean to be full of
doubt, in the sense of a fullness and a plenitude of other possible incalculable
paths.
142 What might a method of reinstating doubt, or giving doubt a presence
that it had lost, look like? Returning to Feynman’s investigations into the Chal-
Chapter Five

lenger disaster, the seals of the solid fuel rocket boosters of the shuttle were
secured by pairs of rubber O-­rings. It was known to nasa that, during some
launches, the hot gases from the boosters could leak from the seals, causing
what was called blowby over the liquid oxygen and nitrogen of the orbiter’s en-
gines. Though the engineers observed corrosion of the O-­rings over time, they
assumed that the heating of the rubber O-­rings during launch caused small ex-
pansions to close the gaps in the seal. At the fatal launch, the temperature was
28 degrees Fahrenheit (−2 Celsius); the coldest launch recorded prior to this
had been 53 degrees Fahrenheit. On the evening of January 27, the night before
the launch, the engineers at Thiokol, which manufactured the seals, warned
nasa that the launch should not take place if the temperature was below 53
degrees. Yet, the public record shows that nasa “took a management decision”
because “the evidence was incomplete” that “blowby and erosion had been
documented above 53 degrees,” so “temperature data should be excluded in
the decision.”26 As Feynman writes, nasa testified that “the analysis of exist-
ing data indicated that it [was] safe to continue flying.”27 In effect, the aggre-
gate data of twenty-­four past flights without a mission failure had placed the
launch decision beyond the doubt that would otherwise have been reinstated
by attentiveness to the components, and to the inscribed traces they embodied.
A sense of doubtfulness that might be considered fully posthuman — dwelling
in the material marks of blowby and erosion, and in the touch of engineers on
fissures and cracks — was aggregated out in the calculation of a singular output.
In historical instances such as Feynman’s account of Challenger, we can see
some of the fallacies of an objective risk-­based or data-­driven decision placed
beyond doubt. Remarking on the computer model used by nasa in support of
their launch decision, Feynman writes, “It was a computer model with various
assumptions. You know the danger of computers, it’s called gigo: garbage in,
garbage out.”28 To question the assumptions of the computer model, Feynman
began to focus on the fragments of data and their associated probabilities, re-
instating doubt within each element. Understood in this way, the probabilis-
tic calculation of risk gives way to what Donna Haraway has called “partial,
locatable knowledges,” in which there is the “possibility of new webs of con-
nections.”29 nasa had testified to the commission that the probability of fail-
ure of a mission was calculated to be 1 in 100,000. “Did you say 1 in 100,000?,”
queried Feynman. “That means you could fly the shuttle every day for 300
years between accidents?”30 To open up a breach in the risk calculus, Feynman
gathered the scientists who had worked on the various different components of
Challenger, asking them to estimate the probability of failure of the shuttle and 143
to write this probability on a slip of paper. When he collected the fragments

The Doubtful Algorithm

of paper with their pieces of data, he found that the calculus reflected the par-
ticular and situated relationship to a material component and its properties.
Thus, for the seal engineers, the probability of failure was felt to be 1 in 25; for
the orbiter’s engines, 1 in 200; with none of the data elements showing nasa’s
aggregate probability of 1 in 100,000. Feynman’s point was not to correct the
calculation but rather to dramatize its impossibilities and incalculables. From
the paper fragments of each embodied likelihood, gathered as incongruous
scraps, the possibility of new webs of connections emerges.
The second aspect of Feynman’s method I discuss here is the affordance of
capacities to technical devices — in particular, an affordance of the capacity to
give accounts of themselves, their breaches, and their limits. One approach to
such affordances would be to say that, put simply, when something fails, it also
speaks to us and tells us something of its limit points.31 Yet, this is not merely
a question of how the technical object has capacity to “gather to itself ” a com-
munity of other human and nonhuman beings.32 Instead, as Katherine Hayles
reminds us, “Cognition is much broader than human thinking,” so that cogni-
tive capacities are afforded to other life forms, animals, and technical devices,
forming a “rich ecology of collaborating, reinforcing, contesting and conflict-
ing interpretations.”33 Within this broad ecology of cognition, the always pres-
ent ethicopolitical difficulty of humans giving an account of themselves ex-
tends also to the partial and incomplete accounts of other beings.
When Feynman found that the Rogers Commission was unwilling or un-
able to register his critical account, he brought with him to the public assem-
bly part of his material scientific community, inviting the material device of
the O-­ring to give an account in the forum.34 As the government officials and
press gathered for a meeting with the commission, Feynman placed a piece of
rubber O-­ring in a clamp bought in a downtown hardware store. Requesting
a glass of ice water, he dropped in the materials and waited for them to cool.
When the rubber was removed from the glass, brittle and compressed, it broke
apart before the assembled audience (figure 5.1). One reading of the event of
144
Chapter Five

Figure 5.1  Richard Feynman removes the O-­ring material from ice water at the
Rogers Commission inquiry, February 11, 1986. “Richard Feynman (Articulate
Scientist) — Space Shuttle Challenger Testimony,” https://www.youtube.com
/watch?v=MWZs8l2AMps, posted October 3, 2013.

the O-­ring giving an account of itself is that Feynman follows the “structure
of heroic action” in science, as Donna Haraway has put it, speaking objectively
and “self-­invisibly” through the “clarity and purity of objects.”35 Certainly the
now famous New York Times report of the O-­ring event, citing Feynman’s few
words, “There is no resilience in this material when it is at a temperature of
32 degrees,” bears a form of witness close to that of Haraway’s critique. I lo-
cate a different form of account here, however, one with no unified authorial
source or transparent picture of the truth, but rather a distributed and oblique
account in which we see the political impossibility of resolving the truth be-
fore the public. What is dramatized with the rubber O-­ring in the ice water
is a set of relations that are material and political. Refusing the mode of ac-
countability in favor of this different form of giving an account, there is what
Karen Barad calls a “condensation of responsibility” in matter, wherein the
multiple contingent past decisions are lodged within the object, engaging us
in “a felt sense of causality.”36 In this way, a material limit that was ordinary
in the embodied experiences of engineers and materials scientists — familiar to
their touch — could be expressed as a claim that could be heard in the world.
Feynman’s claim did not express the voice of Haraway’s “victor with the stron-
gest story”; indeed his account was derided and rejected by the commission
throughout.37 At the specific moment I describe, Feynman speaks against the
grain of Cartesian doubt and instantiates a mode of doubtfulness more com-
monly feminized and annexed within subjectivity. It is precisely this mode of
intuitive causality and embodied doubtfulness that I am seeking as a resistant 145
and critical form of responsibility.

The Doubtful Algorithm

Risky Speech

To speak of the excesses and limits of technoscience in our contemporary pres-

ent has become extraordinarily difficult. Every output of the algorithm, even
when it leads to wrongful detention or racialized false positives, is produc-
tively reincorporated into the adjustment of the weights of a future model.
The method of reinstating doubtfulness within algorithmic arrangements,
however, has the potential to yield a kind of “risky speech,” or “parrhesia,”
where, as Michel Foucault contends, an account is given that places itself at
risk.38 Significantly, Foucault considers it possible for science to give parrhesi-
atic accounts, suggesting that “when deployed as criticism of prejudices” and
“current ways of doing things,” science can place itself at risk and play the
“parrhesiatic role.”39 For Foucault, the defining character of parrhesia is that
the parrhesiast is bound to the truth of their claim, but they place themselves
at risk in so doing, because they speak against the grain of prevailing thought.
In this sense, the parrhesiast can never “witness with authority” but only with
a curiosity that is not counted before the public assembly. As Judith Butler
comments on Foucault’s mode of criticism, the practice “involves putting one-
self at risk,” “imperilling the very possibility of being recognized by others,”
and risking unrecognizability as a subject.”40 Understood thus, risky speech
does not come from an already authorized subject, but only from a subject who
risks making themselves unrecognizable to others. To place oneself at risk of
being unrecognizable is what seems to be necessary in an age of algorithmic
calculations. What marks out the risky speech of parrhesia from other forms
of political claims, other modes of discursive and performative speech, is its
capacity to open onto an indeterminate future: “There is a major and crucial
difference. In a performative utterance, the given elements of the situation are
such that when the utterance is made, the effect which follows is known and
ordered in advance, it is codified. . . . In parrhesia, on the other hand, the ir-
ruption determines an open situation, or rather opens the situation and makes
 possible effects which are, precisely[,] not known. Parrhesia does not produce a
codified effect; it opens up an unspecified risk.”41
Understood in these terms, parrhesiatic speech is an irruption that opens a
situation and makes possible effects that are not known in advance. By contrast,
the search for an encoded ethics of algorithms proceeds from a known subject
and a problem that is ordered in advance. For example, when Google employ-
ees protested Project Maven’s “involvement in war” and “biased and weapon-
146 ized ai,” the problem of “data captured by drones to detect vehicles and other
objects” was ordered in advance. Despite the manifest entanglements of the
Chapter Five

drone program in multiple other Google object recognition systems, the protest
delineated good from evil, and war from commerce, in ways that did not place
the subjects and objects of science at risk. Within a cloud ethics, risky speech
permits loosening the hold on subjectivity and relinquishing the grounds from
which the subject speaks. When the “avowing subject,” as Foucault writes,
“loosens its hold,” it becomes possible to envisage a risky speech that forges new
ethicopolitical relations with oneself and others.42 This means that one could
be doubtful of all claims, for example, that the bias or the violence could be ex-
cised from the algorithm and begin instead from the intractable political rela-
tions between the algorithm and the data from which it learns.
Ultimately, on discovering that the Rogers Commission report contained
a previously undisclosed “tenth recommendation,” sheltering the risk algo-
rithm within “error” and commending nasa’s work and mission to the na-
tion, Richard Feynman refused to add his signature to the report. His fellow
commissioners sought to persuade him, suggesting that the tenth recommen-
dation was a feminized “only motherhood and apple pie,” and that “we must
say something for the president.”43 The tenth recommendation, which would
authorize a particular decided future of continued nasa missions, was demar-
cated from the real science of objective findings and rendered maternal and
nation building.
In response, Feynman asked what kind of truth could be spoken to the
government: “Why can I not tell the truth about my science to the President?”
In the event, his own report appeared as an appendix to the published docu-
ment, titled “personal observations on the reliability of the shuttle,” as though
his merely personal situated account could only be subjective and without
grounds. Feynman’s appendix is risky speech in Foucault’s sense, an argument
that places the subject at risk, runs against the grain of the commission report,
and challenges the form of truth telling of the public report. In this sense, to be
doubtful is to open onto the full contingency of a situation and to be responsive
to unspecified future effects. The fullness of an embodied posthuman doubt
I envisage here is a critical cousin to Lauren Berlant’s notion of a “cruel opti-
mism,” where a “cluster of promises could seem embedded in a person, a thing,
an institution, a text, a norm, a bunch of cells, smells, a good idea.” The cru-
elty Berlant so vividly describes is a “relation of attachment to compromised
conditions of possibility whose realization is discovered to be impossible, sheer
fantasy, or too possible, and toxic.”44 Though Berlant does not probe directly
what it might mean politically to be doubtful of the promises as they embed 147
in a person, a thing, or a technology, her work does envisage the making of

The Doubtful Algorithm

claims in the present so that there remains “the possibility of the event taking
shape otherwise.”45 Understood thus, to be doubtful could be to experience a
fullness or multiplicity of the present moment and the many ways it might un-
fold, such that the cruelly optimistic promises of technoscience do not cling
so tightly to ideas of the optimal or the good enough output. The optimism of
the algorithm is founded on its optimization, that is, on its capacity to reduce
the fullness of the present moment to an output to be acted on in the future.

Doubtfulness and Ethics

If a signaling of the multiple errors and faults in contemporary machine learn-

ing algorithms cannot meaningfully provide grounds for critique, then can
there be risky speech amid algorithmic techniques? I wish to note here that
throughout all my discussions with algorithm designers, there has been a cu-
rious mood of twinned optimism and doubt: optimism that a good enough
model can always be found, and doubt that the “junk data” elements of social
media sentiment or unlabeled image files could ever be adequately prepared.46
If, however, the many misidentified faces in biometric systems, or wrongly
seized assets in a Palantir immigration system, merely “unleash computable
probabilities into everyday culture,” as Luciana Parisi reminds us, then how
does one speak against the grain of prevailing algorithmic logics?47 Put differ-
ently, Richard Feynman excavated and amplified the doubt within the ana-
log data elements by painstakingly reconstructing the plural probabilities of
each finite piece. If this method of gathering calculative fragments allows us
to pause and dwell within the doubtfulness of the calculus, can we imagine to-
day an equivalent method of asking the scientists we research to please “write
the probability of the failure of your piece of the machine learning software
on this piece of paper.” How would the reinstatement of doubt account for
the adjustments of weights that are conducted by the algorithm on itself? If
one sought to reinstate the doubtfulness of test data, then what kind of truth-­
 telling practice could this be? How would it extend to the Amazon Turk out-
sourced workers who had labeled the images for a machine learning algorithm
to recognize a future human face? Could the doubtfulness of a numeric prob-
ability account for the future uses of one algorithm used to teach another?
When Richard Feynman asked the scientists on the shuttle program why
the fragments of their doubts had not found their way into the future calcula-
tions of risk assembled by nasa, he was informed that “it is better if they don’t
148 hear the truth, then they cannot be in a position to lie to Congress.”48 The
problem of truth telling in this instance lies less with science as such, and more
Chapter Five

with the various ways in which embodied doubtfulness in science does not give
political accounts of itself. Similarly, when the former US director of national
intelligence James Clapper was found to have lied to Congress on the analysis
of data on US citizens in March 2013, he later argued that he had provided the
“least untruthful answer possible in a public hearing.”49 To speak against the
grain of the promise algorithms make in the name of political resolution is to
bring to the public assembly a kind of contemporary heresy. Following the UK
terror attacks in Manchester and London in 2017, for example, where some of
the perpetrators had previously been known to intelligence agencies, the offi-
cial inquiry by David Anderson has concluded that what is required is a more
effective deep machine learning algorithm for the identification of high-­risk
individuals among twenty thousand so-­called closed subjects of interest.50 In
effect, this implies training an algorithm to recognize the attributes of known
individuals in the dataset to identify similar attributes among unknown fu-
ture subjects. The weighted probabilities among the twenty thousand will be
used to generate target outputs for future security action. To signal the doubt-
fulness of such an algorithm is precisely to engage in risky speech and annul
the promise of a securable future.
The form of doubtfulness I have described would demand that algorithms
give political accounts that are marked by contingency and ungroundedness.
In place of an ethics that seeks defined grounds for algorithmic decisions, a
cloud ethics begins from the ungroundedness of all forms of decision, all polit-
ical claims, human and algorithmic. As Thomas Keenan writes, “What we call
ethics and politics only come into being or have any force and meaning thanks
to their ungroundedness,” so that “we have politics” precisely because we have
no “reliable standpoints.” If a future political claim is to be possible, then “its
difficulty and its persistence stem from its terrifying, challenging, removal of
guarantees.”51 If there are to be future claims to human rights not yet registered
or heard, for example, then there must always also be doubts haunting the pos-
sibility of rights, doubts that acknowledge the irresolvability of all political
 149

The Doubtful Algorithm

Figure 5.2  The threat centaur model of intelligence gathering and threat
detection, as presented by Recorded Future. Author photograph.

claims. Here, doubt pervades what Keenan describes as a “darkened frontier,”

where the essential difficulties of politics must take place, and where all deci-
sions must, following Derrida, “act in the night of non-­k nowledge and non-­
rule.”52 In holding out the promise of resolving political difficulty — coming
out of the darkness with the target output of a machine learning algorithm —
 technoscience harms the terrain of the political. The algorithm promises to
resolve the political by holding together a multiplicity of possibilities within
the neural net, while carving out a space for a single optimized decision. This
is not to say that the decision is made without doubt. On the contrary, it is
doubtful, full of doubt, and yet it decides with indifference.
Seated in the audience at a UK technology company’s presentation, I
watched the pitch made to UK government officials in the name of resolving
political difficulty with machine learning algorithms.53 Recorded Future ex-
plained how their “threat centaur” machine learning algorithms (figure 5.2)
would “scrape the web for geopolitical events” and send “real time alerts” on
“political protests, terrorist attacks, and civil unrest.”54 As the algorithm de-
 signers describe the system, it deploys a combination of rule-­based and deep
learning algorithms. The rule-­based algorithms are said to be based on “hu-
man intuition” about whether “an entity is associated with some kind of risk,”
while the machine learning algorithms are trained on a “large dataset, using
trusted threat list sources as ground truth for what constitutes a malicious
entity.” The figure of the “centaur” in the algorithmic system is a posthuman
body with a capacity to reach an unsupervised algorithmic decision (“this
150 event is critical”) but also to display “a human readable motivation for that
judgement,” so that action can be authorized against a threat.
Chapter Five

At one level the algorithms of the threat detection centaurs appear to run
counter to the form of risk calculation observed by Feynman at nasa. The
smallest of infractions or the subtlest of signals of a component failure or a
change in online clickstream patterns, for example, is said to yield a threat
alert to the centaurs. Like the Rolls Royce jet engines that yield real-­time data
on the performance of aircraft as they fly, the twenty-­fi rst-­century O-­ring
would almost certainly be given a data voice. And yet, at the level of practices
of truth telling, there is something of significance here. The Challenger launch
decision teemed with a cacophony of embodied doubt — on the part of the en-
gineers, physicists, and mathematical modelers, in communion with their
nonhuman partners — a nd yet, the decision was placed beyond doubt. What
we are seeing emerge in contemporary algorithmic technoscience is an orien-
tation to truth telling that says that all political difficulty can be resolved, all
decisions can be placed beyond doubt.
How might one intervene in models such as the threat centaur in ways
that reinstate the doubtfulness of the algorithm? It is necessary to engage the
traces of machine learning algorithms, to be curious about them, to doubt
them, but also to listen to the doubt they themselves express in the world. In
engaging research methods that follow elements of an algorithmic solution as
they travel and have onward life, I am trying to be a receptive listener when
the piece of data appears to say “you do know that I cannot really be cleaned; I
remain in there, muddying the calculation”; or when the biometric-­matching
algorithm says “in 14 percent of cases, where ambient lighting exceeds these
parameters, I cannot be read”; or when the pattern recognition algorithm says
“the training dataset I have been exposed to has established what looks normal
or abnormal to me.” There is doubt, and it proliferates everywhere; at every
step in the application of algorithmic solutions to political questions, doubt
lives and thrives and multiplies. And, where there is an ineradicable doubt,
there is also the darkened frontier and, therefore, the potential for politics. In-
spired by Richard Feynman’s method of reinstating doubt in the data points,
and being an interlocutor for the people and things “in the engine” who ac-
tively give accounts of limit points of all kinds, I urge that we experiment as
the algorithms do, iteratively and recursively, giving doubtful accounts of the
calculation.

The Open Channel

To speak in praise of an embodied form of posthuman doubt is to draw to-

gether several things that do not sit together easily in terms of critique. The 151
form of doubt I am proposing is to be distinguished from Cartesian skepticism,

The Doubtful Algorithm

for it does not seek to annex fallible sensory doubts in the pursuit of foun-
dational truths. As posthuman composite life forms, the doubtful subjects I
depict here dwell uneasily within partial and situated knowledges, and when
they make a claim in the world, they do so in ways that stray from calculable
paths. I include within this category of doubtful subjects all those whose sci-
ence bears a responsibility for an apparently risk-­free, error-­tolerant political
decision, from drone strikes to border controls, and from voter targeting to
immigration decisions. In so doing, I follow Foucault’s sense that the scientist
and the geometer do not stand outside the capacity for parrhesiatic, or risky,
speech, alongside Karen Barad’s caution that we are materially immersed in
and inseparable from science. Doubtfulness expresses the many ways in which
algorithms dwell within us, just as we too dwell as data within their layers,
so that we could not stand apart from this science even if we wanted to. The
doubtfulness of our relations to ourselves and to algorithms implies that cri-
tique will always involve “putting oneself at risk,” as Judith Butler describes it,
and risking “unrecognizability as a subject,” as one can “never fully recuperate
the conditions of . . . emergence.”55 In this way, the unrecognizability of the
subject extends into the opacity of the algorithm, so that new forms of subjec-
tivity emerge that are never fully recuperated. The next time one hears that
the “black box” of the algorithm should be opened, one might usefully reflect
on the unrecognizability of all forms of self.
To express doubt in these terms is to try to find a means to respond to
algorithms that incorporate doubt into computation, converting doubts into
weighted probabilities that will yield a singular output that can be actionable.
A critical response cannot merely doubt the algorithm, pointing to its black-­
boxed errors and contingencies, for it is precisely through these variabilities
that the algorithm learns what to do. Instead, an orientation to doubtfulness
decenters the authoring subject, even and especially when this author is a ma-
chine, so that the grounds — and the ground truths — are called into question.
Understood in this way, to doubt is also to reopen a “decision worthy of the
 name,” as Derrida notes, or to open onto “a politics of difficulty,” as Thomas
Keenan observes, in which the “break with the humanist paradigm” does not
close off the ethicopolitical but provides its starting point.56
What could it mean to value doubtfulness as a critical faculty, to reinstate
doubt into the composite creature that is automated algorithmic calculation?
It is not the same thing at all as calling for an ethics of human decision in
machinic security (putting the human back in the loop), or saying that algo-
152 rithmic decisions are intrinsically malevolent. Instead, it expands the space
for doubt beyond what Deleuze, following Bergson, calls “the decisive turn,”
Chapter Five

where lines diverge according to their differences in kind.57 Understood in this

way, the action signaled by the output of the algorithm is never placed beyond
the darkness of doubt and difficulties, for it carries doubt within, is always in-
complete, and does not know what is around the decisive turn. Though con-
temporary algorithms reduce the multiplicity to one at the point of decision,
they do not leave behind the knots of doubt that dwelled within the multiple.
The condition of possibility of a condensed algorithmic output, in short, is
a teeming multiplicity of potentials that remain in play. As Barad writes on
Feynman’s quantum field theory, in which the electron absorbs within itself
its own proton, and “there is something immoral about that,” the action of the
body is never fully beyond doubt.58
In so many ways, of course, the physicist Richard Feynman is a curious fig-
ure to invite into conversation with Haraway, Braidotti, Hayles, and Barad on
science and the ethicopolitical, particularly given his role in the material algo-
rithmic violence of the nuclear age. Yet, the point is that none of us can stand
outside this science and judge it to be good or evil, to say definitively “things
should not be this way.” We are within it, and it is within us, such that critique
cannot begin from an outside to the algorithm. As Barad suggests, “Ethicality
entails noncoincidence with oneself,” so that to give an ethical account is to
resist anchoring the claim in a knowing subject.59 Doubt does not stand out-
side and pronounce judgment; doubt is also interior, essential to the displace-
ment of all subjectivity, for we doubt ourselves and our place in the world. For
Feynman, doubt is what he describes as the “open channel” in politics and in
science: “If we take everything into account . . . then I think we must frankly
admit that we do not know. But, in admitting this, we have probably found the
open channel. . . . The openness of possibilities was an opportunity, and doubt
and discussion were essential. . . . We must leave the door to the unknown
ajar. . . . It is our responsibility to leave the people of the future a free hand.”60
To leave the door ajar is to resist the promises made that the future is resolv-
able through the optimized output of algorithmic decision engines. I open this
book with a discussion of the algorithm as a “hinge” on which the very idea
of politics might turn. To leave the door to the unknown ajar is to reopen the
many points in the algorithm where another future was possible, where the
hinge does not quite limit the possible movement. While the enfolded doubts
of the algorithm’s feedback loop and back propagation deny the impossibility
and difficult politics of decision, an embodied doubt pauses with the undecid-
ability of alternative pathways and their contingent probabilities. The claim
to a ground truth in data that pervades our contemporary geopolitical imagi- 153
nation precisely closes the door to the future, offering algorithmic solutions to

The Doubtful Algorithm

resolve the difficulties of decision. To reinstate embodied doubt within the al-
gorithm, and to allow the components of the badly formed composite to speak
of their limits, is to seek to leave the door ajar for the making of future politi-
cal claims, for these future unknown claims can never appear in the clusters,
attributes, and thresholds of the algorithm.
6
The Unattributable
Strategies for a Cloud Ethics

I return to the question of apartheid. It is exemplary for the questions of

responsibility and for the ethical-­political stakes that underlie this discussion. . . .
This is the moment of strategies, of rhetorics, of ethics, and of politics. Nothing
exists outside context, as I have often said, but also the limit of the frame or the
border of the context always entails a clause of nonclosure.
— Jacques Derrida, Limited Inc.

A description must capture not only the objects contained in an image, but it
must also express how these objects relate to each other as well as their attri-
butes and the activities they are involved in.
— Oriol Vinyals et al., “Show and Tell: A Neural Image Caption Generator,”
my emphasis

“Show, Attend and Tell”

At a major international machine learning conference in 2015, a group of

Google computer scientists presented their model for addressing what was, at
that time, one of the most difficult and important problems for deep learning:
“scene understanding.”1 The algorithmic model they called “show, attend and
tell” is one of several experiments in the use of generative neural networks for
representing and interpreting what is taking place in a scene.2 The machine
learning model combines two sets of algorithms (convolutional neural net-
works, or cnns, and recurrent neural networks, or rnns) that learn together
how to generate a description of the most likely thing of interest in a scene, its
attributes, and its future actions. Giving an account of their “attention frame-
work,” the designers explain how the first set of algorithms — cnns for extract-
ing the feature vectors of a scene — “ distill information in an image down to the
most salient features.”3 In this process of distillation, the algorithms are learn-
ing how to “infer scene elements” in a crowded data array and to foreground
what is considered to matter among the relations of people and objects.4 This
foregrounding is a mode of attentiveness that simultaneously deploys its con-
textual data environment to learn and discards that context as background
noise. In this sense, it is a form of modern attention that necessarily also in-
volves practices of distraction and annulment.5 Once the first neural networks 155
have extracted the feature vectors for what matters in the scene, a second set

The Unattributable
of algorithms — r nns designed for natural language processing — then act on
the output of the first to “decode those representations in natural language
sentences.”6 Working together in collaboration, the algorithms are said to gen-
erate outputs deciding “not only the objects contained in an image,” but also
how they “relate to each other,” what their essences or “attributes” are, and
what probable “activities” they may be “involved in.”7 The training images of
the algorithms — supplying the features of the recognizable — thus display a cu-
rious and compelling politics of showing and telling for our times.
Let us reflect on how these algorithms are condensing the question of
what matters, “inferring the scene,” as the computer scientists describe it,
and “mimicking the remarkable human ability” to filter out “the clutter” and
attend to the thing of interest.8 In figure 6.1, a group of people gathers in a
marketplace, their relations uncertain and indeterminate. One cannot pos-
sibly know from the image how these bodies relate one to another, what their
qualities or attributes might be, and especially why they are there. Indeed,
perhaps the incalculability and unknowability of their relations is the defin-
ing character of the urban scene, the crowd, and the marketplace. And yet,
from the multiplicity of relations in the scene — the people, the vegetables, the
earth, the clouded backdrop of market stalls — the neural nets cluster the data
and define the attributes necessary to generate a single condensed output of
meaning: “A group of people shopping at an outdoor market. There are many
vegetables at the fruit stand.” The algorithms generate the field of meaning;
they are the mise-­en-­scène of the event; they decide what is to be in the sin-
gle output signal. Understood in this way, there is indeed nothing outside the
text in the neural net, the context supplying the feature space from which all
possible future attention can be extracted, all meaning derived. Simultane-
ously, though, as the algorithms mistakenly infer the “vegetables” at the “fruit
stand” from all the past data elements of images and video of markets they
have been trained on, the “outside penetrates and determines the inside,” so
that the borders of context entail also an opening, or nonclosure.9 As the al-
Figure 6.1  The show and tell model for scene understanding, comprising
two algorithms: a convolutional neural network for image recognition and a
recursive neural network for natural language generation. LeCun, Bengio, and
Hinton, “Deep Learning,” 2048.

gorithm travels from its laboratory site of test images of South Asian markets,
how might the presence of a particular object in the scene — a turban, some
plantain, a pile of baskets — imply the attributes for future instances in Middle
Eastern marketplaces? If the Google computer scientists’ show and tell algo-
rithms are animating the scene understanding of the US drone program, what
action is authorized based on attributes?10 When algorithms learn the “latent
alignments,” what happens to that which is considered to be dormant or lying
hidden within a scene? What is a normal alignment in the marketplace, the
border landscape, or the city plaza?
The techniques deployed within the neural net algorithms to condense
the features of a scene to a single output — “a woman is throwing a Frisbee in
a park,” “a stop sign is on a road with a mountain in the background” — give
a compelling account of the ethicopolitics of algorithms for our times (figure
6.2). The output of the algorithms reduces the intractable difficulties and du-
ress of living, the undecidability of what could be happening in a scene, into
a single human-­readable and actionable meaning. We have ethical and politi-
cal relationships with other beings in the world because the meaning of those
relations, their mediation through every scene of life, cannot be condensed.
They are precisely irreducible. And so, at the moment that the algorithm out-
puts a single meaning from an irreducible scene, at this border limit is also a
“clause of nonclosure,” as Derrida describes the opening of context.11 The ties
that bind words, concepts, images, and things are only ever relatively stable, so
that there is always also an opening. How does one begin to locate the points
of nonclosure within the algorithm’s program of meaning making? If a cloud
ethics is to be concerned with the political formation of relations to oneself
and to others, how might it locate alternative critical relations of showing and
telling? Are there counter-­methods of attention available to us that could re-
sist the frameworks of attention of machine learning? Amid the technologies
of the attribute, what remains of that which is unattributable in the scene?
In this concluding chapter, I am concerned with how a cloud ethics could 157
be brought into being in a world that is increasingly shaped through the ar-

The Unattributable
rangements of the machine learning algorithm. Though I do not propose a
definitive method for critique or for resistance to the prevailing powerful log-
ics of surfacing, targeting, or “show and tell” in machine learning, I do urge

Figure 6.2  The show, attend, and tell model for scene understanding, display-
ing the test image on the left of each pair, with the machine’s “attended” vision
on the right. The target output is underlined. Xu et al., “Show, Attend and Tell,” 2.
 that it is possible to confront the specific fabulatory functions of today’s algo-
rithms with a kind of fabulation of our own. “It is the task of the fabulatory
function,” writes Deleuze, “to invent a people,” to write into being a collective
or a people who are otherwise “missing.”12 Now, of course, algorithms engage
in fabulation — they invent a people, write them into being as a curious body of
correlated attributes, grouped into clusters derived from data that are them-
selves fabulatory devices. Indeed, the fabulation of the algorithm involves a
158 form of iterative writing, a collective enunciation of which we are all a com-
posite part. To consider the writing of algorithms, as I argue in this book, is to
Chapter Six

locate the ethicopolitical acts of relating narratives of oneself and others. So,
the methodology of a critical cloud ethics must also fabulate, must also neces-
sarily invent a people whose concrete particularities are precisely unattribut-
able, never fully enclosed within the attribute and its output. To fabulate in
this way means to take the arrangements of the algorithm (in this instance,
the show and tell neural nets) and imagine their mode of existence beyond the
decisive point of the target output. As method, a cloud ethics engages fabu-
lation in that it does not anchor narrative in the authorial source but rather
“digs under stories, cracking them open,” so that the reader might enter into
the space of the becoming-­political of the algorithm.13 What this means is that
one cannot stand outside the algorithm to judge its morality, its role in doing
good or evil. Instead, one must begin from the iterative writing that is itself
generative of fungible thresholds of the good and the bad.
Just as computer science has become comfortable with experimental fabu-
lation and narrative, it is necessary for others to play with the arrangements,
thresholds, and assumptions of algorithms. The computer scientists are, for
example, apparently comfortably drawing on philosophy and ethics in their as-
sertions that machine learning contains within it a capacity for something like
machine reasoning.14 Methodologically, we social scientists, and humanities
and ethics scholars, must also find ways to engage experimentally in algorith-
mic fabulation. Let us begin by digging under the narratives of the show and
tell algorithms, cracking them open to find the moments of nonclosure, writ-
ing them in ways that lever open their ethicopolitical orientations. Put simply,
as algorithms work to derive future propensities from collections of attributes,
so we too must curate the attributes of the algorithm to signal its propensities
to colonize the unattributable future. A fabulatory method would take the ar-
rangements, perceptions, and accounts of the algorithm and rework them to
deepen the political intensity of what it means to output something into the
world (figure 6.3):
 159

The Unattributable
Figure 6.3  The errors in the algorithm are used to gain intuition into what
the model saw and how it perceived the scene. Xu et al., “Show, Attend
and Tell,” 5.

A woman is throwing a Frisbee in the park.

A woman is holding a child at the border fence.
A stop sign is on a road with a mountain in the background.
A placard is held in Tahrir Square with an unauthorized crowd in the foreground.
A man is talking on his cell phone while another man watches.
A man is opening his rucksack while another man watches.
A group of people shopping at an outdoor market. There are many
vegetables at the fruit stand.
A group of people running at an outdoor market. There are many weapons in some
hands.
 To advance a cloud ethics is to work with the propensities that the algo-
rithm embodies.15 In place of an ethics that seeks to make an illegible algo-
rithm legible to the world, a cloud ethics recognizes the nonclosure in all forms
of writing and pushes the fabulation of the algorithm beyond what can cur-
rently be read of its logic. The single output of the algorithm is reopened and
reimagined as an already ethicopolitical being in the world. What could this
small adjustment in the weighting of the cnn do to the language of the output
160 meaning? What is a “salient feature,” and how does it come to the forefront of
the crowded scene? In the following sections, I address the potential strategies
Chapter Six

of a cloud ethics across three aspects: apertures, opacity, and the unattributable.

Apertures

The computer science problem of how to understand the space of a scene — to

show, to attend, and to tell what is happening — is not limited to the genera-
tion of meaning from images. Indeed, for machine learning algorithms, all
space is feature space. That is to say, whether the scene is aerial video from a
drone or images from an mri scan, the algorithm acts on a feature space un-
derstood as the n-­dimensions where data variables reside. The neural networks
that extract the salient features from images, for example, do not meaningfully
differentiate the sources of their input data. At the level of the algorithm, the
image-­derived input data that generate a given meaning (e.g., “This is a cat,” or
“A placard is held aloft in the city square”) take the form of numeric values for
the presence or absence of edges in the pattern of pixels.16 It would make little
technical difference if the input data to the algorithm included other forms
of data, such as the sentiment analysis of social media text, travel data, or cell
phone location data. The neural network converts all these data into gradients
or feature vectors that can then be recognized as similar to, or different from,
the feature vectors present in other known scenes or situations. What mat-
ters to the algorithm, and what the algorithm makes matter, is the capacity to
generate an output from the feature vectors, to be able to “tell” what is latent
in the scene. For example, when Facebook supplied their users’ data to Cam-
bridge Analytica for the generation of models of voter propensities, they sup-
plied the feature vectors necessary to train algorithms to “tell” what is latent
in the democratic political “scene.”17 It is precisely this emerging alliance — 
between algorithms hungry for new feature spaces to yield attributes, and a
modern politics anticipating the latent attributes of populations — that poses
one of the principal ethicopolitical challenges of our times. What is at stake
is not only, or not primarily, the predictive power of algorithms to undermine
the democratic process, or to determine the outcome of an election, a judicial
process, a policing deployment, or an immigration decision. Of greater signifi-
cance even than these manifest harms, algorithms are generating the bounded
conditions of what a democracy, a border crossing, a protest march, a deporta-
tion, or an election can be in the world.
I do not wish to understate the potential that algorithms have for profound
cruelty and violence. Indeed, when I say that the harms of algorithms are not
principally located in the much-­publicized invasions of privacy or absence of
transparency, I am identifying a profound violence in the algorithm’s foreclo- 161
sure of alternative futures. This is a different kind of crisis of democratic poli-

The Unattributable
tics, one in which deep neural networks do not merely undermine the already
decided functions of democracy (“decide the US president,” or “decide to leave
the European Union”) but actively place a limit on the potential for any other
form of political claim to be made in the future. What new entity or alterna-
tive model of citizenship could come into being after the algorithm has opti-
mized all outcomes? What new political claim, not yet registered as claimable,
could ever be made if its attributes are known in advance? These are not only
philosophical questions but also practical ones with a real purchase on the
politics of our times. Given the machine learning articulations I have followed
in the researching of this book, one would have to contemplate, for example,
whether a set of algorithms designed for the US presidential election actually
learned to recognize the attributes of “clusters” of anti-­immigration and na-
tionalist sentiment via the “ground truth” data of the attributes of UK citi-
zens who voted to leave the European Union. In short, the “salient features” of
groups within the UK EU referendum may have gone on to have specific and
violent ethicopolitical relations to future families who find themselves sepa-
rated at the US-­Mexico border.
What can a society be in the future if an algorithm is optimizing an output
that imagines the resolution of all future claims? The multiplicity of potential
pathways within the neural net — millions of potential alternative connections
and correlations — is reduced to one, condensed to a single output. Understood
in this way, contemporary algorithms are simultaneously constructing the ar-
chitecture of the problem space (What should the state do at the border? How
ought a society decide creditworthiness? What should a police force attend
to?) and generating the solution to the problem. “The problem always has the
solution it deserves,” writes Deleuze, so that the solution (output) is generated
by “the conditions under which it is determined as a problem.”18 If one under-
stands politics as expressing the impossibility of resolving all problems with a
solution, then algorithms appear to circumvent the political, because they ad-
just the parameters of the problem to experiment abductively with the “good
 enough” solution. Alternative political arrangements become more difficult to
imagine, in part because — reversing Deleuze’s formulation — the solution al-
ways gets the problem it deserves; that is, the propositional arrangements of
the algorithm (the model of the problem) are immanently adjustable in rela-
tion to the target output (solution). If, for example, a cnn “shows” that the en-
tity “vehicle” extracted from a video feed has a 0.65 probability of being part of
a military convoy, and then an rnn “tells” that this is an actionable threat, at
162 the narrowed aperture of the output there is no place for a 0.35 probability of a
civilian bus in the scene. How could there be an alternative relation, a nonclo-
Chapter Six

sure, when the output has so intimately and abductively generated the model?
A strategy for cloud ethics would dwell for some time with the aperture of
the algorithm, the point where the vast multiplicity of parameters and hidden
layers becomes reduced and condensed to the thing of interest.19 To be clear,
the aperture is not a matter only of vision or opticality but, as Jonathan Crary
has detailed in his histories of perception, is specifically a means of dividing,
selecting, and narrowing the focus of attention.20 The aperture of perception
matters to my cloud ethics because it is a point of simultaneous closure and
nonclosure. It is a closure in the sense that there is no outside of the algorithm,
so that the weights and the outputs interpenetrate and mediate one another.
The aperture of the algorithm is an ethicopolitical closure of all the alternative
possible values, assumptions, and connections that the hidden layers may con-
tain. And yet, at the limit of the aperture is also an opening onto a moment of
decision that is always already fully ethicopolitical. From its roots in the Latin
aperire, or to open, the aperture is an opening or breach through which light
can enter. Understood as a form of aperture, the algorithm simultaneously fil-
ters down and closes off as it opens a gap or a breach. What this means in terms
of a strategy of cloud ethics is that one can envisage attending to the teeming
multiplicity of all past moments that have yielded their probabilities to the
model. A cloud ethics opens up the space of condensed output and makes the
decision difficult and freighted with the traces of rejected alternatives. The
aperture is the moment at which space is divided and the multiplicity is re-
duced to one, and yet it must also contain within it the traces of the opening
onto lost pathways. A cloud ethics alters the relation to a future that says “here
is a good enough solution” by revaluing the aperture as a moment of decision
in the dark, a calculation of the incalculable. The strategy involves reopening
the multiplicity of the algorithm, digging under the stories, and attending to the
branching pathways that continue to run beneath the surface.
How does one begin to locate the nonclosure in the arrangements of the
algorithms I have discussed in the chapters of this book? When I have given
lectures on cloud ethics, a number of generous computer scientists have asked
questions from the audience, urging me not to pursue the notion of revisiting
the weightings within a machine learning algorithm. For them, the adjustment
of weights within the hidden layers is an impenetrable process that retains its
opacity even to those who undertake it. “You cannot make the weights politi-
cal, Louise,” they have cautioned me, “because they are not really a thing. We
don’t know how they work. We are just messing around with them.” Yet, it is
exactly this kind of opaque, messy, and embodied relation to the weightings of 163
the algorithm that interests me. Moreover, the sense of nonknowledge — where

The Unattributable
the designers of algorithms “don’t know how they work” — is fertile ground for
insisting on a cloud ethics that has to decide in the darkness of not knowing.
Put simply, my cloud ethics would make the adjustment of a weight in a neural
net weightier, heavier in the sense of carrying the burden of difficulty and in-
transigence. Such a heaviness of the weighting of probabilities could shift the
ethical focus from the spatial arrangement of combinatorial possibilities to a
temporal and durational force of becoming. The weighting within one layer of
an algorithm becomes not only a spatial weight of probability but also a tem-
poral weight of duration. There is a kind of injunction here: the algorithm must
carry the weight of its weightings. As Deleuze has reflected on Bergson’s famous
division between space and duration, “Take a lump of sugar: it has a spatial
configuration. But if we approach it from that angle, all we will ever grasp
are differences in degree between that sugar and any another thing.”21 Via the
spatial differences of degree (via gradients), an rnn, for example, generates
the “attend” and “tell” outputs, deciding that “this is a woman on a sidewalk”
because of its spatial distance from any other thing. Returning to Deleuze’s
account of Bergson’s sugar cube: “But, it also has a duration, a rhythm of du-
ration, a way of being in time that is at least partially revealed in the process
of its dissolving, and that shows how this sugar differs in kind not only from
other things, but first and foremost from itself. This alteration, which is one
with the essence or substance of a thing, is what we grasp when we consider it
in terms of Duration.”22
It is possible to reimagine the attributes of something — “the essence or
substance of a thing” — that the machine learning algorithm derives entirely
from spatial clusters as durational qualities that differ in kind from themselves
and others. The algorithm’s generation of attributes carries only the lightness
of likelihoods, of differences of degree. A cloud ethics freights the machine
learning algorithm with the full durational weight — the waiting, the hesita-
tion, the doubtfulness — of the essence or substance of a thing. In this way, the
weight could be made to count as a meaningful moment of ethicopolitical de-
 cision. And so, when an algorithm for immigration decisions is said to weight
the data inputs from past visa refusals, for example, the durational weight
would surface a different kind of attribute, something unattributable. All the
past moments of waiting in line (the wait in the weight), every past gathering
of an incomplete visa application, all the persistent indeterminacies that differ
in kind from themselves — these are the materials of a cloud ethics.
In revisiting the weights of algorithms, I am proposing that there are mo-
164 ments of undecidability within all technologies that enter the world as algo-
rithmic decision systems. With the experience of undecidability comes also a
Chapter Six

mood of doubtful hesitation at the opening of the aperture.23 Such hesitation

is crucial to the forging of new ethicopolitical arrangements, because it ex-
presses the weight of undecidability that is never eradicated entirely by the al-
gorithm’s output. I cannot know in advance the full consequences of what will
be decided, but nonetheless a decision will be made. It seems that contempo-
rary algorithmic decisions are precisely seeking to lighten the ethicopolitical
load of undecidability. They must “feel the weight” of their decisions, which is
a matter of not only weighting alternatives in a calculation but also weighing
what is risked of ourselves and our relations to others in that calculus.24
With cloud ethics, what is condensed of the future by algorithms becomes
something materially heavier and more burdensome, a kind of residue in the
present. All algorithmic decisions contain within them the residue, or the sedi-
ment, of past political weightings. As we saw with the racialized policing and
security algorithms with which I opened the introduction to this book, the
condensed future (“he will pose a risk to the crowd”) contains within it the
residue of all the violence of past colonial histories, migrations, journeys, and
border crossings, a fulsome sediment of all the actions and transactions of
past movements in the name of justice. And in this residue one can trace the
fork in the road that was not taken, the “trace of the rejected alternative” that
continues to lodge itself within the cracks of the machine’s capacity to learn.25

Opacity

Amid the widespread moral panic surrounding the black box of the algorithm
and the political demands for transparency, accountability, and explainabil-
ity, in this book I make a counter case for opacity and the giving of partial ac-
counts. I have argued that responsibility for algorithmic decisions can never
take the form of a clear-­sighted account of the action. “Responsibility,” as
Thomas Keenan writes, “is not a moment of security or cognitive certainty”
in which one makes a “choice between yes and no, this or that.”26 To estab-
lish a moment of cognitive certainty regarding the algorithm — to demand its
transparency and accountability — is not at all the politics of a responsibil-
ity worthy of the name. Indeed, with its new forms of cognitive computing
and cloud reasoning, as I discuss in chapter 1, the contemporary algorithm is
well attuned to generating clear-­sighted precise outputs from otherwise oc-
cluded situations. Contra the demand for transparency, then, the opacity of
the scene, and the undecidability of its salient features and attributes, is the
precondition for politics and ethics. “What could responsibility mean,” asks
Keenan, “without the risk of exposure to chance, without vulnerability,” with- 165
out the undecidability of what is taking place in a scene?27 Running against the

The Unattributable
grain of algorithmic logics that show, attend, and tell the meaning of a situa-
tion, responsibility means that one cannot see a clear path ahead because the
meaning of the situation is undecidable. The multiple branching points of the
decision tree or random forest algorithm do in fact carry the risk of exposure
to chance and vulnerability. A different kind of responsibility dwells at these
forks in the path, bearing the opacity of future consequences of the weighting
of one route over another.
In this book I differentiate the search for an encoded ethical framework
for algorithms from an ethicopolitics that is already present within the algo-
rithm’s arrangements. In so doing, I draw a distinction between a moral code of
prohibition and permission (what algorithms may or may not do) and an ethics
of the orientation one has to oneself and to others (how algorithms are generating
relations of self to self, and self to others).28 This distinction follows Spinoza’s
broader philosophical concern that “ethics has nothing to do with a moral-
ity.”29 Our ethicopolitical relations do not follow strictly from some already
formulated moral code, for we never have the security of these grounds from
which to act. As Deleuze has reflected on Spinoza’s ethics, “You do not know
beforehand what good or bad you are capable of; you do not know beforehand
what a body or a mind can do, in a given encounter, a given arrangement, a
given combination.”30 Similarly, for Isabelle Stengers, “every time we use the
term ‘ethics’ we must obviously distinguish it from the term ‘morality,’ ” so
that her speculative ambition is ethical in the sense of ethos or habit.31
Just as Spinoza’s ethics are ever in formation, forged through ethos, en-
counters, arrangements, and combinations, so a cloud ethics is located in the
encounters, arrangements, and combinations through which the algorithm is
generated. Some of these encounters and arrangements involve human beings;
others are between unsupervised algorithms and a corpus of data; and more
still involve algorithms interacting with other algorithms. This is an ethics,
then, that puts into question the authority of the knowing subject and opens
onto the plural and distributed forms of the writing of algorithms. Understood
 in this way, “ethics is able to exceed its traditional determination” and be “dis-
placed” and considered “anew” in the relations forged with other beings.32 The
possibility of what Derrida calls “the opening of another ethics” is a different
kind of opening to the imagination of the prizing open of a black box or the il-
lumination of a dark calculus with moral oversight.33 It is an opening onto the
edges and limits of an iterative writing that is profoundly political, each one of
us drawn into the proximities of the output. Algorithms engage in writing in
166 and through their dense relations with people and things in a data scene, their
authorship uncertain and their emergence opaque. Here, in the opacity of not
Chapter Six

knowing in advance what the relations could be, we have found an opening for
taking responsibility.
If an ethical response to algorithmic decisions is to begin from opacity
and not from the demand for transparency, then what does this mean for the
strategies of a cloud ethics? Where the political demands for transparency re-
assert the objective vision of a disembodied observer, to stay with opacity is to
insist on the situated, embodied, and partial perspective of all forms of scien-
tific knowledge.34 Understood in Donna Haraway’s terms, to confront opacity
would not be a matter of failing to open the black box or to see something, but
rather it would acknowledge that “the knowing self is partial in all its guises,”
that practices of visuality necessarily enroll the opacity of ourselves and oth-
ers.35 Far from a lack of transparency limiting accountability, the condition of
opacity “allows us to become answerable for what we learn how to see.”36 This
“we” who learns, as I discuss in the machine learning practices of chapter 2, is
a composite figure, “stitched together imperfectly” from the iterative learning
of humans and machines.37
In the chapters of this book, I have been concerned to show how machine
learning algorithms generate their outputs through the opaque relations of
selves to selves, and selves to others, and how these outputs then annul and
disavow the opacity of their underlying relations in the name of a crystalline
clarity of the output. One mode of resistance to the clarity of the algorithmic
decision resides within the opacity of the subjects and what cannot be known
of their relations. Let us consider a practical example of such a strategy. In the
experimental development of the neural networks for show, attend, and tell,
the algorithm designers explain that when their model makes a mistake, they
are able to “exploit” what it manifests to gain “intuition into what the model
saw.”38 This kind of seeing on behalf of the algorithm sustains the god’s-­eye vi-
sion of a computer scientist who may redistribute their way of seeing at will. A
cloud ethics can stay with the difficulty of intuiting what the model saw as a
form of partial, situated, and locally specific account. Among the examples of
the mistakes used to productively generate new learning is “a woman holding
a clock in her hand” and “a man wearing a hat and a hat on a skateboard.” The
algorithms have learned to recognize a clock and a hat, even when these are
absent in the scene, via their exposure to a corpus of data that has supplied the
feature vectors for these objects. The algorithm’s relations with a contingent
and opaque set of multiple other gradients have generated the mistaken pres-
ence of a clock and a hat (figures 6.4 and 6.5). 167
The cnns have generated an output of attentiveness to the focus of the

The Unattributable
scene (shown in the images as an area of white light) — this output itself con-
tingent on millions of past parameters — which enters the rnn as a data in-
put. The rnn then generates the “tell,” or the natural language description
of the scene, assigning a probability for each word in the descriptor. In the
(mis)recognition of “A man is talking on his cell phone while another man
watches,” the algorithm has assigned a probability of 1.00 (or a certainty) to
the indefinite article A, for example, but probabilities of 0.46, 0.24, and 0.17
to the nouns man, cell, and phone. These likelihood scores — low and yet appar-
ently the most probable — are entirely contingent on multiple algorithms’ past
relations with unknown other scenes. As I explain in chapter 4, the mistaken
recognition of the algorithm is not merely error or errancy, for the output is
entirely reasonable given the algorithm’s relations with a data world. In terms
of resistant practices, a cloud ethics is concerned not with fixing the algorithm
or eradicating its errancy, but rather with attending to how these moments
give accounts of algorithmic reason. What will take place when the recogni-
tion of a clock that is not present becomes the recognition of a weapon? What
will be authorized by the most likely cell phone that is an absent presence in
the scene?
To be clear, a cloud ethics will point to the moments when partial and sit-
uated algorithmic reason is sheltered as “mistake” or “error” and will reopen
what is incomputable in the making of a likelihood score. Such a strategy is
not the same as making a calculation transparent or accountable because it
recognizes that the 0.14 assigned to child or 0.46 and 0.37 to border and fence
are contingent on a vast multiplicity of necessarily opaque relations between
other people, algorithms, things, data, gradients, features, and images. When
the algorithm “tells” us that this is what matters in a scene, one could never
have a transparent account of how the story has been made tellable. Instead,
there are only the intransigent political relations that persist in the scene, the
moments of nonclosure that reopen the problem and the politics.
The Unattributable

Contemporary machine learning algorithms operate with a specific concept

of what it means to be together or to have things in common with others. In
the ontology of algorithms, the mechanism that connects people, entities, and
events is the attribute. The sphere of society and social relations thus exists for
algorithms only as a series of groups, or clusters, defined by the spatial prox-
imities and distances of their attributes. To govern a population through its 169
attributes involves a rather different imagination of the relationship between

The Unattributable
the universal (e.g., shared values or moral principles, ideas of economy, com-
mon laws) and the particular (that which is particular to the subject as indi-
vidual) than that which characterizes accounts of neoliberalism or biopolitics.
The machine learning algorithm iteratively moves back and forth between the
ground truth attributes of a known population (known in terms of the at-
tribution of qualities) and the unknown feature vector that has not yet been
encountered. In the space between a computer science universal of attributed
feature space and the particularity of the new entity, the algorithms learn how
to recognize and describe what is happening in the world. Thus, for example,
the arrested person in the custody suite is rendered knowable by algorithms
that calibrate their feature vector against the risk-­scored attributes of another
population — these encounters unrelated except via this new relation of attri-
bution and adjudication.39 Likewise, when a group of teenagers try to board a
bus to attend a protest against the killing of a young African American man
in their city, their political rights to gather with unknown others are annulled
by algorithms that render them knowable via the clustering of their shared risk
attributes.40 What kind of relation to oneself and to others is manifest in this
curious relation of known attributes to unknown others? How can the algo-
rithm’s dissolution of other forms of being together be resisted? How might a
cloud ethics find political spaces for gatherings and collectivities that are pre-
cisely unattributable?
In the opening lines of this chapter, I cite Derrida’s remarks on the ethi-
copolitical stakes involved in resisting apartheid. “This is a moment,” he re-
flects, “of strategies, of rhetorics, of ethics, and of politics,” a moment when it

Figures 6.4 and 6.5  Neural image caption generation

with visual attention. The two examples display inaccurate
outputs, with each word in the descriptor assigned a likeli-
hood score. Xu et al., “Show, Attend and Tell,” 14, 16.
 is imperative to find the nonclosures at the limits of the border.41 It seems that
the struggles of our contemporary moment — resisting the rise of far-­right na-
tionalist groups, racist doctrines of anti-­immigration, the everyday violences
of sexism, racism, and bigotry — confront also the profound difficulties of find-
ing an opening at the limit of the frame. Indeed, many of our most impor-
tant and difficult historical political struggles — antiapartheid movements, the
civil rights movement, campaigns for lgbtq+ rights — would arguably have
170 been impeded by state access to algorithms that could learn to attribute future
threats. In short, machine learning algorithms are actively making it more
Chapter Six

difficult for new ethicopolitical claims — those claims not already registered

as claimable — to be made in the world. In this context, a cloud ethics must
be able to locate ways of being together that resist the algorithmic forces of
attribution.
In an attributive society, what is perhaps most imperiled and at risk, and
yet also most fugitive and precious to ethicopolitical life, is that which is un-
attributable. The unattributable — or a potentiality that cannot be attributed
to a unitary subject — matters to ethics and politics because it does not locate
its grounds in a clear I who makes a claim, but rather remains open to other
claims not yet made. When machine learning algorithms segment a social
scene, generating clusters of propensities, for example, to vote for a particular
political party, or to respond to targeted social media, or to pose a high risk
of recidivism, everything must be attributed, even the outliers understood as
distant gradients from the curve of normality. The unattributable remains
within the scene, however, exceeding the algorithm’s capacity to show and
tell, and opening onto a different kind of community and a different mode of
being together. The attributive function that is the animating device of ma-
chine learning is, in fact, a more slippery and open device than its role in com-
puter science might suggest. A cloud ethics will call this to mind each time it
engages with the algorithm’s showing and telling: the attribute is a cluster of
features that remain persistently unattributable to any subject as such. There
is a gap, a breach, an opening between the output of the algorithm and the
body of the subject in whom it becomes actualized, at the border, in the police
custody suite, at the employment center, on the city street.
The resistant orientation to attribution I have in mind does not stand out-
side or in opposition to the attributive power of the algorithm. In common
with all other forms of resistance, this is an orientation that shares the same
horizon as the mode of power. Indeed, I have explained in this book the extent
to which human subjects dwell within the algorithm, as it comes into being
also through relations with subjects and objects. What is necessary is not a re-
sistant overturning of the logics of algorithm but a fuller and more amplified
attunement to the ethicopolitics that are always already present in algorithmic
reason. Thus, I envisage a cloud ethics that takes seriously the governmental
form of the attribute and dwells with it. Yet, a cloud ethics must also refuse
the attachment of attributes to entities in favor of the immanence of all forms
of writing and attribution. As Deleuze has captured the attribute, following a
Spinozan ethics, it expresses “an essence that is an unlimited, infinite quality”
so that a “single substance” has an “infinity of attributes.”42 The unlimited 171
and infinite quality of the attribute offers a resistant breach in the algorithm’s

The Unattributable
society of attributes, for the algorithm’s attribute must always be a finite and
fixed essence to be classifiable and recognizable. A cloud ethics recognizes, in
contrast, the unlimited and infinite emergence of attributes that will defy the
boundaries of clusters and feature vectors. Put simply, the attribute is unat-
tributable to those on whom its effects are exerted.
There is a political refusal involved in claiming the unattributable. It is a
refusal to be governed by the attribute, an affirmation of “how not to be gov-
erned like that,” as Foucault has described it, a certain “will not to be governed”
as a cluster of attributes.43 My ethicopolitical relations to myself and to others,
then, cannot be limited to recognizable attributes but must begin from a rec-
ognition of singularity. As Simon Critchley writes, “Ethics is first and foremost
a respect for the concrete particularity of the other person in his or her singu-
larity.”44 This means that the other person cannot simply be an “example” of
some set of finite qualities derived from the gradients of norm and anomaly of
others. “Ethics begins as a relation with a singular, other person,” continues
Critchley, this singular other person who then “calls me into question” so that
“politics begins with ethics.”45 As the algorithm presents us with an attributive
other who is never singular or particular, it is crucial that we refuse the para-
digm of attribution and amplify the unattributable. I cannot know the essence
of the other as a cluster of attributes, and when the algorithm presents me with
the likelihood of their attribution, I must begin again from their singularity.
The question of an ethics of the unattributable is not only a matter of
thought or of the philosophy of algorithms in society. It must also be a mat-
ter of the practice of critique and the situated struggle for alternative routes
that are foreclosed in the calculation of an output. There have been many oc-
casions when I have asked the same question of border and policing authori-
ties, newly equipped with their predictive recognition algorithms: “How will
you know the characteristics of this unknown person who appears?” Always,
the response has been some variant on the same notion: “We do not need to
know them in particular; we have seen their attributes before.” The unattrib-
 utable is thus at the very heart of the struggles over the place of algorithms in
our societies. A cloud ethics must pursue a conception of being together as a
society of the unattributable, where what we share in common is precisely our
singularities, and where our encounters with others can never be limited to a
finite collection of attributes.46 All the many moments when algorithms gather
to cluster our attributes, to assemble against us, when they seek to silence the
claim before it could even be made — these are the moments that must remain
172 open to what cannot be attributed. Ethicopolitical life is about irresolvable
struggles, intransigence, duress, and opacity, and it must continue to be so
Chapter Six

if a future possibility for politics is not to be eclipsed by the output signals of

algorithms.
NO T E S

Introduction
1 Observations of tech start-­ups’ pitches for government procurement of algorith-
mic systems, London, March 2016. An earlier iteration of the algorithms is de-
scribed in a 2012 video, “Monitoring Protests from Unstructured Text.”
2 Vice News journalist Jason Leopold made a Freedom of Information Act request
seeking all records pertaining to Freddie Gray and the subsequent protests and
riots in Baltimore, Maryland. For the full disclosed documents, see “dhs foia
Documents Baltimore Protests Freddie Gray.”
3 For the text of the Geofeedia case study, see aclu, “Baltimore County Police De-
partment and Geofeedia Partner.”
4 Computer scientists give accounts of how they use samples of training data to
train their model to recognize and extract text when it appears in an image.
See, for example, Misra, Swain, and Mantri, “Text Extraction and Recognition,”
13 – 19.
5 Browne, Dark Matters, 26.
6 Laughland, “Baltimore Unrest.”
7 Though many specific algorithms become secret and proprietary when bought
and licensed by corporations and governments, the upstream science of their
genesis is much more public, incremental, and debated. Methodologically, a de-
gree of speculation is necessarily involved in identifying which broad family of
algorithms is being used in a particular application. Among the many classic
computer science papers on contemporary machine learning algorithms I have
used to infer the operations of “secret” or “proprietary” algorithms are LeCun,
Bengio, and Hinton, “Deep Learning,” 436 – 4 4; Hinton and Salakhutdinov, “Re-
ducing the Dimensionality of Data,” 504 – 7; Krizhevsky, Sustskever, and Hinton,
“ImageNet Classification”; Simonyan and Zisserman, “Very Deep Convolutional
Networks”; and Hinton, “Deep Neural Networks for Acoustic Modeling,” 82 – 97.
8 “Amid the uncertainties, potentialities, and dangers created by the regime of
computation,” writes N. Katherine Hayles, “simulations — computational and
narrative — can serve as potent resources with which to explore and understand
the entanglement of language with code . . . and subjectivity with computation.”
Hayles, My Mother Was a Computer, 242.
9 Clustering techniques in machine learning are experimental and exploratory in
the sense that they discover previously unknown structure in a large volume of
data. The algorithms iteratively build a model by grouping the data points that
are most similar in their attributes. The model is then used to predict likely be-
havior in the future, such as the propensity for a group to “churn,” or to switch
commercial providers of, say, energy, credit, insurance, or telecommunications.
The aim in clustering, as Ethem Alpaydin describes, “is to find clusters of input”
data and to define each group as “similar in their attributes.” Alpaydin, Machine
 Learning, 112. A set of algorithms designed for customer segmentation models are
also widely used for security and policing. For example, computer scientists have
trained machine learning algorithms on Twitter data to determine the attributes
of support for the Islamic State, training their classifier to “predict future sup-
port or opposition of isis [Islamic State of Iraq and Syria] with 87% accuracy.”
Emerging Technology from the arXiv, “Twitter Data Mining”; Caruso, “Can a
Social Media Algorithm Predict a Terror Attack?”
10 Crawford, “Artificial Intelligence’s White Guy Problem.” See also John Markoff’s
174 interview with Crawford on the ethical conduct of the designers of machine
learning algorithms; Markoff, “Artificial Intelligence Is Far from Matching Hu-
mans.” See also Metcalf, Keller, and boyd’s report on the ethical design of big
Notes to Introduction

data analysis, “Perspectives on Big Data.”

11 O’Neil, Weapons of Math Destruction.
12 Pasquale, Black Box Society.
13 In Michel Foucault’s 1981 series of lectures at the Catholic University of Louvain,
Mal faire, dire vrai [Wrong-­Doing, Truth-­Telling], he explains that his interest lies not
in establishing the basis or grounds of truth but instead in “the forms of the en-
terprise of truth-­telling” of particular scientific or technical claims. He proposes
a counterpoint to positivist claims in what he calls “astonishment before the pro-
liferation of truth-­telling, and the dispersal of regimes of veridiction.” This pro-
liferation of truth telling by algorithms, and the regimes of veridiction that make
it possible, is of interest to me in this book. Foucault, Wrong-­Doing, Truth-­Telling,
20 – 21.
1 4 Cheney-­Lippold, We Are Data, 9.
15 Connolly, “Beyond Good and Evil,” 366.
16 Michel Foucault, in addressing his genealogy of ethics, depicts a “great change”
from Greek society to Christian society dwelling “not in the code but in the eth-
ics,” which is “their relations to themselves and to others.” Foucault, “On the
Genealogy of Ethics,” 253 – 80. A related distinction between ethics as code and
ethics as practice can be found in the work of Gilles Deleuze, for whom the term
morality denotes a set of rules or codes, whereas ethics is immanent to a “mode of
existence,” or a form of life. Deleuze, Negotiations, 100; and Deleuze, “On the Dif-
ference between the Ethics and a Morality,” 17 – 28.
17 The “cloud layer,” as Bratton describes it, forms “the second from the bottom of
the Stack,” so that “cloud platforms may structure Cloud polities.” Bratton, The
Stack, 9, 369. See also Hu, Prehistory of the Cloud.
18 Peters, Marvelous Clouds, 2. Media theorist Jussi Parikka similarly directs atten-
tion to the material and geophysical making of media; Parikka, Geology of Media.
19 Butler, Giving an Account of Oneself, 12.
20 Butler, Giving an Account of Oneself, 17.
21 Amaro, “Race, Surveillance.”
22 Mathematician A. A. Markov defined the ideal algorithm as one that offered
“precision,” “definiteness,” and the “conclusiveness” of a desired result. Markov,
Theory of Algorithms, 1.
 23 Wittgenstein, On Certainty, §653 – 57.
2 4 Wittgenstein, Philosophical Investigations, §136. I am indebted to Jeremy Schmidt
for conversations on Wittgenstein and mathematics as praxis.
25 Wittgenstein, On Certainty, §655.
26 Wittgenstein, On Certainty, §651.
27 As Karen Barad has argued compellingly, “It is not merely the case that human
concepts are embodied in apparatuses, but rather that apparatuses are discur-
sive practices . . . through which ‘objects’ and ‘subjects’ are produced.” In Barad’s
reading, scientific apparatuses are agents that enact boundaries, becoming the 175
condition of possibility of claims to cause and effect. Barad, Meeting the Universe
Halfway, 148, 175.

Notes to Introduction
28 Foucault, “Ethics of the Concern for Self,” 299.
29 Seaver, “Algorithms as Culture,” 2.
30 Among the many depictions of the algorithm as a sequence of steps, for Tarleton
Gillespie, an algorithm is “a recipe composed in programmable steps.” Gillespie
“Algorithm,” 19. Pedro Domingos defines an algorithm as “a sequence of instruc-
tions.” Domingos, Master Algorithm, 1. David Berlinski identifies “an algorithm
[as] a finite procedure.” Berlinski, Advent of the Algorithm, xviii.
31 DeLanda, Intensive Science and Virtual Philosophy.
32 David Berry has distinguished between different algorithmic forms to show their
different implications for ordering and acting on society. Berry, Philosophy of
Software.
33 In Noel Sharkey’s assessment of the different juridical options for the control of
autonomous weapons systems, he writes that “we must urgently begin to develop
a principle for the human control of weapons that is founded on an understand-
ing of the process of human reasoning.” For Sharkey, the “reframing” of “auton-
omy in terms of human control” has the effect of making it “clear who is in con-
trol, where and when.” Sharkey “Staying in the Loop,” 23 – 38.
3 4 DeLanda, Philosophy and Simulation, 44 – 45.
35 For Totaro and Ninno the recursive logic defines “the logic of algorithms” and is
the “specific form in which the concept of function occurs in everyday life.” To-
taro and Ninno, “Concept of Algorithm,” 32. See also Beer, “Power through the
Algorithm?,” 985 – 1002.
36 The current state-­of-­the-­art design of neural networks for voice and image rec-
ognition, for example, is experimenting with teaching algorithms how to teach
themselves which hidden layers could be bypassed without significantly altering
the output. LeCun, Bengio, and Hinton, “Deep Learning.”
37 So-­called risk assessment engines powered by machine learning algorithms are
being used by judges to predict the likely probability of a convicted person reof­
fending if they return to the community. Research has shown that the algorithms
are overestimating the risk of future crimes by black defendants — and therefore
proposing prison as the optimal judgment — even where race is not an input da-
tum in the algorithm. See Angwin et al., “Machine Bias”; and Liptak, “Sent to
Prison.” The United Kingdom’s Durham Police Constabulary deploys the Harm
 Assessment Risk Tool (hart) to assess the likelihood of an individual reoffend-
ing if they are granted bail. See Bridge and Swerling, “Bail or Jail?”
38 Mackenzie, Machine Learners, 27.
39 Parisi, Contagious Architecture, 2.
40 Winslett, “Rakesh Agrawal Speaks Out.”
4 1 Amoore, Politics of Possibility, 50 – 52, 139 – 4 1.
4 2 Turing, “Systems of Logic Based on Ordinals,” 161–228. In letters sent to Max
Newman in 1940, Turing is “not sure whether my use of the word ‘intuition’ is
176 right or whether your ‘inspiration’ would be better” and considers that Newman
is “using ‘inspiration’ to cover what I call ‘ingenuity.’ ” Turing, “Letters on Logic
to Max Newman,” 212.
Notes to Introduction

43 Diamond, Wittgenstein’s Lectures, 150, 251.

4 4 Devlin, Math Gene, 110.
45 Stengers, Thinking with Whitehead, 413.
46 Intelligence and Security Committee of Parliament, Privacy and Security, 8.
47 Schneier, “nsa Robots.”
48 O. Halpern, Beautiful Data, 21.
49 Crary, Suspensions of Perception, 2.
50 Kittler, Discourse Networks 1800/1900; Kittler, Literature, Media, Information Systems.
51 “The whole difficulty of the problem,” writes Bergson, “comes from the fact that
we imagine perception to be a kind of photographic view of things, taken from
a fixed point by that special apparatus which is called an organ of perception.”
Bergson, Matter and Memory, 31.
52 Bergson, Creative Mind, 120.
53 Bergson, Matter and Memory, 34.
54 Interview with designer of algorithms for the extraction and analysis of anoma-
lous objects from video data, northwest England, June 2014.
55 As Lisa Parks and Caren Kaplan write, “Drones are not idle machines hovering
above; they are loaded with certain assumptions and ideologies . . . they are also
ideas, designs, visions, plans, and strategies that affect civilians on the ground.”
The assumptions and weights of deep neural network algorithms form a central
element of just such a loading of ideas and design. Parks and Kaplan, Life in the
Age of Drone Warfare, 9. For a compelling account of the embodied and durational
experience of drone technologies, see also Wilcox, “Embodying Algorithmic
War,” 11 – 28.
56 Hansen, New Philosophy for New Media, 104.
57 Deleuze, Bergsonism, 32.
58 Eubanks, Automating Inequality.
59 Nature Editors, “More Accountability for Big Data Algorithms.”
60 Brandom, “New System Can Measure the Hidden Bias.”
61 Foucault, “What Is an Author?,” 207.
62 Keenan, Fables of Responsibility, 3.
63 Keenan, Fables of Responsibility, 51.
64 In N. Katherine Hayles’s study of the forms of nonconscious cognition distrib-
 uted across human and technical systems, she argues that “nonconscious cogni-
tions in biological organisms and technical systems share certain structural and
functional similarities, specifically in building up layers of interactions.” Hayles,
Unthought, 13.
65 Haraway, Staying with the Trouble, 3 – 4 .
66 Butler, Giving an Account of Oneself, 18, 29.
67 Haraway, “Situated Knowledges,” 583.
68 Hayles, How We Became Posthuman, 287.
69 “If we understand machine learning as a data practice that reconfigures local 177
centers of power and knowledge,” writes Adrian Mackenzie, “then differences as-
sociated with machine learners in the production of knowledge should be a focus

Notes to Chapter 1
of attention.” Mackenzie, Machine Learners, 10.

1. The Cloud Chambers

1 C. T. R. Wilson, “On the Cloud Method,” 194.
2 C. T. R. Wilson, “On the Cloud Method,” 195.
3 C. T. R. Wilson, “On the Cloud Method,” 196.
4 Galison, Image and Logic, 140.
5 C. T. R. Wilson, “On the Cloud Method”, 199.
6 Dodge, “Seeing Inside the Cloud”; see also Dodge and Kitchin, Mapping
Cyberspace.
7 Konkel, “The cia’s Deal with Amazon,” 2.
8 The data of the seventeen icite agencies include information from sensors, sat-
ellites, uav images, open-­source social media, internet image files, text, video,
and voiceover internet files. For further discussion of how these different forms
of structured and unstructured data are ingested and analyzed, see Amoore and
Piotukh, “Life beyond Big Data.”
9 Konkel, “The cia’s Deal with Amazon,” 2.
10 Central Intelligence Agency, “cia Creates a Cloud.”
11 Alpers, “The Studio, the Laboratory,” 415.
12 Hayes, “Cloud Computing,” 9.
13 boyd and Crawford, “Critical Questions for Big Data,” 662 – 79; Mayer-­
Schönberger and Cukier, Big Data; and Kitchin, Data Revolution.
1 4 McCarthy, “Architects of the Information Society,” 2.
15 Human geographer Sam Kinsley has urged greater attention to the materiality of
the virtual and, specifically, to the details of where and how virtual geographies
become actualized. Kinsley, “Matter of Virtual Geographies,” 364 – 84.
16 Bratton, The Stack, 29.
17 O’Brien, Definitive Guide to the Data Lake.
18 Jaeger, “Where Is the Cloud?,” 4.
19 Though Sun Microsystems brought to market the first mobile and modular data
centers inside shipping containers, Google later patented designs for floating data
centers of shipping container servers stacked within cargo ships. Appealing to
the capacity of the floating data center to respond to emergencies such as the Fu-
 kushima earthquake, and to store and analyze data in international waters, the
Google cloud enters the geopolitics of logistics and exceptional spaces.
20 European Commission, “What Does the Commission Mean?”
21 Intelligence and Security Committee of Parliament, “Transcript of Evidence,”
with my additions from video of proceedings.
22 Crawford and Schultz, “Big Data and Data Protection,” 99, 105.
23 Crawford and Schultz, “Big Data and Data Protection,” 110.
2 4 Paglen, “Trevor Paglen Offers a Glimpse.”
178 25 Paglen, Invisible.
26 Galison, Image and Logic, 72.
27 Paglen, “Trevor Paglen Offers a Glimpse,” 4; my emphasis.
Notes to Chapter 1

28 Bridle, New Dark Age, 8.

29 Crary, 24/7; Rose, “Rethinking the Geographies”; and Mattern, Code and Clay,
Data and Dirt.
30 The design of algorithms for a national border-­control system was outsourced
through a layered series of small contractors and start-­ups. The major software
company that held the government contract had no working understanding of
how the algorithms were interacting across databases. In this case, a “place to
look” was a nondescript rented office in North London, hosting a small group
working on the design, validation, and testing of a series of interacting algo-
rithms. See Amoore, “Data Derivatives,” 24 – 43.
31 In Luciana Parisi’s description of how the spatial outline of an object is over-
flowed by digital computational models, she notes, “What is at stake with these
generative algorithms is that the notion of discreteness has changed, and now
includes a model of interactive agents that evolve in and through time.” Under-
stood in these terms, there could never be a definitive building or object in which
to locate cloud computation, for the forms of these places would be (computa-
tionally) malleable and evolving. Parisi, Contagious Architecture, 45.
32 Graham, Vertical, 66.
33 Scorer and Wexler, Cloud Studies in Colour.
3 4 Haraway, “Situated Knowledges,” 581.
35 Donna Haraway traces the etymology of “trouble” to its French roots of “to stir
up,” “to make cloudy,” or “to disturb.” Writing against the notion of trouble as
something to be averted through security and safety, Haraway develops a form
of “staying with the trouble” that learns how to “be truly present” in the “myriad
unfinished configurations.” Haraway, Staying with the Trouble, 1; and Haraway,
“Situated Knowledges,” 583. Understood in this way, to pursue a cloud ethics is
also to disturb the sediment of algorithmic logics, to make cloudy the otherwise
crystalline definiteness of the output.
36 Galison, Image and Logic, 97.
37 Galison, Image and Logic, 67.
38 John Durham Peters proposes that digital media extend historical distributed
infrastructures of the environment as media, so that “media are perhaps more in-
teresting when they reveal what defies materialization.” Peters, Marvelous Clouds,
 11. Similarly, for Derek McCormack, techniques of remote sensing are better un-
derstood as “sensing spectrality” rather than “a project of techno-­scientific mas-
tery.” McCormack, “Remotely Sensing Affective Afterlives,” 650.
39 Barad, Meeting the Universe Halfway, 176.
40 Erickson et al., How Reason Almost Lost Its Mind, 9, 29.
4 1 Bergson, Matter and Memory, 31.
4 2 Digital Reasoning, Synthesys Mission Analytics, 4. See also “Machine Learning for
Cybersecurity.”
43 Leopold, “cia Embraces Cloudera Data Hub.” 179
4 4 For Hayles, cognizers perform as “actors embedded in cognitive assemblages
with moral and ethical implications.” Hayles, Unthought, 31.

Notes to Chapter 1
45 Leopold, “cia Embraces Cloudera Data Hub,” 2.
46 Galison, Image and Logic, 67. Office of the Director of National Intelligence, Intel-
ligence Community.
47 C. Anderson, “The End of Theory.”
48 J. G. Wilson, Principles of Cloud Chamber Technique, 3.
49 Gentner, Maier-­Leibnitz, and Bothe, Atlas of Typical Expansion Chamber Photo-
graphs, 11.
50 Foucault, Security, Territory, Population, 6.
51 Barad, Meeting the Universe Halfway, 148.
52 Elden, “Secure the Volume,” 35 – 51; Crampton, “Cartographic Calculations of
Territory,” 92 – 103; Weizman, “Politics of Verticality.”
53 O’Brien, Definitive Guide to the Data Lake, 4.
54 The data lake architecture underpins the ability of commercial and government
agencies to gather and analyze “vast amounts of social networking platforms
emerging data” in a “single, centralized repository.” O’Brien, Definitive Guide to
the Data Lake, 3.
55 Parisi, Contagious Architecture, 2.
56 Josephson and Josephson, Abductive Inference, 12.
57 Konkel, “The cia’s Deal with Amazon.”
58 Dunning and Friedman, Practical Machine Learning, 23.
59 Interview with designers of cnn algorithms for facial recognition, London, May
2017.
60 J. G. Wilson, Principles of Cloud Chamber Technique, 122.
61 Rochester and Wilson, Cloud Chamber Photographs, vii.
62 Galison, Image and Logic, 130.
63 In the past, the use of data-­mining techniques for security and intelligence was
limited, in part, by the lack of available training datasets. The storage of data in
the cloud provides a readily available supply of training data (and a flexible com-
putational capacity) through which humans and algorithms can advance their
capacities to perceive new emergent events.
64 Interview with UK government procurement officers, London, October 2017.
65 Office of the Director of National Intelligence, Intelligence Community, ii.
66 Hayles, How We Think, 230.
 7 Office of the Director of National Intelligence, “icite Fact Sheet.”
6
68 “Web Intelligence”; my emphasis.
69 O. Halpern, Beautiful Data, 16, 40.
70 Derrida, Archive Fever, 17.
71 James Bridle’s interactive installation is available at Hyper-­Stacks, accessed June
2016, http://hyper-­stacks.com.
72 Derrida, Archive Fever, 18.
73 Observations conducted in the Department of Physics, Durham University, Au-
180 gust 2015.
74 Interviews and observations with the designers of credit card fraud detection al-
gorithms, Frankfurt, April 2014.
Notes to Chapter 1

75 Stengers, Cosmopolitics I, 22.

76 Bal, “Commitment to Look,” 145 – 62; and W. J. T. Mitchell, “There Are No Visual
Media,” 257 – 66.
77 Halpern, Beautiful Data, 156.
78 Stengers, Cosmopolitics I, 31.
79 D. Mitchell, Cloud Atlas, 389.

2. The Learning Machines

1 Newman et al., “Can Automatic Calculating Machines Be Said to Think?,” 10.
2 Three years earlier, Jefferson had delivered his famous lecture “The Mind of Me-
chanical Man” at the Royal College of Surgeons. In the lecture he had reflected
that neuroscience was being “invaded by the physicists and mathematicians,”
so that “we are being pushed to accept the great likeness between the actions of
electronic machines and those of the nervous system.” Jefferson, “Mind of Me-
chanical Man,” 1105 – 10. For Jefferson, distinct and important differences be-
tween human and machine persisted, not least that the human brain could sus-
tain “vast cell losses without serious loss of memory,” and that the machine “can
answer only problems given to it, prearranged by its operator” (1109). During
this postwar era of twinned cybernetics and Cold War rationalities, the neuro-
surgeon’s world of the material structures of the brain ran against the grain of a
belief that “machines might outdo humans in executing algorithms.” Erickson et
al., How Reason Almost Lost Its Mind, 9.
3 “Can Automatic Calculating Machines Be Said to Think?,” 11; my emphasis.
4 Turing, “Systems of Logic Based on Ordinals,” 192.
5 “Can Automatic Calculating Machines Be Said to Think?,” 13.
6 John Cheney-­Lippold’s argument that “we are data” emphasizes how we hu-
mans are “temporary members of different emergent categories,” these catego-
ries having been assigned meaning “without our direct participation.” I am less
persuaded that humans do not have direct participation in emergent forms of
classification and categorization, and I wish to attend to the entangled learning
that situates humans within algorithms and algorithms within humans. Cheney-­
Lippold, We Are Data, 5.
7 Anne Marie Mol’s ethnographic study of one disease — atherosclerosis — suggests
 how the multiplicity of the disease is rendered coherent, or made to cohere,
through the plural practices of imaging, scientific discussion and disagreement,
and everyday negotiations. Mol, Body Multiple. This making of a cohering whole
from multiplicity also characterizes the actions of algorithms. Brian Rotman
notes that digital data environments are generating “a network ‘I’ ” who is “plural
and distributed,” “heterogeneous and multiple.” Rotman, Becoming Beside Our-
selves, 8.
8 A feature space is a mapping of input data conducted by processing subsets of data
so that “after the entire input is processed, significant features correspond to 181
denser regions in the feature space.” These denser regions are “clusters.” Coman-
iciu and Meer, “Mean Shift,” 605.

Notes to Chapter 2
9 Lin et al., “Automatic Detection,” 802 – 10.
10 In machine learning and pattern recognition applications, a feature vector is
a means of numerically representing an n-­dimensional object in space so that
it can be computed. Thus, for example, the image of a human face is rendered
computable in a biometrics application via the extraction of feature vectors cor-
responding to pixelated data points. In the movement of the hands of surgeons
and surgical robots, the feature vectors denote the position along a trajectory in
three-­dimensional space.
11 Lin et al., “Automatic Detection,” 806.
12 Reiley, Plaku, and Hager, “Motion Generation of Robotic Surgical Tasks.”
13 N. Katherine Hayles describes a “cognitive assemblage” of human and techni-
cal “cognizers” who are “flexibly attending to new situations,” incorporating
modes of knowledge, and “evolving through experience to create new responses.”
Hayles, Unthought, 32. Adrian Mackenzie proposes that the “various subject po-
sitions” of machine learning are “neither unified nor fixed” but malleable and
transformative. Mackenzie, Machine Learners, 207.
1 4 Van den Berg et al., “Superhuman Performance of Surgical Tasks.”
15 Robotic surgery actively premediates the actualized surgery in the sense intended
by Richard Grusin in his work on premediation. The robot’s virtual architecture
allows the surgeon to explore the trajectories of the surgery in advance of the ac-
tual surgery, anticipating the optimal trajectories of movement, experimenting,
and preempting potential problems. Grusin, Premediation.
16 Prentice, Bodies in Formation, 227.
17 Prentice, Bodies in Formation, 227 – 29.
18 Prentice, Bodies in Formation, 231.
19 Popescu and Mastorakis, “Simulation of da Vinci Surgical Robot,” 141.
20 Interview with obstetric surgical team, Newcastle-­upon-­Tyne, April 2016.
21 Barad, Meeting the Universe Halfway, 208.
22 Jochum and Goldberg, “Cultivating the Uncanny.”
23 Following Henri Bergson, the cut and the multiplicity are actually two distinct
forms of multiplicity. As Gilles Deleuze writes, “Remember that Bergson opposed
two types of multiplicity — actual multiplicities that are numerical and discon-
tinuous and virtual multiplicities that are continuous and qualitative.” Deleuze,
 Bergsonism, 79. For the singular cut to take place, the virtual multiplicities of past
encounters between people and algorithms must be rendered as actual divisible
numerical multiplicities, or feature spaces.
2 4 In a computer science paper on machine learning for robot surgery, the authors
present their cognate findings in military robots, landmine detection, and ro-
bot inspection of nuclear contamination. The development of machine learn-
ing algorithms takes place indifferent to the environment as such because it
understands the environment explicitly in terms of feature space and not place.
182 Popescu and Mastorakis, “Simulation of da Vinci Surgical Robot,” 137.
25 Hayles, How We Became Posthuman, 4.
26 Hayles, How We Became Posthuman, 287.
Notes to Chapter 2

27 As Kant writes, “The I think must be able to accompany all my representa-

tions,” so that “the thought that these representations given in intuition all
together belong to me means, accordingly, the same as that I unite them in a
self-­consciousness.” The reasoning subject is the anchor that secures all repre-
sentations, all scientific and political claims, so that one conceives of oneself as a
unity of thought and action. Kant, Critique of Pure Reason, 247.
28 In 2016, Intuitive Surgical reserved $100 million to settle an undisclosed num-
ber of claims for damages. The company’s insurer, Illinois Union, sued Intuitive
Surgical for its alleged failure to disclose the extent of medical injuries associated
with its robot systems. Mehotra, “Maker of Surgical Robot.”
29 As Noel Sharkey has proposed in his accounts of the human control of robot weap-
ons, “There will be a human in the loop for lethality decisions.” For Sharkey, the
human in the loop is understood to be the locus of “deliberative human control,”
where space and time in the kill chain is afforded for “deliberate reasoning.” Shar-
key, “Staying in the Loop,” 25, 35. See also Strawser, Killing by Remote Control.
30 Lucy Suchman and Jutta Weber detail the limits of the notion of ethics in rela-
tion to autonomous weapons systems. Reflecting on the iterative relationship
between military robots, their machine learning algorithms, and the humans
conducting postprocessing, they argue that “robot behaviour shifts to another
level: it is no longer totally pre-­programmed but, instead, more flexible. The be-
haviour is mediated by random effects of the system’s architecture or learning
algorithms, which can result in emergent effects which are then exploited post-­
processing.” Suchman and Weber, “Human-­Machine Autonomies,” 88.
31 Haraway offers a “method of tracing” or “following a thread in the dark” that
stands in contrast to the illuminations sought by calls for transparency or an
opening of closed systems. A Haraway-­inspired tracing does not seek out the
concealed origins of science, but rather the tracing makes material the threads
that have been necessary to science’s story of origins and objectivity. To follow a
thread or a trace is thus not comparable to an opening of scientific black boxes,
for it does not seek a moment of revelation and exposure. She describes a “touch-
ing silliness about technofixes (or techno-­apocalypses)” and counters this with
a determination to “follow the threads where they lead” in “real and particular
places and times.” Haraway, Staying with the Trouble, 3.
 32 Interviews and observations with algorithm designers for a national border and
immigration control system, July 2016.
33 I am grateful to Phil Garnett for pointing out the specificity of the notion of
“good enough” in the practice of machine learning and generously discussing the
potential consequences.
3 4 Dunning and Friedman, Practical Machine Learning, 14.
35 Dunning and Friedman, Practical Machine Learning, 16; my emphasis.
36 There is extremely important work being done to address how racism, sexism,
and economic inequalities are becoming inscribed into the actions of algorithms. 183
See Noble, Algorithms of Oppression; and Eubanks, Automating Inequality.
37 Ramon Amaro’s work has focused on the capacities of the algorithm to recognize

Notes to Chapter 2
blackness and the new regimes of racialized recognition that accompany algo-
rithmic techniques. Amaro, “Pre-­emptive Citizenship.”
38 Bucher, If . . . Then, 4.
39 Domingos, Master Algorithm, 189.
40 Quinlan, “Induction of Decision Trees,” 82.
4 1 Quinlan, “Induction of Decision Trees,” 83.
4 2 Quinlan, “Induction of Decision Trees,” 104.
43 Though it may appear inconsequential that an algorithm may be less able to
identify shape or outline than it is able to detect texture or color, in fact it is be-
coming increasingly important to our contemporary regime of recognition. For
example, neural networks used in the assessment of medical images are more
attuned to the texture or color of a human organ than they are to its shape or
indeed to the contours of a lesion. Hendrycks et al., “Natural Adversarial Exam-
ples,” 2.
4 4 LeCun, Bengio, and Hinton, “Deep Learning,” 436.
45 Krizhevsky, Sutskever, and Hinton, “ImageNet Classification,” 1097 – 1105.
46 Across a wide range of recognition algorithms, including tumor detection in medi-
cal imaging and facial recognition, the AlexNet algorithm is used as the pretrained
model, with the domain-­specific images (e.g., datasets of surgical photographs)
added only to the training of the last hidden layer of the neural network.
47 Krizhevsky, Sutskever, and Hinton, “ImageNet Classification,” 1097.
48 Amazon Mechanical Turk has become a vehicle for the human labeling of images
that will be used to train a machine learning algorithm. Described by Amazon as
“a marketplace for work that requires human intelligence,” the casualization and
precarity of this form of work means that a population of workers is engaged in
the preparation of data for algorithmic systems that may at some future moment
be used against them or others in their community.
49 Observations of the process of training algorithms for image recognition, Go-
thenburg, November 2016.
50 MacKay, Information Theory, 345.
51 Interviews and observations with designers of algorithms for airport border secu-
rity, September 2016.
52 LeCun, Bengio, and Hinton, “Deep Learning,” 436.
 53 McGeoch, Guide to Experimental Algorithmics, 87.
54 Interview with test computer scientist working with ibm on facial recognition,
London, June 2018.
55 All forms of recognition, as political claims, carry within them the possibility of
misrecognition or being unrecognized. As Judith Butler has put the problem, “To
ask for recognition, or to offer it, is precisely not to ask for recognition for what
one already is. It is to solicit a becoming, to petition the future always in relation
to the Other.” Butler, Precarious Life, 44.
184 56 The Point Cloud Library is an open source repository of point cloud 2d and 3d
images and object recognition algorithms. The Point Cloud Library offers tech-
niques to “stitch 3d point clouds together” and to “segment relevant parts of a
Notes to Chapter 2

scene.” As laser-­scanning techniques generate datasets of 3d objects, from ar-

chitects’ scans of buildings to mri scans of tumors, the 3d point clouds become
training data for future neural network algorithms. See “About.”
57 The algorithmic capacity to recognize 3d objects from multiple angles and view-
points has also become the condition of possibility for facial recognition systems
that can be deployed on the variegated and oblique angles of the cctv camera
surveillance architecture.
58 Mahler et al., “Dex-­Net 1.0.”
59 Mahler et al., “Dex-­Net 1.0,” 1.
60 Mahler et al., “Dex-­Net 1.0,” 2.
61 Interview with algorithm design team working on surgical robotics, detection
of problem gambling, and biometrics recognition systems, London, September
2014.
62 Algorithm design team interview, London, September 2014.
63 Reflecting on his experiences of the events of May 1968, Foucault expresses how
a set of political claims can be brought that have never previously been registered
as claimable rights or encoded norms. Such claims that are “not part of the statu-
tory domain” included “questions about women, about relations between the
sexes, about medicine, about mental illness, about the environment, about mi-
norities, about delinquency.” In the bringing of these political questions,
Foucault does not identify an appeal to political rights but rather a “liberation
of the act of questioning” from existing political doctrines. One can imagine a
bringing of questions that are not traditionally part of the statutory domain of
politics as one element of a cloud ethics. Foucault, “Polemics, Politics, Problema-
tizations,” 115.

3. The Uncertain Author

1 Kirchner, “New York City Moves.”
2 ProPublica’s investigation obtained the risk scores assigned to seven thousand
people arrested in Broward County, Florida, in 2013 and 2014 and analyzed the
reoffending rate. Their study found that the algorithm used to predict the risk of
reoffending was of not greater predictive value than a coin toss, but that the ef-
fects were acutely divided along racialized lines. “The formula was particularly
 likely to falsely flag black defendants as future criminals, wrongly labeling them
this way at almost twice the rate of white defendants,” the ProPublica report
concluded. Angwin et al., “Machine Bias.”
3 The Council for Big Data, Ethics, and Society calls for a public discussion of how
the ethical conduct of algorithm designers could be enshrined in a program of
training and regulation. Metcalf, Keller, and boyd, “Perspectives on Big Data.”
4 An influential editorial in Nature proposed that algorithmic bias could be scruti-
nized through the disclosure of the data sources used by the designers. The ap-
peal to a unified body of data and a known designer similarly invokes the line of 185
accountability to an authorial source. Nature Editors, “More Accountability.”
5 For Foucault the author function does not consist merely of the attribution of a

Notes to Chapter 3
work to an individual, but rather “constructs a certain being of reason that we
call ‘author.’ ” The author function unifies the oeuvre, or the body of work, so
that unresolved tensions or contradictions appear resolvable and “incompatible
elements are at last tied together.” Foucault, “What Is an Author?,” 384 – 85.
6 Foucault, “What Is an Author?,” 377.
7 Lev Manovich has depicted the distinction between narrative and database as
one of competing forms, even as “enemies.” Manovich, Language of New Media, 228.
Contra Manovich, N. Katherine Hayles develops an understanding of the interde-
pendent mediations of narrative and database, nonetheless proposing that “they
remain different species, like bird and water buffalo.” Hayles, How We Think, 178.
With contemporary machine learning algorithms, however, the indeterminacy
that Hayles suggests “databases find difficult to tolerate” moves closer to the “inex-
plicable” and “ineffable” she associates with the narrative form (178 – 79).
8 Kirby, Quantum Anthropologies, xi.
9 Reflecting on the Greco-­Roman practice of writing the hupomnemata, or indi-
vidual notebooks, Foucault describes writing that “resists scattering” by the “se-
lecting of heterogeneous elements.” The writing becomes a “matter of unifying
heterogeneous fragments” into a body that is recognizable not as a “body of doc-
trine” but as a “certain way of manifesting oneself to oneself and to others.” Fou-
cault, “Self Writing,” 212 – 13, 216.
10 The computational practice of natural language processing is historically in-
tertwined with linguistic theory and, specifically, the debates on how language
is learned by humans. Noam Chomsky’s Platonic approach to linguistics influ-
enced strongly nlp methods from the 1950s to the 1990s. Chomsky’s approach
to the “generative grammar” of language learning favored rules-­based models
of syntactic structure and, thus, supplied to computer science the grounds for
devising algorithmic rules for the extraction of meaning from language. For an
account of what was at stake in the 1950s debates between behavioral and rules-­
based approaches to language learning, see Chomsky, “Bad News,” 26 – 58.
11 Amoore, Politics of Possibility, 39 – 45.
12 Statistical machine translation advanced significantly with the availability of
labeled training data derived from the translations of all European Union docu-
mentation into the eleven recognized official languages of the EU. In effect, the
 human-­translated documents provided a vast corpus of training data for transla-
tion algorithms. See Koehn, “Europarl.” Advances in deep neural networks have
developed these techniques beyond the statistical to devise predictive models
for the recognition of human language. See Bahdanau, Cho, and Bengio, “Neu-
ral Machine Translation.” For an overview of the contemporary state of the art
in nlp methods, see Hirschburg and Manning, “Advances in Natural Language
Processing,” 261 – 66.
13 Kurzweil, The Singularity Is Near.
186 1 4 For the Google research team’s account of the development of their algorithms
see Pickett, Tar, and Strope, “On the Personalities of Dead Authors.”
15 Juola, “How a Computer Program Helped Show.”
Notes to Chapter 3

16 Burgess, “Google Wants to Predict.”

17 Amoore, “Data Derivatives,” 24 – 43.
18 Cuthbertson, “Google’s ai Predicts.”
19 Foucault, “What Is an Author?,” 378.
20 Foucault, “What Is an Author?,” 378.
21 Foucault, “What Is an Author?,” 378.
22 Foucault, “What Is an Author?,” 379.
23 Foucault, “What Is an Author?,” 377.
2 4 Foucault, “What Is an Author?,” 380.
25 Foucault, “What Is an Author?,” 381.
26 Foucault, “What Is an Author?,” 383 – 85.
27 Foucault, “What Is an Author?,” 383.
28 Foucault, “What Is an Author?,” 383; my emphasis.
29 In Vicki Kirby’s wonderful account of the potential for extension of deconstruc-
tion into the sciences of mathematics, physics, and biology, she argues that “de-
construction’s home is as uncomfortable yet essential to the sciences as it is to
the humanities.” Kirby, Quantum Anthropologies, 6.
30 Foucault, “What Is an Author?,” 380.
31 Foucault, “What Is an Author?,” 385 – 86.
32 Turkle, Second Self.
33 Guizzo, “Google Wants Robots to Acquire.”
3 4 Among the most troubling of the political consequences of the search for source
code, private technology corporations have become very successful at protecting
their algorithms as proprietary sources. What this means is that, even when our
everyday interactions with algorithms generate their emergent capabilities, the
imagination of a proprietary “source code” locks out the public from understand-
ing this process.
35 Kirchner, “Federal Judge Unseals”; and Kirchner, “Traces of Crime.” Judge Caproni
had initially ruled that the vendor of the probabilistic genotyping technology had
legitimate proprietary and copyright interests preventing the release of the source
code. The ruling was subsequently overturned following a challenge from Yale Law
School. The full documents are at “Memorandum in Support of Application.”
36 Pishko, “Impenetrable Program.”
 37 For forensic protocols for the use of the fst algorithm, see “Forensic Biology
Protocols.”
38 Taylor, Bright, and Buckleton, “Source of Error,” 33.
39 Adams, “Exhibit A.”
40 Adams, “Exhibit A,” 9.
4 1 Fowles, “Nature of Nature,” 357.
4 2 Fowles, “Golding and ‘Golding,’ ” 203. Edward Said similarly notes the character
of great works of music and literature that cut against the grain of society’s pre-
vailing styles and sensibilities. Said, On Late Style. 187
43 Fowles, “Nature of Nature,” 358.
4 4 Fowles, “Nature of Nature,” 359.

Notes to Chapter 3
45 Deleuze, Negotiations, 133.
46 Fowles, “Notes on an Unfinished Novel,” 18.
47 “Writing unfolds like a game [jeu] that invariably goes beyond its own rules and
transgresses its limits,” writes Foucault. “In writing, the point is not to manifest
or exalt the act of writing, nor is it to pin a subject within language; it is, rather,
a question of creating a space into which the writing subject constantly disap-
pears.” Foucault, “What Is an Author?,” 378.
48 Gilles Deleuze considers there to be “no literature without fabulation,” so that
“writing is a question of becoming, always incomplete, always in the midst of be-
ing formed.” Deleuze, Essays Critical and Clinical, 1 – 3 .
49 Critchley, Ethics of Deconstruction, 34.
50 Fowles and Vipond, “Unholy Inquisition,” 368, 372.
51 Fowles and Vipond, “Unholy Inquisition,” 372.
52 Critchley, Ethics of Deconstruction, 48.
53 Critchley, Ethics of Deconstruction, 48.
54 Fowles and Vipond, “Unholy Inquisition,” 383.
55 Giorgio Agamben understands the author “as gesture” in the sense that “the au-
thor is present in the text only as a gesture that makes expression possible pre-
cisely by establishing a central emptiness within this expression.” Agamben, “Au-
thor as Gesture,” 66.
56 Mantel, “Day Is for the Living.”
57 The concept of fabulation serves to further displace the figure of the author in
favor of a focus on the power of writing. As Deleuze puts the problem: “There is
no literature without fabulation, but as Bergson was able to see, fabulation — the
fabulatory function — does not consist in imagining or projecting an ego. Rather,
it attains these visions, it raises itself to these becomings and powers.” Deleuze,
Essays Critical and Clinical, 3.
58 Deleuze, Essays Critical and Clinical, 118.
59 Foucault, “Self Writing,” 212.
60 Deleuze, Negotiations, 126.
61 The methods of deconstruction have long supplied to philosophy a means of en-
gaging with an extended sense of writing, or what Jacques Derrida has called
“archi-­writing” (archi-­écriture). Derrida, Of Grammatology, 15 – 20.
 2 Derrida, Limited Inc., 148.
6
63 Derrida, Limited Inc., 19.
64 Derrida, Limited Inc., 20.
65 Derrida, Limited Inc., 152.
66 Hinton et al., “Deep Neural Networks,” 83.
67 Hinton et al., “Deep Neural Networks,” 84.
68 Critchley, Ethics of Deconstruction, 21.
69 Nissenbaum, Privacy in Context.
188
4. The Madness of Algorithms
1 On the notion of the flash crash and the advent of high-­frequency trading algo-
Notes to Chapter 3

rithms, see MacKenzie, “How to Make Money in Microseconds,” 16 – 18; Borch

and Lange, “High Frequency Trader Subjectivity,” 283 – 306; and Lange, Lenglet,
and Seyfert, “Cultures of High-­Frequency Trading,” 149 – 65.
2 Gibbs, “Microsoft’s Racist Chatbot.”
3 Bridle, “Something Is Wrong on the Internet”; Lewis, “Fiction Is Outperforming
Reality.”
4 O’Neil, Weapons of Math Destruction; Parkin, “Teaching Robots Right from
Wrong”; and Stubb, “Why Democracies Should Care.”
5 Among the signatories of the Future of Life Institute’s statement on artificial
intelligence, Elon Musk and Ray Kurzweil adjudicate systems that “must do what
we want them to do.” Ormandy, “Google and Elon Musk to Decide.”
6 When machine learning algorithms begin to generate their own form of lan-
guage, the tipping point of terminating the system is reached when the human
experimenters can no longer understand what the machines are saying to one
another. This marking of a limit point at the threshold of human comprehen-
sion invokes a departure from what is considered reasonable, as seen, for exam-
ple, when algorithms communicate with other algorithms. Walker, “Researchers
Shut Down ai.”
7 Derrida, “Cogito and the History of Madness,” 46.
8 Derrida, “Cogito and the History of Madness,” 39; and Foucault, History of Madness.
9 The recommendation system’s architecture “funnels” candidate videos re-
trieved and ranked from a corpus of millions of videos, presenting a targeted set
of recommendations to the user. Covington, Adams, and Sargin, “Deep Neural
Networks.”
10 Jacques Derrida famously opens his lecture on Foucault’s History of Madness with
an epigraph he attributes to Kierkegaard: “The instant of the decision is mad-
ness.” He uses different formulations of Kierkegaard’s decision across a number
of his works. For Derrida, all decision worthy of the name is a decision made
without reference to programmed knowledge, in a state of nonknowledge that is
darkness. Derrida’s use of Kierkegaard does not quite accurately capture the par-
adox of decision as Kierkegaard formulates it. Derrida, “Cogito and the History
of Madness,” 36; and Bennington, “Moment of Madness,” 103 – 27.
11 Thomas Keenan addresses directly the politics of decision as an “ethics and poli-
 tics” without grounds, so that “we have politics because we have no grounds, no
reliable standpoints.” Keenan, Fables of Responsibility, 4.
12 Alpaydin, Machine Learning, 79.
13 Norbert Wiener, “Atomic Knowledge of Good and Evil” [1950], folder 631, box
28D, mc 22, Norbert Wiener Papers, Institute Archives and Special Collections,
mit Libraries, Cambridge, Massachusetts (hereafter Wiener Papers).
1 4 Kate Crawford, “Dark Days”; Wiener, “Scientist Rebels”; Norbert Wiener, “Re-
bellious Scientist after Two Years,” Bulletin of the Atomic Sciences, folder 601, box
28C, Wiener Papers. 189
15 Norbert Wiener, “The Human Use of Human Beings” [1950], folder 639, box
29A, mc 22, Wiener Papers.

Notes to Chapter 4
16 Erickson et al., How Reason Almost Lost Its Mind, 2.
17 Erickson et al., How Reason Almost Lost Its Mind, 30.
18 Norbert Wiener, “Second Industrial Revolution and the New Concept of the Ma-
chine,” folder 619, box 28D, mc 22, Wiener Papers.
19 O. Halpern, Beautiful Data, 150.
20 O. Halpern, Beautiful Data, 151.
21 Chun, Control and Freedom, 300; Liu, Freudian Robot, 202.
22 M. Beatrice Fazi argues that computation is a process of determining indetermi-
nacy, where the generative potential of algorithms works with Turing’s notion of
the incomputable. Fazi, Contingent Computation. The kind of generative potential
identified by Fazi also drives the capacity of algorithmic systems to derive value
from speculative and experimental models.
23 Foucault, History of Madness, 164.
2 4 Foucault, History of Madness, xxxii.
25 Foucault, History of Madness, 169.
26 Jacques Derrida describes Foucault’s determination to avoid the trap of a history
of madness untamed by reason as “the maddest aspect of his project.” Derrida,
“Cogito and the History of Madness,” 45; and Foucault, “Reply to Derrida,” 575.
27 Derrida, “Cogito and the History of Madness,” 46.
28 Derrida, “Cogito and the History of Madness,” 46.
29 “To the extent that madness is a form of knowledge,” writes Foucault, “it was
also a set of norms, both norms against which madness could be picked out as
a phenomenon of deviance within society, and, at the same time, norms of be-
haviour for normal individuals.” The idea of madness, for Foucault, defined “a
certain mode of being of the normal subject, as opposed to and in relation to the
mad subject.” The dissension between madness and reason, then, also marks the
relation of selves to others in relation to thresholds of norm and anomaly. Fou-
cault, Government of Self and Others, 3.
30 Foucault, History of Madness, 426.
31 Foucault, History of Madness, 426.
32 Foucault, History of Madness, 461.
33 Foucault, History of Madness, 159.
3 4 Foucault, History of Madness, 527.
 35 Foucault, History of Madness, 528.
36 Foucault, History of Madness, 164.
37 Derrida, “Cogito and the History of Madness,” 62 – 63.
38 In place of the notion that the madness of algorithms denotes a departure from
some normal course, I envisage something closer to Brian Rotman’s concept
of “becoming beside oneself.” Rotman’s concept depicts a plurality of being in
which the self is “distributed and in excess of unity.” Rotman, Becoming Beside
Ourselves, 103.
190 39 The nsf funded Nicholas Evans’s study “Ethical Algorithms in Autonomous
Vehicles” in 2017. See nsf Award Abstract 1734521. See also Nyholm and Snids,
“Ethics of Accident-­A lgorithms,” 1275 – 89.
Notes to Chapter 4

40 Phys.org, “UMass Lowell Professor Steers Ethical Debate.” The encoded ethics
of algorithms for autonomous vehicles poses questions such as “Should your self-­
driving car protect you at all costs? Or should it steer you into a ditch to avoid
hitting a school bus full of children?” The notion of a moral choice to be made
and to be encoded within the algorithm looms large in these accounts of the eth-
ics of decision. Webster, “Philosophy Professor Wins nsf Grant.”
4 1 Phys.org, “UMass Lowell Professor Steers Ethical Debate.”
4 2 The school bus features discursively within many experimental scenarios for
the development of ethical codes governing algorithmic action. The image of a
school bus embodies the fears of algorithms recognizing objects for drone sur-
veillance, and the school bus similarly serves as the epitome of a mad decision
within autonomous vehicles research.
43 Derrida, Gift of Death, 24.
4 4 Eklund, Nichols, and Knutson, “Cluster Failure,” 7900 – 7905.
45 Timmer, “Software Faults Raise Questions.”
46 Brackeen, “Facial Recognition Software.”
47 Foucault, History of Madness, 241.
48 Domingos, Master Algorithm, 83.
49 Kourou et al., “Machine Learning Applications,” 8 – 17.
50 As both Derrida and Foucault elaborate, madness is what cannot be said, the un-
spoken that is absent from a body of work. Derrida, “Cogito and the History of
Madness,” 51.
51 In all the domains where I have followed the development of algorithms for state
security — border controls, counter-­terror, policing, and immigration control — 
my discussions with algorithm designers have included some reference to the use
of random forest algorithms as preferred by operational security teams. Though
I cannot definitively explain the allure these algorithms hold for the state, in
many cases, the random forest algorithm was using proxy data to identify pat-
terns associated with potential threats.
52 Grothoff and Porup, “nsa’s skynet Program.”
53 Grothoff and Porup, “nsa’s skynet Program,” 2.
54 Amoore, “Data Derivatives,” 24 – 43.
55 Breiman, “Random Forests,” 5 – 32. Breiman’s paper has had extraordinary in-
 fluence on the development of machine learning, having been cited more than
thirty-­five thousand times.
56 Breiman, “Random Forests,” 6.
57 Breiman, “Random Forests,” 7. See also Abdulsalam, Skillicorn, and Martin,
“Classification Using Streaming Random Forests,” 24.
58 Daston, “Probabilistic Expectation,” 234 – 60.
59 Grothoff and Porup, “nsa’s skynet Program,” 3.
60 Abdulsalam, Skillicorn, and Martin, “Classification Using Streaming Random
Forests,” 25. 191
61 Tucker, “New ai Learns through Observation Alone.”
62 Foucault, History of Madness, 521.

Notes to Chapter 5
63 It has long been acknowledged in relation to financial markets that notions of
irrationality or chance have a place within the making of ideas of rational eco-
nomic humans. De Goede, Virtue, Fortune, and Faith. Derrida, Death Penalty, 21.
64 Foucault, History of Madness, 521.
65 For Derrida, “calculation is always calculating with what is incalculable as well
as what is non-­calculable” so that there is always also indecision and undecidabil-
ity. Derrida, Death Penalty, 197.
66 Derrida, Death Penalty, 4, 42.
67 Derrida, Death Penalty, 93.
68 Derrida’s discussion of the scaffold invokes the material scaffold of the sovereign
scene of the death penalty as well as the “scaffolding” that renders an act as an
unconditional principle. Derrida, Death Penalty, 21.
69 Derrida, Death Penalty, 26.
70 Derrida, Death Penalty, 6, 139.

5. The Doubtful Algorithm

1 While working on the Manhattan Project, Feynman and his colleagues calcu-
lated the energy release from the atomic bomb’s implosion using ibm punch card
machines. Feynman describes how they learned to compute “around the error”
when the “machine made a mistake.” The use of such calculating machines al-
lowed the scientists to conduct multiple calculations in parallel so that “a mis-
take made at some point in the cycle only affects the nearby numbers.” Whereas
in their early attempts at the mathematics of the atomic bomb, the linearity of
the calculation meant that errors caused them to “go back and do it over again,”
the parallel processing of the punch cards allowed them to incorporate the er-
ror and continue to progress with the program. In short, intrinsic to the science
of the atomic bomb was a form of computation that could incorporate error and
continue to a useable output. Some years later, after the events of Hiroshima and
Nagasaki, Feynman reflected on the relationship between his partial and experi-
mental computation and the science of the atomic bomb: “I looked out at the
buildings and I began to think, you know, about how much the radius of the Hi-
roshima bomb damage was. . . . How far from here was 34th Street? It’s a terrible
thing that we made.” Feynman, Surely You’re Joking, 131 – 36.
 2 Feynman, What Do You Care What Other People Think?, 245, 247.
3 Braidotti, “Posthuman, All Too Human,” 197.
4 Haraway, Modest_Witness@Second_Millennium, 7.
5 I owe the development of this point on the enactment of a decision indifferent
to persistent doubt to discussions with Susan Schuppli, Eyal Weizman, and their
wonderful postgraduate forensic architecture students, February 19, 2016.
6 As Rosi Braidotti argues, the posthuman subject need not be anthropocentric,
but “it must be the site for political and ethical accountability.” Braidotti, Posthu-
192 man, 103.
7 Haraway, Modest_Witness@Second_Millennium, 23 – 2 4.
8 Butler, Giving an Account of Oneself, 29.
Notes to Chapter 5

9 House of Commons Science and Technology Select Committee, Algorithms in De-

cision Making.
10 Fast learning algorithms used in systems such as facial recognition are increas-
ingly generating their own “ground truths” through their exposure to data.
Among the many computer science examples of this process, see Yang et al.,
“Fast 𝓁1-­Minimization Algorithms,” 3234 – 46; and Hinton, Osindero, and Teh,
“Fast Learning Algorithm,” 1527 – 54.
11 Alpaydin, Machine Learning, 116.
12 Bottou, “From Machine Learning to Machine Reasoning,” 133 – 49.
13 Phil Garnett, of York University UK, put this point to me in response to my ac-
count of how an image recognition neural network learns to recognize form. He
captures here the significance of the algorithm’s indifference to the attributes of
the cat as qualities or essence. The relation to the ground truth data means that
the algorithm does not need to have a truth of the essence of “catness,” just as it
is indifferent to the truthfulness of a person who is subject to an algorithmic de-
cision about their propensity for crime or the repaying of debts.
1 4 B. Anderson, “Preemption, Precaution, Preparedness”; Aradau and Van Mun-
ster, Politics of Catastrophe; de Goede, Speculative Security; and Amoore, Politics of
Possibility.
15 Haraway, “Situated Knowledges,” 581; and Hayles, How We Became Posthuman, 287.
16 Interview with designers of deep neural networks, London, September 2014.
17 The target output of the algorithm is, for this reason, akin to Samuel Weber’s
concept of a “target of opportunity,” in which the singularity of the targeted
strike is defined in terms of the realization of opportunities. With regard to al-
gorithms, the target of opportunity involves an extractive economy of profit and
sovereign power from the target output. S. Weber, Targets of Opportunity, 5.
18 Woodman, “Palantir Provides the Engine.”
19 House of Representatives Committee on Science and Technology, Investigation of
the Challenger Accident.
20 Feynman, What Do You Care What Other People Think?, 113.
21 Erickson et al., How Reason Almost Lost Its Mind, 29 – 30.
22 Feynman, What Do You Care What Other People Think?, 17.
23 Vaughan, Challenger Launch Decision, xiv – xv.
2 4 P. Halpern, Quantum Labyrinth, 114.
25 Barad, “On Touching,” 207.
26 Feynman, What Do You Care What Other People Think?, 135.
27 Feynman, What Do You Care What Other People Think?, 137.
28 Feynman, What Do You Care What Other People Think?, 107.
29 Haraway, “Situated Knowledges,” 584.
30 Feynman, What Do You Care What Other People Think?, 180.
31 Bennett, Vibrant Matter.
32 Heidegger, “Bauen, Wohnen, Denken.” 193
33 Hayles, Unthought, 29.
3 4 Latour and Weibel, Making Things Public.

Notes to Chapter 6
35 Haraway, Modest_Witness@Second_Millennium, 24, 34.
36 Barad, “On Touching,” 208.
37 Haraway, Modest_Witness@Second_Millennium, 35.
38 Foucault, Courage of Truth, 79.
39 Foucault, Courage of Truth, 30.
40 Butler, Giving an Account of Oneself, 19.
4 1 Foucault, Government of Self and Others, 62.
4 2 Foucault, Wrong-­Doing, Truth-­Telling, 200.
43 Feynman, What Do You Care What Other People Think?, 139.
4 4 Berlant, Cruel Optimism, 23 – 2 4.
45 Berlant, Cruel Optimism, 262.
46 There is widespread acknowledgement among self-­styled data scientists that 90
percent of the work of building a model remains a task of data preparation.
47 Parisi, Contagious Architecture, 95.
48 Feynman, What Do You Care What Other People Think?, 214.
49 Harding, Snowden Files; and Greenwald, No Place to Hide.
50 D. Anderson, Attacks in London and Manchester.
51 Keenan, Fables of Responsibility, 3, 42.
52 Keenan, Fables of Responsibility, 12; and Derrida, “Force of Law,” 26.
53 Observation of presentations of “threat intelligence” software to UK govern-
ment, London, June 2017.
54 Truvé, “Building Threat Analyst Centaurs.”
55 Butler, Giving an Account of Oneself, 19, 45.
56 Derrida, Gift of Death, 24; and Keenan, Fables of Responsibility, 2.
57 Deleuze, Bergsonism, 29.
58 Barad, “On Touching,” 212.
59 Barad, Meeting the Universe Halfway, 265.
60 Feynman, Surely You’re Joking, 247; my emphasis.

6. The Unattributable
1 Xu et al., “Show, Attend and Tell.”
2 Vinyals et al., “Show and Tell,” 2048 – 57.
3 Xu et al., “Show, Attend and Tell,” 2048.
 4 Vinyals et al., “Show and Tell,” 653.
5 Crary, Suspensions of Perception.
6 Xu et al., “Show, Attend and Tell,” 2048.
7 Vinyals et al., “Show and Tell,” 652.
8 Xu et al., “Show, Attend and Tell,” 2048.
9 Derrida, Limited Inc., 152 – 53.
10 Shane and Wakabayashi, “Business of War.”
11 Derrida, Limited Inc., 151.
194 12 Deleuze, Essays Critical and Clinical, 4.
13 Deleuze, Essays Critical and Clinical, 113.
1 4 Bottou, “From Machine Learning to Machine Reasoning,” 133 – 49.
Notes to Chapter 6

15 To begin from the propensities of the algorithm, and not from strategies to over-
turn it, is to echo Isabelle Stengers’s sense that “the diagnosis of becomings” is
not a strategy for alternatives but rather a “speculative operation, thought ex-
periment” in which the possibilities are not determined. Stengers, Cosmopolitics I,
12 – 13.
16 Addressing the machine learner “kittydar,” Adrian Mackenzie describes how
“the software finds cats by cutting the images into smaller windows. For each
window, it measures a set of gradients” and then “compares these measurements
to the gradients of known cat images.” Mackenzie, Machine Learners, 4 – 5.
17 The UK Information Commissioner’s Office (ico) is investigating the use of mil-
lions of Facebook profiles to train the Cambridge Analytica algorithms. At the
time of writing, the maximum fine has been imposed on Facebook for the unlaw-
ful sharing of data. See Greenfield, “Cambridge Analytica Files.”
18 Deleuze, Bergsonism, 16.
19 The question, as Henri Bergson formulated it, “is not how perception arises, but
how is it limited,” so that the image of a whole scene is “reduced to the image of
that which interests you.” Bergson, Matter and Memory, 34.
20 For Crary, perception exceeds the “single-­sense modality of sight,” extending
to “hearing and touch” and “most importantly, irreducibly mixed modalities.”
Crary, Suspensions of Perception, 1, 4.
21 Deleuze, Bergsonism, 31.
22 Deleuze, Bergsonism, 32.
23 Simon Critchley argues that “deconstruction is a philosophy of hesitation”
in that it carries with it “the experience of undecidability” essential to ethics.
Critchley, Ethics of Deconstruction, 42.
2 4 For Derrida, there is a close relation between “feeling the weight” of something,
“peser,” and thought, “penser.” Derrida, Death Penalty, 93.
25 For Derrida, the “trace of the rejected alternative” is not a matter of “either this
or that” in the sense that some alternative path could be reopened, but a “neither
this . . . nor that,” where the decisive moment is unknowable. Derrida, Writing
and Difference, 112.
26 Keenan, Fables of Responsibility, 1, 51.
27 Keenan, Fables of Responsibility, 51.
 28 Writing on the genealogy of ethics, Michel Foucault describes the transforma-
tion in how Greek and Christian societies viewed themselves: these changes
“are not in the code, but in what I call the ‘ethics’ ” which is in “their relations to
themselves and to others.” Foucault, “On the Genealogy of Ethics,” 255, 266.
29 Deleuze, Spinoza, 125.
30 Deleuze, Spinoza, 125.
31 Stengers, Thinking with Whitehead, 515.
32 Critchley, Ethics of Deconstruction, 18.
33 Derrida, Limited Inc., 122. 195
3 4 Haraway, “Situated Knowledges,” 575 – 99.
35 Haraway, “Situated Knowledges,” 586.

Notes to Chapter 6
36 Haraway, “Situated Knowledges,” 583.
37 Haraway, “Situated Knowledges,” 586.
38 Xu et al., “Show, Attend and Tell,” 2052, 2055.
39 Burgess, “UK Police Are Using ai.”
40 Laughland, “Baltimore Unrest.”
4 1 Derrida, Limited Inc., 152 – 53.
4 2 Deleuze, Spinoza, 17, 51.
43 Foucault, “What Is Critique?,” 265, 278.
4 4 Critchley, Ethics of Deconstruction, 48.
45 Critchley, Ethics of Deconstruction, 48.
46 As Deleuze has written, the singularities “connect with one another in a manner
entirely different from how individuals connect” and can be neither clustered
nor “grouped [or] divided in the same way.” Deleuze, Pure Immanence, 30.
This page intentionally left blank
BIBLIO GR A PH Y

Abdulsalam, Hanady, David B. Skillicorn, and Patrick Martin. “Classification Using

Streaming Random Forests.” ieee Transactions on Knowledge and Data Engineering
23, no. 1 (2011): 24.
“About.” pcl. Accessed April 2018. http://www.pointclouds.org/about/.
aclu. “Baltimore County Police Department and Geofeedia Partner to Protect the
Public during Freddie Gray Riots.” aclu, October 11, 2016. https://www.aclunc
.org/docs/20161011_geofeedia_baltimore_case_study.pdf.
Adams, Nathaniel. “Exhibit A: Declaration of Nathaniel Adams.” October 17, 2017.
https://www.documentcloud.org/documents/4112649-­10-­17-­17-­Unredacted-­NA
-­Exhibit-­A .html.
Adey, Peter. Aerial Life: Spaces, Mobilities, Affects. Oxford: Wiley-­Blackwell, 2010.
Agamben, Giorgio. “The Author as Gesture.” In Profanations, translated by Jeff Fort,
61 – 72. New York: Zone Books, 2007.
Agamben, Giorgio. Potentialities: Collected Essays in Philosophy. Stanford, CA: Stanford
University Press, 2007.
Allen, John. “The Whereabouts of Power: Politics, Government and Space.” Geograf-
iska Annaler: Series B 86, no. 1 (2004): 19 – 32.
Alpaydin, Ethem. Machine Learning. Cambridge, MA: mit Press, 2016.
Alpers, Svetlana. “The Studio, the Laboratory and the Vexations of Art.” In Picturing
Science, Producing Art, edited by P. Galison and C. Jones, 401 – 17. New Haven, CT:
Yale University Press, 1998.
Amaro, Ramon. “Pre-­emptive Citizenship: Crime Prediction and Statistical Inter-
vention in Racialized Ecologies.” Paper presented at the Radical Ecologies Confer-
ence, Goldsmiths University of London, May 4, 2015.
Amaro, Ramon. “Race, Surveillance and Autopoetically Instituted Being.” Paper pre-
sented at the Race and Surveillance Conference, Birkbeck University of London,
May 4, 2018.
Amoore, Louise. “Data Derivatives: On the Emergence of a Security Risk Calculus
for Our Times.” Theory, Culture and Society 28, no. 6 (2011): 24 – 43.
Amoore, Louise. The Politics of Possibility: Risk and Security beyond Probability. Durham,
NC: Duke University Press, 2013.
Amoore, Louise, and Marieke de Goede. “Transactions after 9/11: The Banal Face of
the Preemptive Strike.” Transactions of the Institute of British Geographers 33, no. 2
(2008): 173 – 85.
Amoore, Louise, and Volha Piotukh. “Life beyond Big Data: Governing with Little
Analytics.” Economy and Society 44, no. 3 (2015): 341 – 66.
Anderson, Ben. “Preemption, Precaution, Preparedness: Anticipatory Action and Fu-
ture Geographies.” Progress in Human Geography 34, no. 6 (2010): 777 – 98.
Anderson, C. “The End of Theory: The Data Deluge Makes Scientific Method Obso-
lete.” Wired, June 23, 2008.
 Anderson, David. Attacks in London and Manchester. London: Brick Court Chambers,
2017. https://assets.publishing.service.gov.uk/government/uploads/system/uploads
/attachment_data/file/664682/Attacks_in_London_and_Manchester_Open
_Report.pdf.
Angwin, Julia, Jeff Larson, Surya Mattu, and Lauren Kirchner. “Machine Bias.”
ProPublica, May 23, 2016. https://www.propublica.org/article/machine-­bias-­risk
-­assessments-­in-­criminal-­sentencing.
Aradau, Claudia, and Rens Van Munster. Politics of Catastrophe: Genealogies of the Un-
198 known. London: Routledge, 2011.
Ash, James. The Interface Envelope: Gaming, Technology, Power. London: Bloomsbury,
2015.
Bibliography

Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. “Neural Machine Trans-
lation by Jointly Learning to Align and Translate.” Presented at the International
Conference on Learning Representations, San Diego, California, May 7 – 9, 2015.
Bal, Mieke. “The Commitment to Look.” Journal of Visual Culture 4, no. 2 (2005):
145 – 62.
Barad, Karen. Meeting the Universe Halfway: Quantum Physics and the Entanglement of
Matter and Meaning. Durham, NC: Duke University Press, 2007.
Barad, Karen. “On Touching: The Inhuman That Therefore I Am.” Differences: A Jour-
nal of Feminist Cultural Studies 23, no. 3 (2012): 206 – 23.
Beer, David. “Power through the Algorithm? Participatory Web Cultures and the
Technological Unconscious.” New Media and Society 11, no. 6 (2009): 985 – 1002.
Bennett, Jane. Vibrant Matter. Durham, NC: Duke University Press, 2010.
Bennington, Geoffrey. “A Moment of Madness: Derrida’s Kierkegaard.” Oxford Liter-
ary Review 3, no. 1 (2011): 103 – 27.
Bergson, Henri. The Creative Mind: An Introduction to Metaphysics. 1946. New York: Do-
ver, 2007.
Bergson, Henri. Matter and Memory. 1912. Mineola, NY: Dover Philosophical Classics,
2004.
Berlant, Lauren. Cruel Optimism. Durham, NC: Duke University Press, 2011.
Berlinski, David. The Advent of the Algorithm. New York: Harcourt, 2000.
Berry, David. The Philosophy of Software. Basingstoke: Macmillan, 2011.
Borch, Christian, and Ann-­Christina Lange. “High Frequency Trader Subjectivity:
Emotional Attachment and Discipline in an Era of Algorithms.” Socio-­Economic
Review 15 (2017): 283 – 306.
Bottou, L. “From Machine Learning to Machine Reasoning.” Machine Learning 94
(2014): 133 – 49.
boyd, danah, and Kate Crawford. “Critical Questions for Big Data.” Information, Com-
munication and Society 15, no. 5 (2012): 662 – 79.
Brackeen, Brian. “Facial Recognition Software Is Not Ready for Use by Law Enforce-
ment.” TechCrunch, June 22, 2018.
Braidotti, Rosi. The Posthuman. Cambridge: Polity, 2013.
Braidotti, Rosi. “Posthuman, All Too Human: Towards a New Process Ontology.”
Theory, Culture and Society 23, no. 7 – 8 (2006): 197 – 208.
Brandom, Russell. “A New System Can Measure the Hidden Bias in Otherwise Secret
Algorithms.” Verge, May 25, 2016. https://www.theverge.com/2016/5/25/11773108
/research-­method-­measure-­algorithm-­bias.
Bratton, Benjamin. The Stack: On Software and Sovereignty. Cambridge, MA: mit Press,
2015.
Breiman, Leo. “Random Forests.” Machine Learning 45, no. 1 (2001): 5 – 32.
Bridge, Mark, and Gabriella Swerling. “Bail or Jail? App Helps Police Make Decision
about Suspect.” Sunday Times (London), May 11, 2017. https://www.thetimes.co.uk
/article/bail-­or-­jail-­app-­helps-­police-­make-­decision-­about-­suspect-­kv766zjc9. 199
Bridle, James. New Dark Age: Technology and the End of the Future. London: Verso, 2018.
Bridle, James. “Something Is Wrong on the Internet.” Medium, November 6, 2017.

Bibliography
Browne, Simone. Dark Matters: On the Surveillance of Blackness. Durham, NC: Duke
University Press, 2015.
Bucher, Taina. If . . . Then: Algorithmic Power and Politics. Oxford: Oxford University
Press, 2018.
Burgess, Matt. “Google Wants to Predict the Next Sentences of Dead Authors.”
Wired (UK), February 26, 2016. https://www.wired.co.uk/article/google-­author
-­create-­artificial-­intelligence.
Burgess, Matt. “UK Police Are Using ai to Inform Custodial Decisions — But It Could
Be Discriminating against the Poor.” Wired, March 1, 2018.
Butler, Judith. Giving an Account of Oneself: A Critique of Ethical Violence. Amsterdam:
Van Gorcum Press, 2003.
Butler, Judith. Precarious Life: The Powers of Mourning and Violence. London: Verso, 2004.
Caruso, Catherine. “Can a Social Media Algorithm Predict a Terror Attack?” mit
Technology Review, June 16, 2016. https://www.technologyreview.com/s/601700
/can-­a-­social-­media-­algorithm-­predict-­a-­terror-­attack/.
Central Intelligence Agency. “cia Creates a Cloud: An Interview with cia’s Chief
Intelligence Officer, Doug Wolfe, on Cloud Computing at the Agency.” cia Fea-
tured Story Archive, 2014. https://www.cia.gov/news-­information/featured-­story
-­archive/2014-­featured-­story-­archive/cia-­creates-­a-­cloud.html.
Cheney-­Lippold, John. We Are Data: Algorithms and the Making of Our Digital Selves. New
York: New York University Press, 2019.
Chomsky, Noam. “Bad News: Noam Chomsky — A Review of B. F. Skinner’s Verbal
Behavior.” Language 35, no. 1 (1959): 26 – 58.
Chun, Wendy Hui Kyong. Control and Freedom: Power and Paranoia in the Age of Fiber
Optics. Cambridge, MA: mit Press, 2006.
Comaniciu, Dorin, and Peter Meer. “Mean Shift: A Robust Approach toward Feature
Space Analysis.” ieee Transaction on Pattern Analysis and Machine Intelligence 24, no.
5 (2002): 605.
Connolly, William. “Beyond Good and Evil: The Ethical Sensibility of Michel Fou-
cault.” Political Theory 21, no. 3 (1993): 365 – 89.
Covington, Paul, Jay Adams, and Emre Sargin. “Deep Neural Networks for YouTube
Recommendations.” Paper presented at the 10th acm Conference on Recom-
mender Systems, Boston, Massachusetts, September 15 – 19, 2016.
 Cowen, Deborah. The Deadly Life of Logistics. Minneapolis: University of Minnesota
Press, 2014.
Crampton, Jeremy. “Cartographic Calculations of Territory.” Progress in Human Geog-
raphy 35, no. 1 (2010): 92 – 103.
Crary, Jonathan. 24/7: Late Capitalism and the End of Sleep. London: Verso, 2013.
Crary, Jonathan. Suspensions of Perception: Attention, Spectacle, and Modern Culture.
Cambridge, MA: mit Press, 2001.
Crawford, Kate. “Artificial Intelligence’s White Guy Problem.” New York Times, June
200 25, 2016.
Crawford, Kate. “Dark Days: Artificial Intelligence and the Rise of Fascism.” Pre-
sented at the sxsw Conference, Austin, Texas, March 12, 2017.
Bibliography

Crawford, Kate, and Jason Schultz. “Big Data and Data Protection: Toward a Frame-
work to Redress Predictive Privacy Harms.” Boston College Law Review 55, no. 1
(2014): 93 – 128.
Critchley, Simon. The Ethics of Deconstruction: Derrida and Levinas. Oxford: Blackwell,
1992.
Cuthbertson, Anthony. “Google’s ai Predicts the Next Sentence of Dead Authors.”
Newsweek, February 29, 2016.
Daston, Lorraine. “The Moral Economy of Science.” Osiris 10 (1995): 2 – 2 4.
Daston, Lorraine. “Probabilistic Expectation and Rationality in Classical Probability
Theory.” Historia Mathematica 7 (1980): 234 – 60.
Daston, Lorraine, and Peter Galison. Objectivity. Cambridge, MA: mit Press, 2007.
de Goede, Marieke. Speculative Security: The Politics of Pursuing Terrorist Monies. Minne-
apolis: University of Minnesota Press, 2012.
de Goede, Marieke. Virtue, Fortune, and Faith: A Genealogy of Finance. Minneapolis:
University of Minnesota Press, 2005.
de Goede, Marieke, and Samuel Randalls. “Precaution, Preemption: Arts and Tech-
nologies of the Actionable Future.” Environment and Planning D: Society and Space
27, no. 5 (2009): 859 – 78.
DeLanda, Manuel. Intensive Science and Virtual Philosophy. London: Continuum, 2002.
DeLanda, Manuel. Philosophy and Simulation. London: Bloomsbury, 2011.
Deleuze, Gilles. Bergsonism. New York: Zone Books, 1991.
Deleuze, Gilles. Essays Critical and Clinical. Translated by Ariel Greco and Daniel
Smith. Minneapolis: University of Minnesota Press, 1998.
Deleuze, Gilles. Negotiations: 1972 – 1990. Translated by Martin Joughin. New York:
Columbia University Press, 1995.
Deleuze, Gilles. “On the Difference between the Ethics and a Morality.” In Spinoza:
Practical Philosophy, translated by Robert Hurley, 17 – 28. San Francisco, CA: City
Lights, 1988.
Deleuze, Gilles. Pure Immanence: Essays on a Life. New York: Zone Books, 2001.
Deleuze, Gilles. Spinoza: Practical Philosophy. Translated by Robert Hurley. San Fran-
cisco, CA: City Lights, 1988.
Derrida, Jacques. Archive Fever: A Freudian Impression. Chicago: University of Chicago
Press, 1995.
Derrida, Jacques. “Cogito and the History of Madness.” In Writing and Difference,
translated by A. Bass, 31 – 63. London: Routledge, 2001.
Derrida, Jacques. The Death Penalty. Vol. 2. Chicago: University of Chicago Press,
2016.
Derrida, Jacques. “Force of Law: The Mystical Foundations of Authority.” In Decon-
struction and the Possibility of Justice, edited by D. Cornell, M. Rosenfeld, and D. G.
Carlson, 26. London: Routledge, 1992.
Derrida, Jacques. The Gift of Death. Chicago: University of Chicago Press, 1996.
Derrida, Jacques. Limited Inc. Evanston, IL: Northwestern University Press, 1988. 201
Derrida, Jacques. Of Grammatology. Translated by Gayatri Spivak. Baltimore, MD:
Johns Hopkins University Press, 1976.

Bibliography
Derrida, Jacques. Writing and Difference. London: Routledge, 2001.
Devlin, Keith. The Math Gene. London: Weidenfeld and Nicolson, 2000.
“dhs foia Documents Baltimore Protests Freddie Gray.” Scribd. Uploaded by Jason
Leopold. Accessed July 10, 2017. https://www.scribd.com/document/274209838
/DHS-­FOIA-­Documents-­Baltimore-­Protests-­Freddie-­Gray.
Diamond, Cora, ed. Wittgenstein’s Lectures on the Foundations of Mathematics, Cambridge,
1939. Chicago: University of Chicago Press, 1976.
Digital Reasoning. Synthesys Mission Analytics. Franklin, TN: Digital Reasoning, 2015.
Dodge, Martin. “Seeing Inside the Cloud: Some Ways to Map the Internet.” Media
Art Net Lectures, Karlsruhe, January 24, 2004.
Dodge, Martin, and Rob Kitchin. Mapping Cyberspace. London: Routledge, 2001.
Domingos, Pedro. The Master Algorithm: How the Quest for the Ultimate Learning Machine
Will Remake Our World. New York: Allen Lane, 2015.
Dunning, Ted, and Ellen Friedman. Practical Machine Learning: A New Look at Anomaly
Detection. Sebastopol, CA: O’Reilly, 2014.
Dürrenmatt, Friedrich. The Physicists. New York: Grove, 2006.
Eklund, Anders, Thomas Nichols, and Hans Knutson. “Cluster Failure: Why fMRI
Inferences for Spatial Extent Have Inflated False-­Positive Rates.” Proceedings of
the National Academy of Sciences of the United States of America 113, no. 28 (2016):
7900 – 7905.
Elden, Stuart. “Land, Terrain, Territory.” Progress in Human Geography 34, no. 6 (2010):
799 – 817.
Elden, Stuart. Mapping the Present: Heidegger, Foucault and the Project of a Spatial History.
London: Continuum, 2001.
Elden, Stuart. “Secure the Volume: Vertical Geopolitics and the Depths of Power.”
Political Geography 34, no. 1 (2013): 35 – 51.
Emerging Technology from the arXiv. “Twitter Data Mining Reveals the Origins of
Support for Islamic State.” mit Technology Review, March 23, 2015. https://www
.technologyreview.com/s/536061/twitter-­data-­mining-­reveals-­the-­origins-­of
-­support-­for-­islamic-­state/.
Erickson, Paul J., Judy Klein, Lorraine Daston, Rebecca Lemov, Thomas Sturm, and
Michael D. Gordin, eds. How Reason Almost Lost Its Mind: The Strange Career of Cold
War Reasoning. Chicago: University of Chicago Press, 2013.
 Eubanks, Virginia. Automating Inequality: How High-­Tech Tools Profile, Police, and Punish
the Poor. New York: Palgrave Macmillan, 2018.
European Commission. “What Does the Commission Mean by Secure Cloud Com-
puting Services in Europe?” Memo 13/898, Brussels, October 15, 2013. http://
europa.eu/rapid/press-­release_MEMO-­13-­898_en.htm.
Fazi, M. Beatrice. Contingent Computation: Abstraction, Experience, and Indeterminacy in
Computational Aesthetics. Lanham, MD: Rowman and Littlefield, 2018.
Feynman, Richard. Surely You’re Joking, Mr. Feynman!: Adventures of a Curious Character.
202 New York: Norton, 1985.
Feynman, Richard. What Do You Care What Other People Think?: Further Adventures of a
Curious Character. New York: Norton, 1988.
Bibliography

Flinders, Karl. “Swift Builds Underground Datacentre to Keep Bank Transactions in

Europe.” Computer Weekly, April 30, 2012.
“Forensic Biology Protocols for Forensic str Analysis.” City of New York website,
June 20, 2016. http://www1.nyc.gov/assets/ocme/downloads/pdf/technical
-­manuals/protocols-­for-­forensic-­str-­analysis/forensic-­statistical-­tool-­fst.pdf.
Foucault, Michel. The Courage of Truth: Lectures at the Collège de France, 1983 – 84. Basing-
stoke: Palgrave, 2011.
Foucault, Michel. “The Ethics of the Concern for Self as a Practice of Freedom.” In
Ethics: Essential Works of Foucault, 1954 – 1984, edited by Paul Rabinow, 281 – 302. Lon-
don: Penguin, 1997.
Foucault, Michel. The Government of Self and Others: Lectures at the Collège de France,
1982 – 83. Basingstoke: Palgrave, 2010.
Foucault, Michel. History of Madness. Edited by J. Khalfa. Translated by J. Murphy
and J. Khalfa. London: Routledge, 2006.
Foucault, Michel. “On the Genealogy of Ethics: An Overview of Work in Progress.”
In Ethics: Essential Works of Foucault 1954 – 1984, edited by Paul Rabinow, 253 – 80.
London: Penguin, 1997.
Foucault, Michel. “Polemics, Politics, Problematizations.” In Ethics: Essential Works of
Foucault, 1954 – 1984, edited by Paul Rabinow, 115. London: Penguin, 1997.
Foucault, Michel. “Reply to Derrida.” In History of Madness, edited by J. Khalfa, trans-
lated by J. Murphy and J. Khalfa, 575. London: Routledge, 2006.
Foucault, Michel. Security, Territory, Population: Lectures at the Collège de France,
1977 – 1978. London: Palgrave Macmillan, 2004.
Foucault, Michel. “Self Writing.” In Ethics: Essential Works of Foucault, 1954 – 1984, ed-
ited by P. Rabinow, 207 – 22. London: Penguin, 2000.
Foucault, Michel. “What Is an Author?” In Aesthetics, Method, and Epistemology, vol.
2 of Essential Works of Foucault, 1954 – 1984, edited by James Faubion, 205 – 22. New
York: Penguin, 2000.
Foucault, Michel. “What Is Critique?” In The Essential Foucault: Selections from the Es-
sential Works of Foucault, 1954 – 1984, edited by Paul Rabinow and Nikolas Rose,
263 – 78. New York: New Press, 2003.
Foucault, Michel. Wrong-­Doing, Truth-­Telling: The Function of Avowal in Justice. Edited by
Fabienne Brion and Bernard Harcourt. Chicago: University of Chicago Press, 2014.
Fowles, John. “Golding and ‘Golding.’ ” In Wormholes: Essays and Occasional Writings,
edited by Jan Relf, 197 – 207. London: Jonathan Cape, 1998.
Fowles, John. “The Nature of Nature.” In Wormholes: Essays and Occasional Writings,
edited by Jan Relf, 343 – 62. London: Jonathan Cape, 1998.
Fowles, John. “Notes on an Unfinished Novel.” In Wormholes: Essays and Occasional
Writings, edited by Jan Relf, 13 – 26. London: Jonathan Cape, 1998.
Fowles, John, and Susanna Onega. Form and Meaning in the Novels of John Fowles. Lon-
don: vmi Press, 1989.
Fowles, John, and Dianne Vipond. “An Unholy Inquisition.” In Wormholes: Essays 203
and Occasional Writings, edited by Jan Relf, 365 – 84. London: Jonathan Cape,
1998.

Bibliography
Friedberg, Anne. The Virtual Window: From Alberti to Microsoft. Cambridge, MA: mit
Press, 2006.
Fuller, Matthew, and Andrew Goffey. Evil Media. Cambridge, MA: mit Press, 2012.
Galison, Peter. Image and Logic: A Material Culture of Microphysics. Chicago: University
of Chicago Press, 1997.
Galloway, Alexander. The Interface Effect. Cambridge: Polity, 2012.
Gentner, W., H. Maier-­Leibnitz, and W. Bothe. An Atlas of Typical Expansion Chamber
Photographs. London: Pergamon Press, 1954.
Gibbs, Samuel. “Microsoft’s Racist Chatbot Returns with a Drug-­Smoking Twitter
Meltdown.” Guardian (London), March 30, 2016.
Gillespie, Tarleton. “Algorithm.” In Digital Keywords: A Vocabulary of Information Society
and Culture, edited by Benjamin Peters, 18 – 30. Princeton, NJ: Princeton Univer-
sity Press, 2016.
Graham, Stephen. Cities under Siege. London: Verso, 2011.
Graham, Stephen. Vertical: The City from Above and Below. London: Verso, 2016.
Greenfield, Patrick. “The Cambridge Analytica Files: The Story So Far.” Guardian
(London), March 25, 2018.
Greenwald, Glenn. No Place to Hide: Edward Snowden, the nsa, and the Surveillance State.
New York: Random House, 2014.
Gregory, Derek. “Drone Geographies.” Radical Philosophy 183 (2014): 7 – 19.
Grothoff, Christian, and Jens Porup. “The nsa’s skynet Program May Be Killing
Thousands of Innocent People.” Ars Technica, February 16, 2016. http://
arstechnica.co.uk/security/2016/02/the-­nsas-­skynet-­program-­may-­be-­killing
-­thousands-­of-­innocent-­people.
Grusin, Richard. Premediation: Affect and Mediality after 9/11. London: Palgrave Mac-
millan, 2010.
Guizzo, Erico. “Google Wants Robots to Acquire New Skills by Learning from Each
Other.” ieee Spectrum, October 5, 2016.
Halpern, Orit. Beautiful Data: A History of Vision and Reason since 1945. Durham, NC:
Duke University Press, 2014.
Halpern, Paul. The Quantum Labyrinth: How Richard Feynman and John Wheeler Revolu-
tionized Time and Reality. New York: Basic Books, 2017.
Hansen, Mark. New Philosophy for New Media. Cambridge, MA: mit Press, 2006.
 Haraway, Donna. Modest_Witness@Second_Millennium. FemaleMan©_Meets_ Onco-
mouse. New York: Routledge, 1997.
Haraway, Donna. “Situated Knowledges: The Science Question in Feminism and the
Privilege of Partial Perspective.” Feminist Studies 14, no. 3 (1988): 575 – 99.
Haraway, Donna. Staying with the Trouble: Making Kin in the Chthulucene. Durham, NC:
Duke University Press, 2016.
Harding, Luke. The Snowden Files. London: Faber and Faber, 2014.
Hayes, Brian. “Cloud Computing.” Communications of the acm 51, no. 7 (2008): 9 – 11.
204 Hayles, N. Katherine. “How We Became Posthuman: Ten Years on, an Interview
with N. Katherine Hayles.” Paragraph 33, no. 3 (2010): 318 – 30.
Hayles, N. Katherine. How We Became Posthuman: Virtual Bodies in Cybernetics, Litera-
Bibliography

ture, and Informatics. Chicago: University of Chicago Press, 1999.

Hayles, N. Katherine. How We Think: Digital Media and Contemporary Technogenesis.
Chicago: University of Chicago Press, 2012.
Hayles, N. Katherine. My Mother Was a Computer: Digital Subjects and Literary Texts.
Chicago: University of Chicago Press, 2005.
Hayles, N. Katherine. Unthought: The Power of the Cognitive Nonconscious. Chicago:
University of Chicago Press, 2017.
Heidegger, Martin. “Bauen, Wohnen, Denken.” In Vorträge und Aufsätze, 145 – 62.
Pfullingen: Neske, 1954.
Hendrycks, Dan, Kevin Zhao, Steven Basent, Jacob Steinhardt, and David Song.
“Natural Adversarial Examples.” arXiv:1907 – 07174v1 (2018): 2 – 16.
Hinton, Geoffrey. “Deep Neural Networks for Acoustic Modeling in Speech Recog-
nition.” ieee Signals Processing 29, no. 6 (2012): 82 – 97.
Hinton, Geoffrey, Li Deng, Dong Yu, George Dahl, Abdel-­Rahman Mohames,
Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sain-
ath, and Brian Kingsbury. “Deep Neural Networks for Acoustic Modeling in
Speech Recognition.” ieee Signal Processing Magazine 82 (2012): 83.
Hinton, Geoffrey, Simon Osindero, and Yee Whye Teh. “A Fast Learning Algorithm
for Deep Belief Networks.” Neural Computation 18, no. 7 (2006): 1527 – 54.
Hinton, Geoffrey, and Ruslan Salakhutdinov. “Reducing the Dimensionality of Data
with Neural Networks.” Science (2006): 504 – 7.
Hirschburg, J., and C. D. Manning. “Advances in Natural Language Processing.” Sci-
ence 349 (2015): 261 – 66.
Holzmann, G. J. “Points of Truth.” ieee Software, July/August 2015.
House of Commons Science and Technology Select Committee. Algorithms in Decision
Making. London: hmso, 2018.
House of Representatives Committee on Science and Technology. Investigation of the
Challenger Accident. H. R. Rep. No. 99-­1016. 99th Cong., 2d sess. Washington, DC:
gpo. https://www.gpo.gov/fdsys/pkg/GPO-­CRPT-­99hrpt1016/pdf/GPO-­CRPT
-­99hrpt1016.pdf.
Hu, Tung-­Hui. A Prehistory of the Cloud. Cambridge, MA: mit Press, 2015.
Intelligence and Security Committee of Parliament. Privacy and Security: A Modern and
Transparent Legal Framework. London: hmso, 2015.
Intelligence and Security Committee of Parliament. “Transcript of Evidence Given
by the Rt. Hon. Philip Hammond mp (Foreign Secretary).” Privacy and Security
Inquiry, October 23, 2014. http://isc.independent.gov.uk/public-­evidence.
Jaeger, Paul. “Where Is the Cloud? Geography, Economics, Environment and Juris-
diction.” First Monday 14, no. 5 (2009). https://firstmonday.org/ojs/index.php/fm
/article/view/2456/2171.
Jay, Martin. “Cultural Relativism and the Visual Turn.” Journal of Visual Culture 1, no.
3 (2002): 267 – 78.
Jefferson, Geoffrey. “The Mind of Mechanical Man.” British Medical Journal 1, no. 4616 205
(1949): 1105 – 10.
Jochum, Elizabeth, and Ken Goldberg. “Cultivating the Uncanny: The Telegarden

Bibliography
and Other Oddities.” In Robots and Art, edited by Damith Herath and Christian
Kroos, 149 – 75. New York: Springer, 2016.
Josephson, J., and S. Josephson. Abductive Inference: Computation, Philosophy, Technology.
Cambridge: Cambridge University Press, 1996.
Juola, Patrick. “How a Computer Program Helped Show J. K. Rowling Wrote a Cuck-
oo’s Calling.” Scientific American, August 20, 2013.
Kant, Immanuel. Critique of Pure Reason. Cambridge: Cambridge University Press,
1998.
Kant, Immanuel. Groundwork of the Metaphysics of Morals. Translated by H. J. Patton.
New York: Harper, 1964.
Keenan, Thomas. Fables of Responsibility: Aberrations and Predicaments in Ethics and Poli-
tics. Stanford, CA: Stanford University Press, 1997.
Kinsley, Sam. “The Matter of Virtual Geographies.” Progress in Human Geography 38,
no. 3 (2014): 364 – 84.
Kirby, Vicki. Quantum Anthropologies: Life at Large. Durham, NC: Duke University
Press, 2011.
Kirchner, Lauren. “Federal Judge Unseals New York Crime Lab’s Software for Ana-
lyzing dna Evidence.” ProPublica, October 20, 2017.
Kirchner, Lauren. “New York City Moves to Create Accountability for Algorithms.”
Ars Technica, December 19, 2017.
Kirchner, Lauren. “Traces of Crime: How New York’s dna Techniques Became
Tainted.” New York Times, September 4, 2017.
Kitchin, Rob. The Data Revolution: Big Data, Open Data, Data Infrastructures and Their
Consequences. London: Sage, 2014.
Kittler, Friedrich. Discourse Networks 1800/1900. Stanford, CA: Stanford University
Press, 1990.
Kittler, Friedrich. Literature, Media, Information Systems. Amsterdam: G + B Arts Inter-
national, 1997.
Koehn, Philipp. “Europarl: A Parallel Corpus for Statistical Machine Translation.”
Machine Translation Summit, Phuket, Thailand, September 12 – 16, 2005.
Konkel, F. “The cia’s Deal with Amazon.” Atlantic, July 17, 2014.
Kourou, Konstantina, Themis P. Exarchos, Konstantinos P. Exarchos, Michalis V.
Karamouzis, and Dimitrios I. Fotiadis. “Machine Learning Applications in Can-
 cer Prognosis and Prediction.” Computational and Structural Biotechnology Journal 13
(2015): 8 – 17.
Krizhevsky, Alex, Ilya Sutskever, and Geoffrey Hinton. “ImageNet Classification
with Deep Convolutional Neural Networks.” Advances in Neural Information Pro-
cessing Systems 2 (2012): 1097 – 1105.
Kurzweil, Ray. The Singularity Is Near: When Humans Transcend Biology. London: Viking,
2005.
Lange, Ann-­Christina, Marc Lenglet, and Robert Seyfert. “Cultures of High-­
206 Frequency Trading: Mapping the Landscape of Algorithmic Developments in
Contemporary Financial Markets.” Economy and Society 45 (2016): 149 – 65.
Latour, Bruno, and Peter Weibel. Making Things Public: Atmospheres of Democracy. Cam-
Bibliography

bridge, MA: mit Press, 2005.

Laughland, Oliver. “Baltimore Unrest: 49 Children Were Arrested and Detained
during Protests.” Guardian (London), May 7, 2015.
LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep Learning.” Nature 521,
no. 28 (2015): 436 – 4 4.
Lemov, Rebecca. Database of Dreams: The Lost Quest to Catalog Humanity. New Haven,
CT: Yale University Press, 2015.
Leopold, G. “cia Embraces Cloudera Data Hub.” EnterpriseTech, February 26, 2015.
Lewis, Paul. “Fiction Is Outperforming Reality: How YouTube’s Algorithm Distorts
the Truth.” Guardian (London), February 3, 2018.
Lin, Henry, Izhak Shafran, Todd Murphy, Allison Okamura, David Yuh, and Greg-
ory Hager. “Automatic Detection and Segmentation of Robot-­Assisted Surgical
Motions.” Medical Image Computing 8, no. 1 (2005): 802 – 10.
Liptak, Adam. “Sent to Prison by a Software Program’s Secret Algorithms.” New York
Times, May 1, 2017. https://www.nytimes.com/2017/05/01/us/politics/sent-­to
-­prison-­by-­a-­software-­programs-­secret-­algorithms.html.
Liu, Lydia H. The Freudian Robot: Digital Media and the Future of the Unconscious. Chi-
cago: University of Chicago Press, 2010.
“Machine Learning for Cybersecurity.” Cloudera. Accessed April 2018. https://www
.cloudera.com/solutions/gallery/digital-­reasoning-­synthesys.html.
MacKay, David J. C. Information Theory, Inference, and Learning Algorithms. Cambridge:
Cambridge University Press, 2017.
Mackenzie, Adrian. Machine Learners: Archaeology of a Data Practice. Cambridge, MA:
mit Press, 2017.
MacKenzie, Donald. “How Algorithms Interact: Goffman’s ‘Interaction Order’ in
Automated Trading.” Theory, Culture and Society 36, no. 2 (2019): 39 – 59.
MacKenzie, Donald. “How to Make Money in Microseconds.” London Review of Books
33, no. 10 (2011): 16 – 18.
Mahler, J., J. Liang, S. Niyaz, M. Laskey, R. Doan, X. Liu, J. A. Ojea, and K. Goldberg.
“Dex-­Net 1.0: A Cloud-­Based Network of 3D Objects for Robust Grasp Planning
Using a Multi-­A rmed Bandit Model with Correlated Rewards.” Paper presented
at the International Conference on Robotics and Automation, Stockholm, May
16 – 21, 2016.
Manovich, Lev. The Language of New Media. Cambridge, MA: mit Press, 2002.
Mantel, Hilary. “The Day Is for the Living: Hilary Mantel on Writing Historical
Fiction.” bbc Reith Lectures, June 13, 2017. Medium. https://medium.com
/@bbcradiofour/hilary-­mantel-­bbc-­reith-­lectures-­2017-­aeff8935ab33.
Markoff, John. “Artificial Intelligence Is Far from Matching Humans, Panel Says.”
New York Times, May 26, 2016. https://www.nytimes.com/2016/05/26/technology
/artificial-­intelligence-­is-­far-­from-­matching-­humans-­panel-­says.html.
Markov, A. A. Theory of Algorithms. Translated by Jacques Schorr-­Kon. Moscow:
Academy of Sciences of the USSR, 1954. 207
Mattern, Shannon. Code and Clay, Data and Dirt: Five Thousand Years of Urban Media.
Minneapolis: University of Minnesota Press, 2017.

Bibliography
Mayer-­Schönberger, Viktor, and Kenneth Cukier. Big Data: A Revolution That Will
Change How We Live, Work, and Think. New York: Houghton Mifflin Harcourt, 2013.
McCarthy, John. “Architects of the Information Society: Thirty-­Five Years of the
Laboratory for Computer Science at mit.” Lecture at the mit Centennial, July
1961.
McCormack, Derek. “Remotely Sensing Affective Afterlives: The Spectral Geogra-
phies of Material Remains.” Annals of the Association of American Geographers 100,
no. 3 (2010): 640 – 54.
McGeoch, Catherine C. A Guide to Experimental Algorithmics. Cambridge: Cambridge
University Press, 2012.
Mehotra, Kartikay. “Maker of Surgical Robot Hid More than 700 Injury Claims, In-
surer Says.” Bloomberg, May 26, 2016.
“Memorandum in Support of Application by ProPublica for Leave to Intervene, Lift
the Protective Order and Unseal Judicial Records.” Filed September 25, 2017. Yale
Law School website, https://law.yale.edu/system/files/area/clinic/document/139
._memorandum_of_law.pdf.
Metcalf, Jacob, Emily F. Keller, and danah boyd. “Perspectives on Big Data, Ethics,
and Society.” Report for the Council for Big Data, Ethics, and Society, May 23,
2016. https://bdes.datasociety.net/wp-­content/uploads/2016/05/Perspectives-­on
-­Big-­Data.pdf.
Misra, C., P. K. Swain, and J. K. Mantri. “Text Extraction and Recognition from Im-
age Using Neural Network.” International Journal of Computer Applications 40, no. 2
(2012): 13 – 19.
Mitchell, David. Cloud Atlas. London: Hodder and Stoughton, 2004.
Mitchell, W. J. T. “There Are No Visual Media.” Journal of Visual Culture 4, no. 2
(2005): 257 – 66.
Mol, Anne Marie. The Body Multiple: Ontology in Medical Practice. Durham, NC: Duke
University Press, 2002.
“Monitoring Protests from Unstructured Text.” 29:54, YouTube, 2012. https://www
.youtube.com/watch?v=6Sbny91NjeA.
National Research Council of the National Academies. Bulk Collection of Signals Intelli-
gence: Technical Options. Washington, DC: National Academies Press, 2015.
Nature Editors. “More Accountability for Big Data Algorithms.” Nature 537, no. 7621
 (2016): 449. https://www.nature.com/news/more-­accountability-­for-­big-­data
-­algorithms-­1.20653.
Newman, M. H. A., A. Turing, G. Jefferson, and R. B. Braithwaite. “Can Automatic
Calculating Machines Be Said to Think?” Transcript of discussion, BBC, recorded
1951, aired January 14, 23, 1952. http://www.turingarchive.org/browse.php/b/6.
Nissenbaum, Helen. Privacy in Context: Technology, Policy, and the Integrity of Social Life.
Stanford, CA: Stanford University Press, 2009.
Noble, Safiya. Algorithms of Oppression: How Search Engines Reinforce Racism. New York:
208 nyu Press, 2018.
nsf Award Abstract 1734521. Last amended October 27, 2017. https://www.nsf.gov
/awardsearch/showAward?AWD_ID=1734521&HistoricalAwards=false.
Bibliography

Nyholm, Sven, and Jilles Snids. “The Ethics of Accident-­A lgorithms for Self-­Driving
Cars.” Ethical Theory and Moral Practice 19, no. 5 (2016): 1275 – 89.
O’Brien, John. The Definitive Guide to the Data Lake. Boulder, CO: Radiant Advisors, 2015.
Office of the Director of National Intelligence. “icite Fact Sheet.” Washington, DC:
odni, 2012.
Office of the Director of National Intelligence. Intelligence Community Information
Technology Enterprise Strategy 2016 – 2020. Washington, DC: odni, 2015.
Ohri, A. R for Cloud Computing: An Approach for Data Scientists. London: Springer, 2014.
O’Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and
Threatens Democracy. New York: Crown, 2016.
Ormandy, Roman. “Google and Elon Musk to Decide What Is Good for Humanity.”
Wired, January 26, 2015.
Paglen, Trevor. Invisible: Covert Operations and Classified Landscapes. New York: Aper-
ture, 2010.
Paglen, Trevor. “Trevor Paglen Offers a Glimpse of America’s Vast Surveillance Infra-
structure.” American Suburb X, December 30, 2014. http://www.american
suburbx.com/2014/12/trevor-­paglen-­offers-­a-­glimpse-­of-­americas-­vast-­surveillance
-­infrastructure.html.
Parikka, Jussi. A Geology of Media. Minneapolis: University of Minnesota Press, 2015.
Parikka, Jussi. Insect Media: An Archaeology of Animals and Technology. Minneapolis: Uni-
versity of Minnesota Press, 2010.
Parisi, Luciana. Contagious Architecture: Computation, Aesthetics, and Space. Cambridge,
MA: mit Press, 2013.
Parkin, Simon. “Teaching Robots Right from Wrong.” Economist 1843 Magazine, June/
July 2017. https://www.1843magazine.com/features/teaching-­robots-­right
-­from-­wrong.
Parks, Lisa, and Caren Kaplan. Life in the Age of Drone Warfare. Durham, NC: Duke
University Press, 2017.
Pasquale, Frank. The Black Box Society: The Secret Algorithms That Control Money and In-
formation. Cambridge, MA: Harvard University Press, 2015.
Peters, John Durham. The Marvelous Clouds: Toward a Philosophy of Elemental Media.
Chicago: University of Chicago Press, 2015.
Phys.org. “UMass Lowell Professor Steers Ethical Debate on Self-­Driving Cars.”
UMass Lowell News, October 5, 2017. https://www.uml.edu/News/news-­articles
/2017/phys-­self-­driving-­ethics.aspx.
Pickett, Marc, Chris Tar, and Brian Strope. “On the Personalities of Dead
Authors.” Google ai Blog, February 24, 2016. https://ai.googleblog.com/2016/02
/on-­personalities-­of-­dead-­authors.html.
Pishko, Jessica. “The Impenetrable Program Transforming How Courts Treat dna
Evidence.” Wired, November 29, 2017.
Popescu, M. C., and N. Mastorakis. “Simulation of da Vinci Surgical Robot Using 209
Mobotsim Program.” International Journal of Biology and Biomedical Engineering 4,
no. 2 (2008): 137 – 46.

Bibliography
Prentice, Rachel. Bodies in Formation: An Ethnography of Anatomy and Surgery Education.
Durham, NC: Duke University Press, 2013.
Puar, Jasbir. Terrorist Assemblages: Homonationalism in Queer Times. Durham, NC: Duke
University Press, 2007.
Quinlan, J. R. “Induction of Decision Trees.” Machine Learning 1, no. 1 (1986): 81 – 106.
Reiley, Carol, Erion Plaku, and Gregory Hager. “Motion Generation of Robotic Surgi-
cal Tasks: Learning from Expert Demonstrations.” Paper presented at the Annual
International Conference of the ieee Engineering in Medicine and Biology, Bue-
nos Aires, Argentina, August 31 – September 4, 2010.
Rochester, G. D., and J. G. Wilson. Cloud Chamber Photographs of the Cosmic Radiation.
London: Pergamon Press, 1952.
Rose, Gillian. “Rethinking the Geographies of Cultural ‘Objects’ through Digital
Technologies: Interface, Network, Friction.” Progress in Human Geography 40, no. 3
(2015): 334 – 51.
Rotman, Brian. Becoming Beside Ourselves: The Alphabet, Ghosts, and Distributed Human
Being. Durham, NC: Duke University Press, 2008.
Said, Edward. On Late Style: Music and Literature against the Grain. London: Blooms-
bury, 2007.
Schneier, Bruce. “nsa Robots Are Collecting Your Data, Too, and They’re Getting
Away with It.” Guardian (London), February 27, 2014.
Scorer, Richard, and Harry Wexler. Cloud Studies in Colour. Oxford: Pergamon Press,
1967.
Seaver, Nick. “Algorithms as Culture: Some Tactics for the Ethnography of Algorith-
mic Systems.” Big Data and Society 3, no. 2 (2017): 2.
Shane, Scott, and Daisuke Wakabayashi. “The Business of War: Google Employees
Protest Work for the Pentagon.” New York Times, April 4, 2018.
Sharkey, Noel. “Staying in the Loop: Human Supervisory Control of Weapons.” In
Autonomous Weapons Systems: Law, Ethics, Policy, edited by Nehal Bhuta, Susanne
Beck, Robin Geiss, Hin-­Yan Liu, and Claus Kress, 23 – 38. Cambridge: Cambridge
University Press, 2016.
Shaw, Ian, and Majed Akhter. “The Dronification of State Violence.” Critical Asian
Studies 46, no. 2 (2014): 211 – 3 4.
Simonyan, Karen, and Andrew Zisserman. “Very Deep Convolutional Networks for
 Large-­Scale Image Recognition.” arXiv:1409.1556 (September 2014). https://arxiv
.org/abs/1409.1556v1.
Stengers, Isabelle. Cosmopolitics I. Minneapolis: University of Minnesota Press, 2010.
Stengers, Isabelle. Thinking with Whitehead: A Free and Wild Creation of Concepts. Cam-
bridge, MA: Harvard University Press, 2011.
Strawser, B. J. Killing by Remote Control: The Ethics of an Unmanned Military. Oxford: Ox-
ford University Press, 2013.
Stubb, Alexander. “Why Democracies Should Care Who Codes the Algorithms.” Fi-
210 nancial Times, June 2, 2017.
Suchman, Lucy, and Jutta Weber. “Human-­Machine Autonomies.” In Autonomous
Weapons Systems: Law, Ethics, Policy, edited by N. Bhuta, Susanne Beck, Robin
Bibliography

Geiss, Hin-­Yan Liu, and Claus Kress, 75 – 102. Cambridge: Cambridge University
Press, 2016.
Taylor, Duncan A., Jo-­A nne Bright, and John Buckleton. “A Source of Error: Com-
puter Code, Criminal Defendants, and the Constitution.” Frontiers in Genetics 8
(2017): 33.
Timmer, John. “Software Faults Raise Questions about the Validity of Brain Studies.”
Ars Technica, July 1, 2016.
Totaro, Paulo, and Domenico Ninno. “The Concept of Algorithm as an Interpreta-
tive Key of Modern Rationality.” Theory, Culture and Society 31, no. 4 (2014): 32.
Truvé, Staffan. “Building Threat Analyst Centaurs Using Artificial Intelligence.”
Recorded Future, January 14, 2016. https://www.recordedfuture.com/artificial
-­threat-­intelligence/.
Tucker, Patrick. “A New ai Learns through Observation Alone.” Defense One, Sep-
tember 6, 2016.
Turing, Alan. “Letters on Logic to Max Newman.” In The Essential Turing, edited by
B. Jack Copeland, 205 – 16. Oxford: Oxford University Press, 2004.
Turing, Alan. “Systems of Logic Based on Ordinals.” Proceedings of the London Math-
ematical Society 2, no. 45 (1939): 161 – 228.
Turkle, Sherry. The Second Self: Computers and the Human Spirit. Cambridge, MA:
mit Press, 2005.
van den Berg, J., S. Miller, D. Duckworth, H. Hu, A. Wan, X. Fu, K. Goldberg, and
P. Abbeel. “Superhuman Performance of Surgical Tasks by Robots Using Iterative
Learning.” Paper presented at the ieee International Conference on Robotics and
Automation, Anchorage, Alaska, May 3 – 7, 2010.
Vaughan, Diane. The Challenger Launch Decision: Risky Technology, Culture and Deviance
at nasa. Cambridge, MA: mit Press, 1996.
Vinyals, Oriol, Alexander Toshev, Samy Bengio, and Dumitru Erhan. “Show and Tell:
A Neural Image Caption Generator.” ieee Transactions on Pattern Analysis and Ma-
chine Intelligence 39, no. 4 (2017): 2048 – 57.
Walker, James. “Researchers Shut Down ai That Invented Its Own Language.”
Digital Journal, July 21, 2017. http://www.digitaljournal.com/tech-­and-­science
/technology/a-­step-­closer-­to-­skynet-­ai-­invents-­a-­language-­humans-­can-­t-­read
/article/498142.
Weber, Jutta. “Keep Adding: On Kill Lists, Drone Warfare, and the Politics of Data-
bases.” Environment and Planning D: Society and Space 34, no. 1 (2016).
Weber, Samuel. Targets of Opportunity: On the Militarization of Thinking. New York:
Fordham University Press, 2005.
“Web Intelligence.” Recorded Future. Accessed January 2018. https://www.recorded
future.com/web-­intelligence/.
Webster, Katharine. “Philosophy Professor Wins nsf Grant on Ethics of Self Driving
Cars.” UMass Lowell News, August 8, 2017. https://www.uml.edu/news/stories/2017
/selfdrivingcars.aspx. 211
Weizman, Eyal. “The Politics of Verticality: The West Bank as an Architectural
Construction.” In Territories: Islands, Camps, and Other States of Utopia, edited by

Bibliography
A. Franke, 106 – 2 4. Cologne: Walther Koenig, 2004.
Wiener, Norbert. “A Scientist Rebels.” Atlantic Monthly, January 1947.
Wilcox, Lauren. “Embodying Algorithmic War: Gender, Race, and the Posthuman in
Drone Warfare.” Security Dialogue 48, no. 1 (2017): 11 – 28.
Wilson, C. T. R. “On the Cloud Method of Making Visible Ions and the Tracks of
Ionizing Particles.” In Nobel Lectures in Physics, 1922 – 1941, 167 – 217. Amsterdam:
Elsevier, 1927.
Wilson, J. G. The Principles of Cloud Chamber Technique. Cambridge: Cambridge Uni-
versity Press, 1951.
Winslett, Marianne. “Rakesh Agrawal Speaks Out.” Interview with Rakesh Agrawal.
sigmod Record 32, no. 3 (2003). https://sigmod.org/publications/interviews/pdf
/D15.rakesh-­final-­final.pdf.
Wittgenstein, Ludwig. On Certainty. Edited by G. E. M. Anscombe and G. H. von
Wright. Oxford: Blackwell, 1975.
Wittgenstein, Ludwig. Philosophical Investigations. Translated by G. E. M. Anscombe,
P. M. S. Hacker, and Joachim Schulter. Oxford: Wiley Blackwell, 2008.
Woodman, Spencer. “Palantir Provides the Engine for Donald Trump’s Deportation
Regime.” Intercept, March 2, 2017.
Xu, K., J. Lei Ba, R. Kiros, K. Cho, A. Courville, R. Salakhutdinov, R. Zemel, and
Y. Bengio. “Show, Attend and Tell: Neural Image Caption Generation and Visual
Attention.” Paper presented at the 32nd International Conference on Machine
Learning, Lille, France, July 6 – 11, 2015.
Yang, A. Y., X. Xhou, A. G. Balasubramanian, S. S. Sastry, and Y. Ma. “Fast 𝓁1-
­Minimization Algorithms for Robust Facial Recognition.” Transactions on Image
Processing 22, no. 8 (2013): 3234 – 46.
Zhang, Q., L. Cheng, and R. Boutaba. “Cloud Computing: State-­of-­the-­A rt and Chal-
lenges.” Journal of Internet Services and Applications 1 (2010): 7 – 18.
IN DE X

aberration, 108 – 29 art, 5

abduction, 47 – 48, 75, 161 – 62 artificial intelligence (ai), 109, 146
abnormality, 6, 43, 47, 68, 78 – 79, 92, 97, associative life, 7
115, 139. See also norms atomic weapons, 112, 133
accidents, 75 – 76, 115 – 16, 119 attention, 15, 42, 154 – 55, 157, 162
accountability: algorithms and, 5, 11, attributes: and attribution, 41, 85 – 107,
18 – 20, 38, 85, 87, 166; episteme of, 118; and clustering of data, 1, 4, 33,
8 – 9; in robot surgery, 66; of science, 48 – 49, 70, 122, 126, 160; racialized,
134 11; and selves and others, 8, 20, 43, 58,
accounts: giving of, 18 – 20, 62, 80, 135, 79, 138, 148; and the unattributable,
140 – 48, 152, 164 – 68 25, 154 – 72
action, 16, 19, 41, 46, 109 – 10; theory of author: algorithm designer as, 18; and
as a “spring,” 59 – 60, 68, 76 – 77 authorship, 85 – 107, 135
AlexNet, 72 – 74, 77, 79, 136 automation, 15, 18, 33, 85
algorithms: accounts of themselves, autonomous machines, 57, 65, 90, 109
18 – 20, 54, 72, 102, 135, 148, 171; as ap- autonomous vehicles, 76, 119 – 22
erture instruments, 15 – 18, 40, 156; as autonomous weapons, 11, 66, 175n33,
arrangements of propositions, 6 – 7, 9, 182n30
11 – 1 4, 40, 57, 87, 96, 99, 104, 136, 145,
158, 162; as composites, 10, 20, 42, 57, back propagation, 66, 106
64, 66, 88, 135; deep learning, 3, 89, Barad, Karen: on algorithms, 141; on in-
136, 150; designers of, 18, 67 – 68, 139, telligibility, 42; on mattering, 10, 46,
147; and experimentation, 12, 40 – 4 1, 144, 152; on touch, 63
43, 48, 53, 67 – 68, 92, 118 – 19, 139, 161; Bergson, Henri, 16, 42, 152, 163
and layers, 13, 35; and modification, 11, Berlant, Lauren, 147
48, 74, 87, 96 – 97, 106, 137; and mul- bias: algorithmic, 5, 8, 18 – 19, 38, 87, 92,
tiplicity, 13, 17, 64, 80, 87, 121, 152, 162; 95 – 96, 146; and gender, 11; and ma-
and steps, 11, 13, 40, 69, 89 chine learning, 74 – 75, 79, 81, 106, 125
Amazon Web Services (aws), 33, 35, 49 bifurcation, 88, 98 – 99, 113
animals, 118, 137, 157 big data, 16, 18, 34 – 35, 77. See also data
anomaly detection, 40, 42 – 43, 68, 79. biometrics, 3, 4, 14, 65, 70, 135, 147. See
See also patterns also facial recognition
apertures, 15 – 18, 42 – 4 4, 46, 48, 71, 94, black box, 5, 40, 73, 110
104 – 5, 160 – 64 borders, 33, 46 – 47, 54, 67 – 68, 112, 155;
application programming interface and immigration, 12, 55, 139, 161
(api), 59 Bratton, Benjamin, 35, 40
applications, 35, 49 Breiman, Leo, 125
architecture, 14, 33, 35, 64, 111, 136 Browne, Simone, 4
archiving, 49 – 51 Butler, Judith, 8, 19, 38, 135, 145, 151
calculation: algorithmic forms of, 11, 21, Cold War, 42, 113 – 1 4, 140
33, 42, 46, 75, 95, 112; archiving of data communication, 37
for, 50, 54; and decision-­making, 120, computation, 7, 34 – 35, 44, 137, 139, 165
128, 140 – 41, 143, 150; and incalculabil- computer science, 11, 34 – 35, 71 – 72,
ity, 162; incorporation of doubt in, 134, 88, 93, 104, 106, 113, 137, 135. See also
142, 151; probabilistic methods of, 125 science
Cambridge Analytica, 4 – 5, 118, 160 condensation, 16, 29 – 30, 43, 44 – 46, 49,
cancer, 123 77, 99, 139, 156, 162
categorization, 14 contingency, 21, 68, 98, 146 213
Central Intelligence Agency (cia), 31, control, 58, 65, 79, 113
33, 50 correlation, 31, 40, 44, 47, 50, 107, 138,

Index
Challenger (space shuttle), 140 – 53 161
cities, 1 – 4 , 20, 85, 155 counterterrorism. See terrorism
classification, 33, 40, 49, 92, 115 – 19; and Crary, Jonathan, 15, 39, 162
classifiers, 9, 11, 14, 68, 70, 125, 139; credit scoring, 18
and clouds, 29 criminal justice, 11, 85, 95, 104, 137
cloud atlas, 44, 48 – 49 Critchley, Simon, 100, 106, 171
Cloud Atlas (Mitchell), 55 crowd: algorithms to assess risk in, 4, 17,
cloud chamber, 29 – 31, 44, 48, 50, 52 – 53 155, 159 – 60; crowded court, 25, 69
cloud computing, 7, 30 – 31, 33, 35, 38, cybernetics, 111 – 15, 140
40 – 4 2, 46 – 47
cloud data: algorithmic analysis of, 7, Daston, Lorraine, 125
42, 49, 62, 79; architectures of, 21, 70, data: and algorithms, 7, 18, 49, 50, 86,
72, 76 – 77; medical images as, 59; stor- 88, 92, 96, 134 – 35, 138, 141 – 53; and fea-
age of, 37 tures, 12, 16, 20, 43, 155, 160; sharing
Cloudera, 33 of, 33, 42, 50; social media and, 1 – 2,
cloud ethics: and doubt, 135, 137 – 38; and 37, 42, 47, 89; storage of, 36, 38; un­
encoded ethics, 121, 123; and partial labeled, 90. See also big data
account, 141; and political claims, data analysis, 31, 33
80 – 81, 146, 148; theory of, 7 – 8, 17, 19, data centers, 33 – 35, 38 – 39
88, 117, 154 – 72. See also ethics data environment, 58, 75
cloud forms, 40 data inputs, 11, 48, 58, 90, 105, 125, 127, 160
cloud imaginations, 36 data lake, 47
cloud robotics, 76 data mining, 36, 179n63
clustering: in anomaly detection, 40, data residue, 46, 59, 64, 164
42 – 4 4, 92, 122, 136; of data for protest da Vinci robot, 59, 63 – 64
detection, 1 – 4 , 169; in elections, 161; decision procedures, 42, 57, 69, 136, 139
and false positives, 128; mapping of decisions: and automation, 11, 57, 67, 85,
features via, 58, 70, 155; in natural lan- 102; doubt and, 134, 137, 150; human,
guage processing, 89 – 9 0; and predic- 22, 62, 65 – 66, 140 – 48; and machine
tions, 138, 173n9 learning, 12, 71, 80, 87, 92, 106; mad-
code, 1, 7, 9, 40, 59; source, 18, 85 – 88, ness of, 112 – 1 4, 119 – 23; and undecid-
91 – 92, 95 – 96, 99, 104, 186n34. See also ability, 19, 20, 88, 98 – 99, 128, 151 – 53,
writing 162 – 65
 decision trees, 10, 70, 89 112 – 113, 125, 165. 88, 90, 105, 111, 118, 135, 147 – 53, 154 – 72;
See also random forests and opacity, 5, 14. See also cloud ethics
deconstruction, 88, 99, 104 European Union, 36
deep learning. See machine learning; experimentation, 12, 30, 40 – 4 1, 43 – 4 4,
neural networks 48, 53, 67, 90, 161
Deleuze, Gilles: on duration, 17, 163; on expertise, 59 – 60, 96
ethics, 165; on fabulation, 98, 102, 158, explainability, 18
161 – 62 extraction. See feature extraction
214 democracy, 4, 38, 118, 160
derivatives, 50, 106, 124 – 25 fabulation, 98 – 99, 102 – 5, 158, 160
Derrida, Jacques: on archives, 51; on the Facebook, 37, 50, 89, 138, 160
Index

author, 99; on calculation, 128; on de- facial recognition, 14, 48, 68 – 70, 135 – 36.
cision, 121, 149, 152; on ethics, 166, 169; See also recognition
on madness, 110, 116; on nonclosure, false positive, 122, 145
104, 154, 156; on text, 104 fascism, 112
DexNet, 77 feature extraction, 14, 16, 58, 160; pro-
the digital, 35 cesses of, 42 – 43, 59, 63, 89 – 91, 139, 154
digital subjects, 94 feature space, 58, 74, 79, 155, 160, 169,
discrimination, 8, 38, 85 1 81n8, 182n24
dna, 95 – 97, 105 feature vectors, 59, 72, 155, 160, 167, 169,
doubt, 9, 10, 13, 133 – 53 171, 181n10
drones, 33, 47, 78, 80, 124 – 26, 176n55; feedback loop, 66, 97, 106
image recognition systems and, 16, feminism, 88, 19, 141
38, 111, 146, 156, 160; targeting of, 55, Feynman, Richard, 133 – 53
127, 128 finance. See economy
forensics, 95
economy, 35, 43, 46, 49, 89, 108, 123 forking pathways, 88, 98 – 99, 118, 139,
embodiment, 57, 59, 62, 64, 137 – 38, 143 141, 153, 161 – 62, 165
entanglement, 16, 57, 59 Foucault, Michel: on authorship, 18,
error and errancy, 47 – 48, 108 – 12, 115, 86 – 87; on ethics, 7, 10, 80 – 81, 171; on
116, 120 – 23, 136, 146 – 47; experimen- madness, 115 – 18; on parrhesia, 145 – 47;
tation and, 74 – 75, 96, 118, 125, 159, on probable events, 44; on truth-­
169; in robot surgery, 58; sources of, telling, 5, 122
20, 86 Fowles, John, 85, 88, 98 – 99, 101
ethicopolitics: of algorithms, 7 – 8, 10 – 11, fraud, 46, 53, 68, 91, 124
47, 53, 58, 71, 87, 88, 110, 146, 156, 165, the future: algorithms and, 7, 17, 46,
171 – 72; bias and, 75, 79; materiality 49 – 5 1, 122; anticipation of, 72, 73;
and, 36; theories of, 19 – 20, 33, 65 – 66, inferring, 88 – 9 1; optimization of,
69, 121; writing and, 22, 92, 105. See 62, 80, 147; past data and, 20, 43,
also politics 46, 138; political claims on, 4, 54,
ethics: codes of, 5 – 7, 15, 19, 86, 95, 110, 152, 153, 161
119, 120 – 21, 128, 146; of deconstruc-
tion, 104, 106; fabulation and, 98 – 99; Galison, Peter, 30, 38, 40 – 4 1, 44, 49
machine learning and, 55 – 56, 66, 80, gender, 85
geography: of cloud computing, 34 – 36, incompleteness, 21, 43, 80, 96
38, 40; of data, 33 the incomputible, 64, 167
geopolitics, 33, 36 – 37, 114, 137 indeterminacy, 14, 87, 96, 156
gestures, 59, 60, 62, 118 inference, 43 – 4 4, 74, 89, 155; causality
Github, 97 and, 47, 107, 122; of intent, 7, 91, 127;
good: appearance of, 78 – 79, enough, 67, of motion, 46
69, 75, 161; and evil, 58, 109 – 10, 112, infrastructure, 30, 33, 35, 39
134, 146, 158 insurance, 138
Google, 17, 35, 77, 89 – 9 0, 92, 146, 154 intelligence: geopolitical, 15, 31, 33, 54; 215
government, 1, 33, 36, 42 – 43, 49, 113, surveillance and, 37 – 39, 42, 47, 126,
140, 149 148

Index
grammar, 13 Intelligence Community Information
ground truth, 4, 133 – 53, 169. See also Technology Enterprise (icite), 31, 42,
truth 49, 54
internet, 30, 37, 50
Halpern, Orit, 15, 50, 54, 114 intuition, 13, 57, 64, 67, 145, 150, 166
Haraway, Donna: on partiality, 20, 135, iteration, 11, 43, 92, 96 – 97, 139; and
137, 143, 166; on staying with the trou- learning, 60, 62, 68, 106, 138
ble, 19, 40, 66, 178n35, 182n31; on tech-
noscience, 133 – 3 4 Keenan, Thomas, 19, 148 – 49, 152, 165
Hayles, N. Katherine, 137; on cognition, knowledge discovery, 47
19, 43; on computation, 4, 173n8; on
data, 49; on the posthuman, 65 – 66 language, 9, 43, 89. See also natural lan-
health, 123 guage processing
Hinton, Geoffrey, 74 latency, 90
Homeland Security, US Department of law, 37 – 38, 66, 95 – 96
(dhs), 2, 31, 43, 50 layers, 13, 35, 62, 85; hidden, 71, 162
human, humans, 57, 60, 62, 64, 79, learning, 8 – 9, 56, 59, 90, 122. See also
88, 94; agency, 65, 112, 135; and algo- machine learning
rithm relations, 9, 19 – 20, 42 – 43, 57, likelihood. See probability
63 – 64, 92, 139; labeling of images, literature: authorship and, 87, 98 – 99,
17, 72, 136, 148; in the loop, 11, 65 – 67, 102; as training data for machine
109, 127, 152; subjects, 5, 118; visibil- learning, 89, 92 – 93
ity, 14, 30, 38. See also posthuman, Los Alamos, New Mexico, 112, 133, 140
posthumanism
machine learning: anomaly detection
icite, 31, 42, 49, 54 and, 42 – 43, 53; deep, 89, 156 – 57; logics
ImageNet, 17, 73, 173n7 of, 14, 40, 87, 95 – 96, 118, 154, 157, 158,
image recognition, 14, 17, 70 – 71, 73, 121, 160, 173n9; materiality and, 46 – 47;
137. See also recognition processes of, 8 – 9, 11, 20, 54, 56 – 81,
images, 30, 44, 48, 91, 155; drone, 16, 38; 133 – 53, 173n7, 181n10; semi-­supervised,
medical, 62 – 63, 122 – 23, 160 68, 124; and society, 5, 12, 109, 169,
incalculability, 92, 110, 120, 141, 143, 155, 170; unsupervised, 68 – 69, 111
162 Mackenzie, Adrian, 12
 madness, 108 – 29 74 – 75, 80, 96, 97, 102, 104, 107, 111,
Mantel, Hilary, 102, 107 123, 128, 145, 155 – 56, 160, 162; decision
matching, 90 and, 7; and optimization, 10, 12 – 13, 17,
mathematics, 9, 13, 42, 56, 62, 94, 112, 43, 54, 60, 64, 147; and target, 9, 43,
140 67 – 68, 90, 105, 125, 139; and truth, 14,
Mechanical Turk, 72, 136, 148 114, 136 – 37. See also targets
media, 7, 15, 24, 59, 111, 178n38
Microsoft, 108 – 9 Palantir, 4, 139, 147
216 migration, 129 parameters, 48, 62, 68, 80, 111, 121, 126,
mistakes. See accidents; error and 161
errancy Parisi, Luciana, 12, 47, 147
Index

metadata. See data partiality, 8, 14, 123; Butler’s conceptual-

models, 5, 59, 67 – 68, 89 – 9 0, 111, 142, 162 ization of, 19, 135; Haraway’s concep-
moral law, 6 tualization of, 20, 135; of images, 72;
moral panic, 16, 57, 110 science and, 134; writing and, 9
moral philosophy, 119 – 20 particles, 30, 44, 48
moral responsibility. See responsibility pathways, 88, 98 – 99, 118, 139, 141, 153,
movement, 59, 62 161 – 62, 165
patterns, 6, 17, 43 – 4 4, 46 – 49, 89, 126.
National Security Agency (nsa), 31, 37, See also anomaly detection
39, 43, 50, 124 perception, 15 – 18, 29 – 30, 33, 40 – 4 1
natural language processing (nlp), performativity, 13, 104, 145
50, 88 – 9 0, 92, 155 – 56, 167, 185n10, photography. See images
185–86n12. See also language physics, 29, 40, 44, 48, 112, 134, 141
nearest neighbor, 77 pixels, 70, 73 – 74, 160, 181n10
neural networks: convolutional, 14, 72, platform, 35, 49, 97, 109
77, 154 – 55, 160, 162, 166; deep, 10, play, 68, 71, 79, 80
56 – 81, 89, 94, 105 – 6, 110, 114, 136, 137, point clouds, 75 – 76, 184n56
139, 161; drones and, 176n55; training police, 2, 33, 94
of, 9, 16 – 17, 89, 111, 137, 192n13 politics, 4, 7, 8, 33, 36, 48, 55, 170; of al-
New York, 85, 91, 93, 95 gorithms, 7, 15, 42 – 43, 87, 139, 153, 160,
norms, 6, 13, 48, 136. See also abnormality 172; of appearance, 5, 15, 46, 72 – 73,
79, 155; and difficulty, 10, 152; and fore-
objectivity, 9, 20, 29, 106, 117 closure, 20, 80, 93, 160 – 61; theory of,
object recognition, 10, 16 – 17, 44, 69, 19, 81, 104, 148 – 49, 150, 165, 167, 169,
77 – 78, 146, 154. See also recognition 171. See also ethicopolitics
opacity: of algorithms, 5, 8, 19; Butler’s possibility: of automation, 59; combina-
conceptualization of, 135, 151; critical torial, 14, 163; as computational layer,
theory of, 23, 67, 98, 101, 105, 110, 117, 137; of errors, 47; of future action, 51,
123, 140, 163, 164 – 68, 171 – 72; data and, 79, 97, 119, 147; perception and, 16,
42, 136 46, 143
optimization, 10, 12–13, 17, 43, 54, 60, posthuman, posthumanism, 65, 88, 118,
64, 147 135, 141, 142, 150 – 51. See also human,
output: algorithms and, 4, 17, 18, 47, humans
potentiality, 7 scene analysis, 16, 25, 43, 88, 126, 154, 157
pre-­computation, 77 – 79, 120 science: apparatus of, 29 – 30, 34, 38, 44,
prediction, 90, 106, 138, 160 53; and knowledge, 16, 35, 87, 93, 98,
privacy, 38 112, 133, 137 – 53; and observation, 29.
probability, 14, 48, 68, 70, 71; algorithms See also computer science
and, 74, 90, 92, 99, 105, 137, 162 – 63, skynet, 124 – 27
167; contingency of, 75, 80, 81, 139, scores: algorithm-­generated, 4, 68, 96,
142 – 43, 147 – 48, 175–76n37; Daston’s 126
conceptualization of, 125; and the security, 12, 40, 43, 50; agencies, 33, 37, 217
event, 44, 107; likelihood ratio and, 89; automation and, 140; military, 16,
95 – 96; robotics and, 77 112, 124 – 27; threats to, 54, 70, 74, 148;

Index
profiles, 69 – 70, 76, 95 – 96 and uncertainty, 47, 134
prism, 37 sentiment analysis, 54, 147
propensities, 5, 17, 49, 91, 160, 194n15 sight, 15, 29, 31, 38, 40 – 4 1. See also
protest, 1 – 4 , 10, 17, 46, 55, 91, 169 vision
social media: chatbot and, 108 – 9; data,
race and racism, 1 – 4 , 11, 69, 85, 97, 111, 1 – 2, 37, 42 – 43, 47, 50, 51, 71, 160, 170;
128, 145, 175n37 images, 3, 54; sentiment analysis of,
random forests, 112, 124 – 27, 165, 190n51. 147, 174n9; text, 70, 89, 91
See also decision trees Society for Worldwide Interbank Finan-
reason, 42, 44, 86, 108 – 29 cial Telecommunications (swift),
recognition, 1, 3, 14, 17, 33; of authorship, 36
90; Butler’s conceptualization of, software, 35, 42, 97
19, 145; of handwritten text, 69 – 70, source code, 18, 85 – 88, 91 – 92, 95 – 96,
173n4; of patterns, 46, 150, 181n10; 99, 104, 186n34
politics of, 4, 8, 41, 126, 127, 167, sovereignty, 31, 36 – 38, 53 – 54, 128 – 29
171, 184n55; regimes of, 48 – 49, 54, spatiality, 11, 14, 33, 35, 37, 40, 71,
56 – 81, 121, 122, 183n43; semantic, 89, 162 – 63
185–86n12; speech, 106. See also facial spoofing, 121
recognition; image recognition; ob- statistics, 12, 47, 95, 122, 124, 139
ject recognition Stengers, Isabelle, 14, 53 – 54, 165
reduction: algorithmic processes of, surgery, 20, 62, 64
16 – 17, 115, 128 surveillance, 16 – 17, 38, 39
resistance, 51, 145, 157, 169 – 70 swift, 36
responsibility: for algorithmic harms,
100, 112; and code, 86, 94, 95; ethics targets, 43, 124, 139; generation of, 5, 9,
of, 8, 81, 118, 121, 123, 164 – 65; human 16, 67, 90
touch and, 63 – 64, 144; locus of, 5, technology, 15 – 16, 31, 36, 40, 43, 50, 101,
18, 19, 66, 92, 99; of science, 114, 145
133 – 3 4 terrorism, 33, 55, 74, 89, 91, 124, 126, 138,
rights, 5, 10, 38, 148 148
risk, 4, 19, 40, 104, 126, 138, 140, 145 – 48 text, 93, 97, 101, 104
robots, 20, 58 – 59, 63 – 64, 77, 94 threshold, 41, 48, 54, 64, 68, 105, 139
rules, 7, 14, 47, 53, 65, 69, 89, 141, 150 time, 16 – 17
 traces: data, 44 – 46, 62, 150 tential for, 17, 111; science and, 134;
training datasets, 4, 9, 17 – 18, 20, 68, 96, sovereign, 53, 128, 164
122, 125, 135 – 36, 150, 155; images used the virtual, 35, 62, 177n15, 181–82n23
as, 49; in natural language process- vision, 14, 39, 44, 53, 159, 162. See also
ing, 90; patterns in, 54; and point sight
clouds, 77 visualization, 14, 30, 38
transparency, 5, 8, 18, 86, 110, 117, 123,
134, 166 war. See violence
218 truth: concept of, 5, 9, 14, 18, 93, 122, weights: adjustment of, 24, 48, 68, 73,
135 – 37, 145 – 46, 148; and falsity, 103, 75, 80, 92, 96, 103, 128, 145; and as-
114. See also ground truth signing value, 129; and indeterminacy,
Index

Turing, Alan, 13, 56 – 57 22, 75, 147; machine learning and, 8,
Twitter, 89, 108 – 9 13 – 1 4, 70 – 71, 75, 79, 106, 111, 128; and
mediation, 162; parameters and, 74,
uncertainty, 12 – 13, 47 – 48, 55; machine 136; politics of, 163 – 64; and rejected
learning and, 21, 22, 64, 88, 95, 111, alternatives, 81, 99
118, 136, 139, 142; political, 80; writing Wiener, Norbert, 111 – 15
and, 98 – 99 Wilson, Charles Thomson Rees, 29 – 31, 44
the unknown, 90, 95, 97, 102, 105 – 7, Wittgenstein, Ludwig, 9, 13
152 – 53, 155 world-­making, 12, 20, 30, 33, 40, 43, 48,
US Department of Homeland Security 88, 97, 161
(dhs), 2, 31, 43, 50 writing, 86 – 88, 95 – 98, 101 – 3 , 158. See
also code
variability, 71 – 73, 126, 134
verticality, 17, 40, 46 xenophobia, 119
violence: algorithmic, 4, 11, 46, 66, 118,
126, 146, 152, 161; meaning of, 91; po- YouTube, 109 – 10