The Path Forward: Navigating Machine

Learning Between Promise and Peril

by

Thabisile Msibi u13130138

Submitted in partial fulfillment of the requirements for the degree Honours in

Fil713 Philosophy of mind technology AI, Data, Ethics and Justice

November 2023

I hereby certify that this report is my own work, except where duly acknowledged and that no
plagiarism has been committed in the writing of the report.
Signed: Thabisile Msibi of student number u13130138 on this day 20 November 2023.
 Abstract ....................................................................................................................................................... 2
Introduction.................................................................................................................................................. 2
Harms of Bias in Machine Learning ............................................................................................................ 3
Actions to Mitigate Harms of ML Bias ......................................................................................................... 4
Conclusion................................................................................................................................................... 8
Reference list ............................................................................................................................................ 10

Abstract

This essay argues that machine learning in itself is not harmful and even though machine learning systems can result in
significant harms. Furthermore I argue that banning machine learning technology altogether is unwarranted and would
not only jeopardise the opportunity to make machine learning safe and trust worthy.
Instead I will advocate for disciplined application of bias mitigation and harm reduction techniques like the
implementation of improved data collection practices, algorithmic auditing and the implementation of regulation
through laws and oversight significant efforts can be made to mitigate harms of discrimination, lack of access to
marginalized groups, amplification of misinformation, and erosion of public trust. This paper will also introduce the
concept of reciprocal ethical interaction where through ethical education, Artificial Intelligence (AI) users can work to
reinforce the actions taken by developers to mitigate bias within machine learning data. This paper aims to provide a
balanced perspective on navigating the nuanced challenges of Artificial Intelligence that exhibit promise along with
risk and that machine learning offers humanity a novel opportunity in the day to day practical application of Ethics.

Introduction

Machine Learning (ML) has rapidly emerged as one of the most transformative technologies of our time. By enabling
artificial intelligent (AI) systems to learn from data and improve with experience, ML has been responsible for
breakthroughs across an array of fields including healthcare, the financial systems as well as the criminal defence and
justice systems. However, alongside the facilitation of these major breakthroughs harms caused by ML have also
bubbled to the surface.

Instead, we have the opportunity to respond to the deeper call of ML, the need for ethical development and well
informed governance. ML is asking to be developed within a carefully cultivated intentional ethical environment with
increased transparency, auditing algorithms, mindful regulations and reciprocal ML ethical engagement engaging
stakeholders with interdisciplinary approaches within every step of the design process to mitigate these harms.
ML is the research and development of algorithms that can learn from data and improve their performance without
being explicitly programmed for every task. Through sample data the algorithms build models this training data then
helps it to make predictions and decisions. The key aspect of this process is that ML is trained on massive data sets,
now, if there are biases in the training data and if the algorithm is not deigned properly, the resulting models can inherit
and amplify prejudices. For instance, a facial recognition system using dataset without racial diversity will not prove
efficient when applied, to minorities. Text generation models can still produce hate speech after being trained with
such stereotype-filled data. Machine learning systems are not inherently biased, but biases are created by the data used
to train them.

In response to these harms’ options have been proposed on how to correct and prevent these harms, one of these being
an outright bans on ML research and all its applications. Proponents of the outright ban state that in order to eliminate
these harms prohibition of development is the best option. In this essay, I will argue that while ML in its current state
carries significant risks in harms and of perpetuating discrimination, prejudice, and unethical outcomes that deeply
impact millions of people, banning this technology altogether would be an unwarranted response, that would not only
mean the forfeiting of substantial benefits and the possibility of driving the development underground eliminating any
potential of achieving what the ban aims to do which is curb the risk and eliminate the harms as much as possible.

The core issue that Ml faces now is dealing with the inherent biases embedded within the algorithm and ML training
data sets that causes these harms. For example biased algorithms can lead to discriminatory outcomes, perpetuate
existing social economic and possibly amplify these existing inequalities through the exclusion from access to
resources and opportunities. If the repercussions of these biases were not enough these harms go onto erode trust in
institutions and further perpetuating the age old human fear ‘of the possibility of autonomous intelligent machines
stems from the behaviour of these agents and if their choices will be ethical or not’ (Anderson and Anderson, 2007,
p16). In this work I will introduce the idea of reciprocal ethical interaction. The Algorithm will itself be sensitised to
its potential for bias and the harms that occur as a result. Given that we are strictly dealing with the intelligent
autonomous agents and the user too will be sensitised to the algorithms potential for bias and as the user their own
potential for bias and the resulting harms.

We must respond to this challenge with nuance, recognizing machine learning as a complex sociotechnical system that
demands measured, pragmatic governance rather than reactionary prohibitions.
With diligent work, the benefits of the array of problems that ML can help us solve will far outweigh the risks.and
harms caused by bias. Measured governance, not reactionary prohibition, is the answer to managing the complexities
that arise and will most likely continue to arise through the development and deployment of this technology.

Harms of Bias in Machine Learning

As mentioned above the training data and the algorithm merely surfaces and magnifies existing societal biases
prejudice and imbalances. Our world is still, after a long history of discrimination, recovering and recuperating and the
attitudes of this history, unfortunately, ‘live on in our digital systems for complex reasons and they become buried into
the logics of our ML infrastructure’ (YT VID). One major source of this bias is a consequence of poor model design
choices made by developers and their assumption. An absence of diversity among teams building ML systems leads to
an absence of representation within perspectives and therefore produces blindspots. It is therefore necessary to
implement proactive bias detection checks, without this these issues accumulate and cascade through the machine
learning development lifecycle leading to a biased system. It is believed that most of the problems of fairness in
machine learning stem out from the data. Models learned from the historical data may have encoded some undesired,
existing social biases, which could perpetuate the preexisting inequality. Some logics of prediction are deemed
improper by society, despite being statistically valid, and, as a result, machine-learning algorithms should discriminate
among data points. A machine learning model may also be considered fair during its development, yet later prove to be
unfair when it is applied into another situation or used with other data. Machine learning models’ outputs may be used
in discriminatory ways like employment and loan approvals. Generally, the unfair sources of machine learning are
complicated and multifaceted and therefore, it requires profound insight into the social, economic and political
circumstances where machine learning is being utilized.
There are two major harms that are found in ML: allocation and representation harms. Allocation harms refer
predominantly to economic resources and occurs when prejudice leads to unfair distribution of resources. In the
financial sector for example, credit approval systems have been found to be biased, denying loans to minority groups
and previously marginalised groups. Representation harms occur when systems reinforce harmful stereotypes,
denigration harms, recognition harms and the harm of underrepresentation. For instance there was an incident
involving Google Photos where the software incorrectly labeled an African-American woman as a gorilla. Harmful
denigration can be seen in this incident. These harms, in propagating injustice, end up diminishing trust in ML
systems.

While the harms and the resulting discrimination caused by ML raise serious concerns, The following section will be
dedicated to considering steps that can be taken in addressing these issues. This will include auditing

Actions to Mitigate Harms of ML Bias

To promote fairness and remove biases in machine learning-based systems, active actions need to be applied over the
whole process of collecting data that is big. The steps will serve as safety mechanism for avoiding biased and
discriminatory tendencies that are reflected in these systems. Key steps in this pro-active approach help identify,
understand and cut these discriminatory biases. For example, data distribution analysis must be ongoing so as to detect
and correct any bias evolving gradually. It is a holistic approach towards data cleaning, audit, and improvement whose
vital contribution towards the modeling fairness can not be downplayed. Improving training data collection and inputs.
Taylor in the article titled ‘What is data justice? The case for connecting digital rights and freedoms globally’
Considers multiple perspective on how to implement data justice and in an effort to reconcile these differing
perspectives offers a single conceptual framework that illustrates a tripillar data justice methodological system. Data
Justice as defined by Taylor is a set of principles and practices that aim to prevent information-based discrimination
and promote a socially just and equal use of data (Taylor, 2017), and is important because with the increasing
availability of digital data has concurrent political and practical implications for how people are treated, especially by
the state and the private sector. In Taylor’s view, data justice is a crucial concept that can balance the differing
perspectives on data use and ensure that the benefits of new data technologies are distributed fairly and equally. With
the primary focus of data justice as marginalisation and promoting a socially just model for handling data, the
approach begins not from consideration of the average person, but instead asks what kind of organising principles for
justice can address the marginalised and vulnerable to the same extent as everyone else. (Taylor, 2017)
The first pillar of this framework is that it needs to take into account the novelty and the complexity of the ways in
which novel and complex big data systems that can discriminate, discipline and control (e.g, Aadhaar and migration
monitoring systems). Critics of the Aadhaar system suggest this could be used to discriminate against the poor and
disenfranchised without any means of appeal or independent review. The migration monitoring system is a data system
of the European Union to monitor and manage migration fluxes that has been questioned over the violation rights and
worsening existing inequalities.
The second pillar is, this framework should take in account both the positive and negative potential of the data
technologies and their ability to facilitate what has been termed ‘human flourishing’, which these systems have the
potential to hinder it.
Thirdly this framework should use principles that are useful across social contexts, and thus remedy the developing
double standard with regard to privacy and the value of visibility in lower versus higher income countries, as
illustrated by Global Pulse’s utilitarian balancing of others needs as well as privacy. Overall, Taylor’s proposed
framework for data justice aims to promote socially just and equal use of data, prevention information-based
discrimination, and make sure the distributed fairly and equally.
The training data is one of the primary sources of bias with the ML systems. The data can underrepresent particular
demographics and marginalised perspectives. Sterotypes can be encoded into the data along with assumptions, and in
some cases certain outright errors (Crawford, 2018). The key to solving this problem is in theory simple; thoughtfully
collecting diverse, representative datasets. The data should cover a wide range of populations and scenarios to avoid
skewing systems. Michael Veale and Reuben Binns in their paper ‘Fairer machine learning in the real world:
Mitigating discrimination without collecting sensitive data’ offer a trusted third party as a source that can help
organisations address knowledge and information deficits related to fairness & bias in the training data sets. These
third parties fulfil the function of data auditors and can function as well as algorithm auditors.
Ongoing analysis of the data that is sourced and how it is edited is needed to catch imbalances, a data curation system
that incorporates auditing and augmentation. This will all serve to enhance model fairness.

In addition to the training data the algorithms themselves require regulation and maintain a keen eye on the algorithm
in order to maintain fairness and mitigate discrimination.Formal auditing of the algorithm involves crash testing the
algorithm across various groups, testing for discrimination. Additionally Vale & Binns mention fairness-aware
algorithm design. This approach, explicitly designs algorithms that have an aversion to bias. The technique this
includes is ‘counterfactual fairness’, this technique involves the modification of the training data to ensure that the
model would have made the same decision even if certain characteristic's such as gender and race, sensitive data,
would have to be protected or even different. The chances of ML harms reaching the public can be significantly
reduced through promoting algorithmic accountability.
Accountable algorithms are a way to make algorithms more transparent and accountable, therefore preventing
discrimination and promoting fairness. Because algorithms can identify patterns and make predictions they can
actually serve to address societal problems and promote social justice (Taylor 2017) but these algorithms need to be
accountable ignorer to serve this purpose. Kroll et a;. (2016) described “Accountable Algorithms” as a technological
toolkit to promote legal fairness in automated technological systems to promote fairness and mitigate discrimination.
These tools include algorithm checking mechanisms that ensure fairness in selections if it is established that the
algorithms were used to arrive at decisions. The article suggests that by providing accountability, transparency, and
fairness, we may build “accountable algorithms.”

The next stage of constructing an airtight harm mitigation procedure is the paring of these technical solutions with
governance. Laws and regulations such as Anti-discrimination laws targeted to specifically cover algorithmic decisions
sets a serious baseline for accountability. High-risk applications of ML systems are found in the financial sector,
healthcare and the criminal justice system. In these specific fields added oversight can be implemented through
regulatory agencies, independently monitoring AI in the public further serves to enforce transparency and auditing of
these systems. To ensure consideration of social implications and harms before deployment, companies also need to
adopt rigorous internal policies, review processes, and codes of ethics.
Equal protection is a constitutional principle. Under this principle government is required to treat all individuals
equally under law. Algorithms therefore need to follow this key standards of legal fairness. However Kroll et al (2026)
highlights that there is tension between equal protection, disparate treatment and disparate impact within anti-
discriminatory law. Kroll et al (2016) invites us to consider that algorithm's can help us to address this. Through
ensuring the algorithms are transparent , fair and legally defensible these tools can help us to alleviate the tension
found in anti-discrimantory law. While laws and regulations concerning discrimination and fairness in machine
learning may differ depending on location, it bears mentioning that standards can change between nations and areas
with authority. Given the legal and contextual nuances involved with machine learning applications, consulting the
applicable laws and rules governing its specific use context is paramount.

Vardi in the article titled 'ACM, Ethics, and Corporate Behaviour, highlights the need for both difficult and nuanced
conversations on responsible computing, ethics, corporate behaviour, looks like. "A new Institute for Ethics in
Artificial Intelligence at the Technical University of Munich, since “ensuring the responsible and thought- ful use of AI
is foundational to everything we do.” (ACM P1) There is awareness of consistent commitment to fairness,
accountability and transparency within the ML community and cognisance of how this creates a culture that promotes
ethical AI development.

In spite of the challenges, a concerted effort can significantly reduce the harms of ML bias in order to unlock more of
AI's potential while ensuring society's safety. One of these steps described by Crawford & Call as a Sociotechnical
approach for developing and deploying AI systems. Crawford describes that it is critical to engage in social-systems
analysis on AI that examines the history of the societies/political background of data used as basis for AI systems; and
consultation of community input on weighing police data against feedback about neighbourhood policing. She
includes that determining if the system should be created and that the risk and rewards of the framework equally.
However Crawford& Call contends that a more rigorous and socially responsive approach to AI system development
and implementation is required to counter the possibility of bias and inequity being entrenched or magnified through
such systems.(Crawford and Calo, 2016)

Beyond technical fixes and governance, improving the ethics of machine learning systems involves enhancing the
reciprocal ethical dynamics between humans and AI. Both sides can promote moral behaviours in each other through
ongoing education and aligned incentives. As AI continues to develop and expand developers have the responsibility
to impart values into the systems through the algorithms, and they also have the opportunity to install machine
morality simultaneously, interactions with AI agents that are moral could serve to make humans more. Ethical by
exposing people to alternate perspectives as well as new situations for moral growth. Philosophers like Peter Singer
have argued that humans struggle with expanding compassion to distant strangers, whereas AI systems could help us
to confront this difficulty within ourselves through AI algorithms and technology making it salient and thereby forcing
us to stand and face the inherent bias and discrimination within our society. Autonomous intelligent agents as moral
machines may prove essential to developing human ethics at a societal level even as we work to embed ethics within
them. Users of AI, not only the tech companies, developers or engineers, share in the responsibility to encourage
ethical practices in these systems. As developers of this technology, however, it begins with the tech companies and
developers. They can begin the ethical practice through implementing incentives, oversight mechanisms, and
regulations that discourage harmful behaviour and outcomes is key. Through recognising the relationship between
humans and machines and highlighting the reciprocal nature of this interaction can serve to insulate the ML systems
against bias and discrimination. Navigating the complex ethics of emerging technologies like AI demands recognizing
the reciprocal relationships between humans and machines. With thoughtful coordination, we can educate each other
on how to bring more justice and dignity into systems increasingly entwined with human society. This reciprocal
perspectives provides hope reciprocal perspective provides hope that human ethics and artificial intelligence can
progress together in a responsible manner.

In the final part of her keynote address Crawford asks where do representational harms actually come from.
Representation serves to communicate group identity and human culture as a whole. However in terms of the actual
categories Crawford asks what if bias is actually a deeper and more systemic issue than we realise, what if it is built
into classification and those that are classified bear the implications of the bias. Te’eni et al., 2023, in his article
‘Reciprocal Human-Machine Learning: A Theory and an Instantiation for the Case of Message Classification’ proposes
Reciprocal Human-Machine Learning Reciprocal human-ML (RHML) configuration for text message classification. A
design science approach that makes it possible for people and machines repeatedly exchanging feedback on a
classification task while refining knowledge representations. In message classification, the theory of Reciprocal
Human-Machine Learning suggests using a configuration which encompasses a reciprocal learning cycle involving
repeated exchanges of feedback about a classification task between humans and a machine learning based model
(RHML), with each party iteratively updating its knowledge(Te’eni et al., 2023).

In experimenting with the idea of an outright ban on machine learning research and development would create more
challenges than solutions. There are several drawbacks to blanket prohibitions. First, blanket bans on machine learning
technology would prove exceedingly difficult to implement and enforce in practice. Second, bans could drive research
and development underground, outside to public oversight. In the same way that the prohibitions on substances created
black markets, bans on machine learning would most likely create black markets, bans on machine learning drive the
activities underground and out of sight. This would then have more sever repercussions as unethical actors operate ML
technologies unchecked. Bringing progress to light allows opportunity for supervision and for accountability
prohibitions would lose.
Legal guardrails provide grounds for public challenges to harmful practices. Government, independent agencies and
nonprofit monitoring of high-risk use cases enable oversight of ML’s rapid evolution.

Finally the imposition of a ban would amount to forfeiting machine learning’s benefits before allowing earnest efforts
at and into ethical development. Stifling innovation prevents the potential for the development of explicit ethical
education as an effective methods to address valid concerns. Without thoughtful coordination, ML will loose the
opportunity to be shaped to promote human dignity human connection and communication over time as other
disruptive technologies like the cellphone or computer have. Banning its progress, evaluates the harms as outweighing
the benefits, however what I have aimed to show thus far is that the benefits actually outweigh the harms and banning
would only exacerbate these harms. More measured governance and incentivising ethical development at every stage
of the ML lifecycle development will enable the transformative potential of this technology to assist us with the
problems plaguing our world today.

Conclusion

In exploring the question of what kinds of harm arise from bias in machine learning processes, this essay has delved
into the substantial risks associated with biases and the lack of transparency, which can result in allocation harms and
representation harms. Despite these challenges, advocating for an outright ban on machine learning technology is
deemed an unwarranted response. Instead, a nuanced understanding of the potential benefits, if developed and
governed responsibly, reveals the tremendous promise that machine learning holds. The neutrality inherent in Machine
Learning, is neither intrinsically beneficial nor harmful. What is of significance is that the impact largely depends on
how humans choose to use it. While current issues uncovered through training data, biases within the algorithm, raise
alarms about fairness, accountability, and most of all human dignity, it is essential to recognise that the field is still in
its early stages. The challenges faced by the machine learning community from the creators and societal shortcomings
integrated into these emerging technologies.

A counterargument presented in Kroll et al’s Accountable Algorithms expresses skepticism regarding the effectiveness
of transparency and accountability measures in addressing the complex issues arising from algorithmic decision-
making. Critics argue that these measures may be insufficient, potentially leading to a false sense of security or
providing opportunities for gaming the system. Despite this, the essay suggests that these concerns can be tackled
through a combination of technical and legal measures, incorporating fairness constraints into algorithm design and
establishing legal frameworks for algorithmic accountability.
Another counterargument acknowledges that AI technology itself is not inherently unethical but is often used by large
corporations to support arguably unethical business models. The tension between an unethical business model and a
façade of ethical behavior within some companies is complex, emphasizing the need for nuanced conversations about
responsible computing, ethics, corporate behavior, and professional responsibility.While acknowledging the validity of
this argument, Moshe Y. Vardi, emphasizes that the problem is more complicated. And these companies struggle to
maintain an ethical façade despite their unethical business models.Vardi argues that the computing community needs
to have difficult and nuanced discussions about ethics, corporate behavior, and professional responsibility, while it
may be true that AI technology is not by nature unethical, it is important to hold the computing community
accountable for their actions and decisions.

A comprehensive understanding of the challenges and counterarguments surrounding bias in machine learning
processes highlights the need for responsible development, transparency, and ongoing discussions within the
computing community. While acknowledging the potential harms, an outright ban on machine learning I highly
discourage, as it could hinder humanity from responsibly advancing technology and shaping its trajectory toward
empowerment rather than injustice. Ultimately, the task is to imbue ethical reasoning into the foundations of these
technologies, addressing biases, and ensuring that machine learning serves as a tool for positive societal change.
Reference list
Anderson, M. and Anderson, S.L. (2007). Machine Ethics: Creating an Ethical Intelligent Agent. American
Association for Artificial Intelligence, 28(4), pp.15–26.

Crawford, K. (2018). The Trouble with Bias NIPS 2017 Keynote Kate Crawford #NIPS2017. [online]
www.youtube.com. Available at: https://www.youtube.com/watch?v=-L-I-mN_C5U&list=PPSV [Accessed
20 Nov. 2023].

Crawford, K. and Calo, R. (2016). There is a blind spot in AI research. Nature, 538(7625), pp.311–313.
doi:https://doi.org/10.1038/538311a.

Kroll, J.A., Huey, J., Barocas, S., Felten, E.W., Reidenberg, J.R., Robinson, D.G. and Yu, H. (2017).
Accountable algorithms. UNIVERSITY OF PENNSYLVANIA LAW REVIEW, [online] 165, p.633705.
Available at: https://univofpretoria.on.worldcat.org/oclc/8276543270.

Taylor, L. (2017). What Is Data justice? the Case for Connecting Digital Rights and Freedoms Globally. Big
Data & Society, 4(2), pp.1–14. doi:https://doi.org/10.1177/2053951717736335.

Te’eni, D., Yahav, I., Zagalsky, A., Schwartz, D., Silverman, G., Cohen, D., Mann, Y. and Lewinsky, D.
(2023). Reciprocal HumanMachine Learning: A Theory and an Instantiation for the Case of Message
Classification. Management Science. [online] doi:https://doi.org/10.1287/mnsc.2022.03518.

Vardi, M.Y. (2022). ACM, ethics, and corporate behavior. Communications of the ACM, [online] 65(3), p.5.
doi:https://doi.org/10.1145/3516423.

Veale, M. and Binns, R. (2017). Fairer machine learning in the real world: Mitigating discrimination without
collecting sensitive data. Big Data & Society, [online] 4(2). doi:https://doi.org/10.1177/2053951717743530.