2022 Ai Report Card
2022 Ai Report Card
EXECUTIVE SUMMARY
In the United States, there are three sets of institutions that are primarily
responsible for initiating, importing, modifying, and diffusing AI: private firms,
publicly funded national laboratories, and universities with funding from
government, industry, and donors. AI policy refers to the public means for
nurturing the capabilities and activities of this AI ecosystem and optimizing its
applications in the service of national goals and the public good. When
government institutions and policies act properly, AI innovation flourishes.
When government fails to act or misfires, so too does AI innovation. 2
This report analyzes how the United States is performing across nine of the
most prominent policies the U.S. government uses to support AI innovation
and competitiveness. We split these policies into two groups. First are
innovation policies that directly spur AI innovation and competitiveness.
These include six types of policies that support AI research, strengthen the AI
workforce, spread AI tech hubs across the country, facilitate access to AI
resources, promote government adoption of AI, and help develop technical
standards for AI. Second are legal and regulatory policies that shape the
environment for AI innovation. These include three types of policies that
regulate the use of AI systems, incentivize AI activity through intellectual
property (IP) rights, and support AI development through international trade.
The following report card summarizes our findings for each policy area and
provides an achievement level. To be failing expectations in a policy area
means the policies in place—or lack thereof—are causing the United States to
fall behind its competitors. To be approaching expectations means the
policies in place are only partially or inconsistently bolstering U.S. AI
innovation and competitiveness. To be meeting expectations means the
The Trump administration updated the plan in 2019 with an eighth strategy
to expand public-private partnerships to accelerate advances in AI and
imbued the existing plan with a stronger focus on maintaining U.S.
leadership. 6
Treasury 0.5
0.8
NOAA 2.2
2.2
NIOSH 3.8
3.8
NASA 6.6
6.5
DOT 11.2
11.3
NIJ 11.5
11.5
DOI 8.5
13
VA 18.4
24.7
NIST 31.8
46.9
FDA 46
50
DHS 56.4
77.8
USDA 161.2
161.2
DOE 278.8
231.7
NIH 337.9
352.2
NSF 639
744.7
Unfortunately, tax incentives in the United States for R&D are quite minimal.
The country ranks 32nd out of 34 comparable Organization for Economic
Cooperation and Development (OECD) and BRIC (Brazil, Russia, India, and
China) nations, having slipped from 24th place in 2020. 16 As of 2022, a
provision in the 2017 Tax Cuts and Jobs Act (TCJA) no longer allows
companies to expense current R&D costs in the first year (to deduct the costs
of R&D from their taxable income in the year they incur those costs) and
instead requires costs to be amortized over a period of five years, effectively
reducing the R&D subsidy by about 5 percentage points, from around 9
percent to 4 percent. 17
Fortunately, the Biden administration has indicated support for enhancing tax
incentives for R&D. In response to a question from Senator Todd Young (R-IN)
about maintaining the immediate deductibility of R&D expenses, U.S.
Treasury Secretary Janet Yellen said:
Many innovation and competitiveness scholars point out that America’s most
innovative firms frequently cluster together in small, relatively well-defined
regions. 21 Innovation clusters offer several benefits for the AI industry. For
one, in the early stages of technology development, clusters tend to produce
and exchange industrial “gossip” about production processes that helps
The problem is, while the clustering of AI skills and firms promotes innovation
and growth within regions, hyperconcentration of the associated economic
gains may further entrench the already imbalanced geography of the nation’s
economy. As ITIF explained in The Case for Growth Centers, the innovation
sector has generated significant technology gains and wealth but has also
helped spawn a growing gap between the nation’s dynamic “superstar”
metropolitan areas and most everywhere else. 23 Among the superstar metro
areas, the “winner take most” dynamics of the innovation economy have led
to dominance, but also to livability and competitiveness crises: spiraling real
estate costs, traffic gridlock, and increasingly uncompetitive wage and salary
costs. Meanwhile, in many of the “left-behind places,” the struggle to keep up
has brought stagnation and frustration. These uneven realities represent a
serious productivity, competitiveness, and equity problem.
AI EDUCATION
On one hand, this decentralized approach allows for greater creativity and
innovation in how schools develop and implement AI curricula, potentially
enabling an increase in the quality of education. But divergent approaches
can exacerbate disparities in how rigorous curricula are and the qualifications
of educators. 30 Indeed, integration of AI curricula in the United States is
already uneven in both depth and scope. Many schools do not teach CS,
which is seen as the first step in AI specialization. 31 Only 51 percent of U.S.
high schools offer foundational CS and only 23 of the 50 states and District
of Columbia require all high schools to offer CS. 32 The few schools that offer
explicit AI courses also vary in the content and scope of their curricula.
Consider the two U.S. public high schools with the most prominent AI
curriculums: The North Carolina School of Science and Mathematics, a
public, two-year boarding school that is administered by North Carolina’s
university system, received a $2 million gift to launch an AI program in 2018
to teach students how to use and create AI systems, with a strong focus on
understanding the ethical implications associated with these technologies. 33
Seckinger High School, established with $79 million in funding by the
Gwinnett County Public Schools district in Georgia, is incorporating AI courses
in core subjects for its entire K-12 cohort. 34 Its program introduces
elementary school students to block coding—a basic form of computer
programming that uses visual instruction blocks to construct games—while
middle and high school students can learn programming languages and apply
robotics and sensors to real-world applications. 35
While the U.S. school system has not fully responded to the increased
importance of AI skills, more employers, parents, and even students
recognize the benefits of learning AI. Several nonprofits and advocacy groups,
learning-programs, and courses have sprung up in response. Nonprofits such
as Girls in Artificial Intelligence and Technology Education (GAITway), AI4ALL,
Black Girls Code, and Girls who Code seek to increase access to AI education
across gender lines and socioeconomic divides, introduce AI to students at a
younger age, train more teachers, and put AI into more schools. The private
sector is also reinforcing AI integration in schools through a number of
different initiatives ranging from after-school programs to hackathons to
summer camps. A 2021 report finds that for-profit companies are responsible
for hosting 59 percent of the close to 450 AI and AI-related summer camps
that exist across the United States. 39 For example, iD Tech Camps is a
summer computer camp held at more than 150 U.S. college campuses that
offers AI and machine learning courses for 13 to 17 year olds. 40 Some of
these programs target populations typically underrepresented in AI and CS,
such as Microsoft’s DigiGirlz High Tech Camps, which are multi-day tech
programs for girls in middle and high school. 41 Other private sector initiatives
support teachers in-classroom learning by providing resources that aid
learning, such as Google’s CS First curriculum, a virtual free-to-use CS
curriculum for students ages 9 to 14. 42
Higher AI Education
Unlike U.S. high schools, where AI and CS education is deemed subpar, some
U.S. institutions of higher education boast strong AI programs, drawing
students from around the world. Moreover, interest in studying AI and AI-
related courses at U.S. universities is surging as the market for these skills
soars. 43 But U.S. institutions are frequently unable to meet demand because
too few universities are willing or able to sufficiently persuade the qualified
professors they need to increase their faculties and offer more classes to
students to pick working in academia over lucrative private sector jobs. 44
Moreover, many colleges and universities respond inadequately to
“customer” demand because they are unwilling to reallocate resources for
less in-demand academic programs to make budget space for more in-
demand ones such as CS.
The private sector has been playing a role in supporting tertiary AI education
just as it does in primary and secondary education, providing everything from
certificate and online learning programs to subsidized access to AI systems
for teaching. 53 Perhaps no public-private AI partnership is as comprehensive
as that of the University of Florida (UF) and Nvidia, a U.S. company that
develops graphics chips. Anchored by a $25 million gift from UF alumnus and
founder of Nvidia Chris Malachowsky, and an additional $25 million in
hardware, software, training, and services from Nvidia, UF has launched an
initiative to become an “AI University.” 54 As part of this initiative, UF has
incorporated AI into all undergraduate majors and graduate programs,
WORKFORCE TRAINING
Providing AI education is important, but national policy needs to also help
incumbent workers. One problem impeding successful workforce policies is
that while schools are primarily responsibility for equipping the future AI
workforce with the requisite skills and knowledge they will need to succeed in
an AI economy, there is little agreement about the respective responsibilities
of individual workers, employers, and government in training the existing
workforce.
It is unlikely that the rate of firms choosing to train employees will reach one
that is optimal from a societal and economic perspective without government
intervention. Corporate investment in workforce training in general has been
on a downward trend as more and more firms seek to simply hire workers
with the requisite skills instead of paying to train them. 61 After all, why would
a company want to invest in training workers in costly AI skills when so many
are leaving their jobs in record numbers? 62 The United States can work to
reverse this trend by creating incentives for firms. One option is to allow
qualified expenditures on workforce training to be taken as a knowledge tax
credit. To ensure companies use this credit to focus on the skills of most of
their workers, and not just managers, firms taking advantage of the credit
could be required to abide by rules such as those for pension program
distribution, which limit focus on highly compensated employees. 63
Fortunately, the government has already begun to recognize the need for
coordinated AI workforce policies at the national level. As part of the National
Defense Authorization Act of 2021, Congress charged NAIIO with developing
a strategic plan that establishes goals, priorities, and metrics to “support and
coordinate federal education and workforce training related to artificial
intelligence.” 65 NAIIO’s interagency committee for education and workforce
training will create AI workforce goals and priorities at the federal level.
OSTP and NSF should prioritize the development of tools and metrics
to quantify the AI computing needs and resources of the academic
community. There is little literature on what level and type of compute
AI researchers need. 100 Without this information, policymakers
cannot effectively make decisions about what resources to invest in
or how much.
Congress and the White House have taken important steps recently to
facilitate greater government adoption of AI, but many agencies still face
unique challenges to becoming more AI-mature organizations. Indeed,
deployment of AI in the federal government is relatively low despite 70
percent of public sector IT leaders agreeing that AI is “mission critical.” 102 The
challenges facing agencies include outdated IT infrastructures, limited
funding for capital expenditures, lack of awareness about the technology, and
risk aversion, among others.
Each federal agency should develop its own AI strategy and appoint a
chief AI officer. One reason certain agencies devote so little attention
to AI is that it is generally not formally recognized as part of agency
agendas or strategic plans. Each agency should explicitly identify
specific steps for how it will connect its data, users, and mission
priorities to support AI transformation, much like the Department of
Defense, Department of Veteran Affairs, and Food and Drug
Administration already have. To really coordinate and drive
implementation of AI, each federal agency should consider appointing
a chief AI officer, like the Department of Health and Human Services
has done.
GAO and Council of the Inspectors General Should Call Out Agencies
for Not Innovating with AI. Rather than looking at waste, fraud, and
abuse alone, these organizations should look at waste and inertia
from lack of AI innovation. The federal government has been aware of
the need to digitize paper form processing and automate manual
processes since the early 2000s, and slow adoption of AI has cost
taxpayers a lot of money. It should therefore hold federal agencies
accountable for not innovating. 109
The U.S. approach to standards development for AI follows the general U.S.
standards system, which has been exceptionally successful in generating
technological innovation in the United States. The U.S. standards system
focuses on voluntary consensus standards that are created by private sector
standards development organizations in response to particular needs or
issues identified by industry stakeholders, government, or consumers. For
instance, the Society of Automotive Engineers has developed definitions and
specifications of autonomy that autonomous vehicle manufactures rely on to
adhere to rules from the Department of Transportation. And the U.S. tech
association brought together 50 technology and health organizations in 2020
to establish the first ever standard for AI in health care accredited by the
American National Standards Institute (ANSI), building consensus on such
terminology as “telehealth” and “remote patient monitoring.” 112
There are two categories that broadly describe how standards achieve
adoption in America. De facto standards achieve adoption through
competition among rival standards consortia. Consider the Open Neural
Network Exchange (ONNX) and the Neural Network Exchange Format (NNEF),
two examples of open data exchange protocols developed by private-sector
consortia to enable interoperability between different frameworks for training,
executing, and deploying machine learning models. 113 In the de facto
method, the market informally decides which protocol achieves the dominant
The role the federal government has played in standards has historically
been limited, primarily focusing on orchestrating and supporting industry-led
efforts through technical assistance with reference materials, data, and
instrumentation. In February 2019, the Trump administration called on NIST
to take a more engaged role in developing AI standards, issuing an executive
order that, among other things, directed the agency to create “a plan for
Federal engagement in the development of technical standards and related
tools in support of reliable, robust, and trustworthy systems that use AI
technologies.” 115 The plan, which was released in August 2019, helped to
establish a federal AI standards coordinator within NIST charged with
gathering and sharing AI standards-related needs and best practices, and to
promote research that underlies technically sound standards for trustworthy
AI. Regarding the latter, NIST drafted “A Taxonomy and Terminology of
Adversarial Machine Learning” in 2019, “Four Principles of Explainable
Artificial Intelligence” in 2020, and “A Proposal for Identifying and Managing
Bias in Artificial Intelligence” in 2021. 116 In response to a directive from
Congress in the 2021 omnibus spending bill, NIST is also developing a
voluntary framework to manage the risks to individuals, organizations, and
society from AI systems. 117
DOC, NIST, ANSI, and other agencies are rightly involved in developing
international standards for AI—the United States plays a leading role in the
international standards committee responsible for developing AI standards
(ISO/IEC JTC 1/SC 42)—but U.S. policymakers should more actively counter
The United States should use the U.S.-EU Trade and Technology
Council (TTC) discussions to counter EU proposals to pursue regional
AI standards. Because the EU’s AI Act creates legal loopholes for the
bloc to create and apply region-specific technical specifications for AI
where it is deemed that relevant standards are insufficient, the
United States should use the TTC working group on tech standards to
establish commitments on AI standards that ensure those that are
developed are based on industry-driven, consensus-built standards.
In general, the U.S. federal government, more so than any other government,
has adhered to the innovation principle in its early regulation of AI, which
holds that because the overwhelming majority of AI innovations benefit
society and pose modest and not irreversible risks, government’s role should
be to pave the way for widespread innovation while building guardrails, where
necessary, to limit harms. 131 This approach recognizes that market forces,
tort law, existing laws and regulations, or light-touch targeted interventions
can usually manage the risks new AI technologies pose.132
Under the Trump administration in early 2020, the White House unveiled a
set of principles for AI regulation as a follow-up to a 2019 executive order
titled “Maintaining American Leadership in Artificial Intelligence,” which
outlined a series of steps for the federal government to ensure the United
States remains at the forefront of technological innovation. The 10 principles
outlined what federal agencies should take into consideration when crafting
their approaches to AI:
President Biden has indicated a shift from this approach, supporting stronger
regulations. In April 2021, the FTC published a widely noted blog post on how
companies can use AI “truthfully, fairly, and equitably” and shortly after
began a rule-making process “to curb lax security practices, limit privacy
abuses, and ensure that algorithmic decision-making does not result in
unlawful discrimination.” 136 Further signaling its shift toward a greater focus
on issues of AI harm, the FTC appointed Meredith Whittaker, cofounder of the
AI Now Institute, who has written that “the vast majority of AI systems and
related technologies are being put in place with minimal oversight, few
accountability mechanisms, and little information about their broader
implications,” to serve as a senior advisor on AI to FTC Chair Lina Khan. 137
Moreover, the president’s science advisor and director of OSTP began
developing an AI Bill of Rights in October 2021 based on the premise that
current AI and biometric technologies have led to serious problems regarding
discrimination and bias. 138
IP rights have long been recognized as fostering innovation. The idea is that
those who combine the spark of imagination with the grit and determination
to see their vision become reality in books, technology, medicines, designs,
sculpture, services, and more deserve the opportunity to reap the benefits of
their innovation—and that these rewards incentivize more creative output. 142
While the United States is a global leader in IP rights protection, certain
developments—both domestically and internationally—are creating some
uncertainty for global entities. 143 In particular, the advent of AI raises the
prospect that some works are now the direct output of computer systems,
including some operating autonomously. Jurisdictions around the world are
divided on how to handle the “AI inventor” while continuing to enable
innovation. 144
That does not mean the overall IP system does not need reform in light of AI.
For one, offices and courts are facing challenges deciding which AI-based
inventions are patent eligible under the law, which becomes more
problematic as the volume and share of AI patents in the United States
increases rapidly (figure 5).
Patent eligibility in the United States is based on section 101 of the Patent
Act, which says that an invention is new, has some practical use, and is non-
obvious. However, the Supreme Court through some landmark cases has
identified three categories of work that are judicial exceptions, meaning they
are ineligible from this broad conception of eligible subject matter. These are
laws of nature, natural phenomena, and abstract ideas. Because AI patents
generally rely strongly on mathematical relationships and algorithms, they
may be considered abstract ideas under patent law. Patent examiners, who
must determine whether an AI invention is patentable, are hindered by such
uncertainties and in turn have rejected a significant portion of AI patents that
should have received protections. To address this issue, the USPTO issued
guidance in 2019 that clarifies what would be considered ineligible concepts
and provides examples to guide the examination process. According to
former USPTO director Andrei Iancu, the guidance has cut rejection rates for
AI from 60 percent to 32 percent.
Still, because of the unpredictable and uncertain nature of the U.S. patent
system, many AI innovators turn to trade secrets to protect their work. 148
Trade secrets have a number of advantages over other IP. The information
protected by trade secrets does not need to be novel, is protectable
immediately without the cost or lengthy registration timelines other IPRs
require, and is protected for as long as the information is commercially
valuable and can be maintained as secret. 149 However, the key distinction
between trade secrets and other forms of IP that is relevant to AI innovation
is the trade secrets work by protecting information that is undisclosed as
opposed to patents, for example, where the IP is publicly disclosed. According
to NSCAI, because trade secrets do not contribute to accessible technical
knowledge in the public domain, they may hinder AI innovation in the long
Congress and the White House should work with the USPTO,
Copyright Office, State Department, and any other relevant agencies
to craft a national strategy for AI that cultivates strong IP rights.
The United States has had mixed success in protecting cross-border data
flows in past trade agreements. The Trans-Pacific Partnership (TPP), now the
Comprehensive and Progressive Agreement for Trans-Pacific Partnership
(CPTPP), was the first international trade agreement with explicit language
governing the flow of data across borders. The CPTPP includes prohibitions
against localization requirements that would force businesses to build data
In the same month USMCA came into force, the European Court of Justice
made a decision to invalidate the EU-U.S. Privacy Shield, which thousands of
organizations relied on to legally transfer data abroad for operations,
customer service, communications, R&D, and human resources. 158 While
both the EU and the United States have agreed on legal tools to establish
transatlantic data flows in the past—initially the U.S.-EU Safe Harbor in 2000,
and more recently the EU-U.S. Privacy Shield—EU courts have undermined
these efforts twice with the Schrems I and Schrems II rulings. If policymakers
do not create an alternative to the EU-U.S. Privacy Shield, firms from a broad
range of sectors on both sides of the Atlantic will suffer. 159
AI CHIPS
Trade disputes can put a nation’s ability to secure semiconductors, including
AI chips, at risk. Having access to state-of-the-art AI chips is important to
ensure AI developers and users can remain competitive in AI R&D and
deployment. There is already an emerging set of AI chips that are specialized
for different tasks, which fall broadly into three categories. The first is
graphics processing units (GPUs), which are mostly used to train and develop
AI algorithms. The second is field programmable gate arrays (FPGAs), which
are mostly used to apply trained AI algorithms to new data inputs. FPGAs are
different from other AI chips because their architecture can be modified by
programmers after fabrication. 162 The third group of AI chips is application-
specific integrated circuits (ASICs), which can be used for either training or
inference tasks. ASICs have hardware that is customized for a specific
algorithm and typically provides more efficiency than FPGAs do, but because
they are so narrow in their application, they grow obsolete more quickly as
new AI algorithms are created.
There is increasing demand from AI developers for specialized chips that are
more efficient for AI because the rate of improvements in traditional
processing chips is getting slower as the ability to pack more transistors onto
a single processor is beginning to reach its physical limits. Fortunately, the
United States is still the world leader in designing chips for AI systems. The
Center for Data Innovation’s 2021 report Who Is Winning the AI Race: China,
the EU, or the United States? — 2021 Update finds that at least 62 firms in
the United States are developing AI chips, compared with 29 firms in China
and 14 in the European Union. 163 The United States has many advantages for
AI chip production, including high-quality infrastructure and logistics,
innovation clusters, leading universities, and a history of leadership in the
field. Moreover, Chinese AI chip firms are reliant on U.S. electronic design
automation software, which is the category of software tools for designing
electronic systems such as integrated circuits. 164
contact: [email protected]
datainnovation.org