National Security Commission on Artificial Intelligence. Final Report

CHAPTER 6

The artificial intelligence (AI) competition will not be won by the side

with the best technology. It will be won by the side with the best,

most diverse and tech-savvy talent. The Department of Defense

(DoD) and the Intelligence Community (IC) both face an alarming

talent deficit. This problem is the greatest impediment to the U.S.

being AI-ready by 2025. National security agencies need more

digital experts now or they will remain unprepared to buy, build, and

use AI and its associated technologies. Digital expertise is the most

important requirement for government modernization, but few parts of

government have adequately invested in building a digital workforce.¹

“DoD and the IC both face an

alarming talent deficit.”

To expand its digital and AI workforce, the government needs to:

• Organize technologists within government through a talent management system

designed to house highly skilled specialists;

• Recruit people who already have the skills the government needs, such as industry

experts, academics, and recent college graduates;

• Build its own workforce by training and educating current and future government

employees; and

• Employ its digital workforce more effectively to ensure digital talent can perform

meaningful work once they are in government.

121

TECHNICAL TALENT IN GOVERNMENT

“Digital expertise is the most

important requirement for

government modernization ...”

The Current Model.

Government organizations responsible for creating AI solutions are struggling to build

their digital workforce. Real obstacles impede recruiting and retaining AI practitioners and

broader digital talent. The government does not compete with private-sector salaries and

suffers from a cumbersome hiring process, and all reforms are hindered by a slow security

clearance process.

We should not accept an undesirable status quo as the inevitable future. The government

can compete with the private sector for talent. The government may not match private-

sector salaries, but it does offer the opportunity to tackle national security challenges

and to make a substantial contribution to society. The biggest obstacle hindering the

recruitment of digital talent is not compensation. It is the perception, and too often the

reality, that it is difficult for digital talent in government to perform meaningful work, with

modern computing tools, at the forefront of a rapidly changing field.²

“We should not accept an

undesirable status quo as

the inevitable future. The

government can compete with

the private sector for talent.”

122

CHAPTER 6

The Commission is not persuaded by the argument that the government should focus

on project management and data collection and management, and outsource all

development. We have heard this argument from leaders who do not believe it is feasible

for the government to hire or train its own AI experts. Interestingly, we have not heard this

argument from industry.

“Government strategies

that do not develop a

government technical

workforce are short-sighted.”

Government strategies that do not develop a government technical workforce are short-

sighted. Government agencies that rely solely on contractors for digital expertise will

become incapable of understanding the underlying technology well enough to make

successful acquisition decisions independent of contractors.³ This situation creates

national security risks. While contractors should continue to play a critical role, they

are incentivized, and in some sense required, to fulfill the terms of their contract, not to

pursue overall system improvements or to disagree with poorly thought-out requirements

or ineffective strategies. As a result, agencies that rely on contractors force their digital

experts to have a secondary voice in key decisions, even those related to their field of

expertise. The government will always have contractors. But the government can and

should grow its own digital workforce.

Organize.

How a digital workforce is organized is as important as the workforce’s level of expertise.

To generate and manage a proficient digital workforce at the scale required by the national

security enterprise, the government needs to establish a talent management framework

tailored to the task.

Recommendation

Departments and select agencies should create Digital Corps. We propose that departments

and select agencies should create Digital Corps that would recruit, train, and educate

personnel; place people in and remove them from digital workforce billets; manage digital

careers; and set standards for digital workforce qualifications. Departments and select

agencies would create billets for members of these Digital Corps and provide guidance to

members about the work they perform for the agencies.

123

TECHNICAL TALENT IN GOVERNMENT

The Digital Corps model is inspired by the Army’s Medical Corps, which organizes

experts with specialized healthcare skills that do not fit into the Army’s traditional talent

management framework. Like the Medical Corps, agency-specific Digital Corps should

have specialized personnel policies, guidelines for promotion, training resources, and

certifications to demonstrate proficiency in new digital areas.

Recruit.

To fill these Digital Corps and to improve its broader digital workforce, the government needs

to improve recruiting and the hiring process, accelerate security clearances, use temporary

hiring vehicles such as the Intergovernmental Personnel Act, and build mechanisms for

part-time civilian service.⁴ Many AI and other digital practitioners are interested in working

with the government as either full-time employees or part-time employees. Of those

desiring full-time employment, some seek an entire career as a government civilian or in

the military. Others are less willing to make long-term commitments and instead desire to

become temporary, full-time employees, fellows, talent exchange participants, or military

reservists. A third group is willing to work with or for the government part-time, but they

are unwilling to become full-time civilian employees and have no desire to serve as part

of the military.

Gaps in the

Recruitment Model.

124

CHAPTER 6

Establish a civilian National Reserve Digital Corps. The government should tap into the

Recommendation

pool of technologists willing to contribute part of their time to public service by creating

a mechanism to hire them. While part-time employees are not a substitute for full-time

employees, they can help improve AI education, perform data triage and acquisition, help

guide projects and frame digital solutions, build bridges between the public and private

sectors, and take on other important tasks.

To eliminate this recruitment gap, the government should establish a civilian National

Reserve Digital Corps (NRDC) modeled after the military reserve’s commitments and

incentive structure. Members of the NRDC would become civilian special government

employees in one of the agency Digital Corps and work at least 38 days each year as

advisors, instructors, or developers across the government.

Streamline the hiring process and expand digital talent pipelines. The government hiring

Recommendation

system’s problems are well known: It moves too slowly, struggles to attract experts in a

competitive market, and makes it difficult for experts who are young or do not have a degree

to be hired, especially at a pay grade matching their level of expertise. These challenges

are not caused by a lack of hiring authorities or an inherently slow hiring process. The

Commission has been unable to identify a gap in hiring authorities for the digital workforce.

To clear this recruiting bottleneck, the government needs to expand science, technology,

engineering, and mathematics

(STEM) and AI talent pipelines from universities to

government service, streamline the hiring process, and create agency- and military

service-specific digital talent recruiting offices either for Digital Corps or agencies. The

recruiting offices would monitor their corps, agency, or service’s need for specific types of

digital talent and be empowered to recruit technologists virtually, by attending conferences

and career fairs, recruiting on college campuses, hosting prize competitions, and offering

scholarships, recruiting bonuses, and referral bonuses.

National Reserve

Digital Corps.

125

TECHNICAL TALENT IN GOVERNMENT

The Commission has been

unable to identify a gap in

hiring authorities for the

digital workforce.

Standing Digital Corps will oversee government-wide progress and make recommendations

to expand and improve digital talent hiring and pipelines. They should also be able to

experiment with new authorities.

Build.

The government will not be able to recruit its way out of its technology workforce deficit.

AI and digital talent are simply too scarce in the United States. In 2020, there were more

than 430,000 open computer science jobs in the United States, while only 71,000 new

computer scientists graduate from American universities each year.⁵ The government

should also make a new commitment to building its workforce from the ground up with a

major initiative.

“The United States needs to

establish a new service academy

to train future civil servants in the

digital skills that are needed to

modernize the government.”

126

CHAPTER 6

Establish a United States Digital Service Academy. The United States needs to establish

Recommendation

a new service academy to train future civil servants in the digital skills that are needed

to modernize the government. The United States Digital Service Academy (USDSA)

would be an accredited, degree-granting university that receives both government and

private funding, is managed by a purpose-built independent agency within the federal

government, and meets the government’s needs for digital expertise--as determined by

an interagency board, assisted by a Federal Advisory Committee composed of private-

sector and academic technology leaders. The USDSA should be modeled off of the five

U.S. military service academies but produce trained and educated government civilians

for all federal government departments and agencies.

Proposed

Phase One

Implementation

(Years 1-2)

Plan for USDSA.

• Identify and secure an appropriate site for initial

• Appropriate $40 million to pay for administrative costs.

USDSA build-out with room for future expansion.

• Satisfy the necessary requirements set by the

• Identify gaps in the government’s current and

Department of Education, as well as the state the

envisioned digital workforce by an interagency task

USDSA is in, for degree-granting approval.

force under Office of Personnel Management

• Apply for degree program-specific accreditation through

leadership.

the Computing Accreditation Commission on Colleges

• Establish the USDSA administration as a new

of Accreditation Board for Engineering and Technology.

Executive Branch agency with an individual

• Apply for accreditation with a regional accrediting

appropriation that will be responsible for the phased

organization approved by the Department of Education

implementation plan and the management of the

and Council for Higher Education Accreditation in order

institution.

to be granted “Candidate” status.

• Recruit tenure-track faculty.

• Construct the initial physical infrastructure.

• Recruit adjunct faculty, primarily from private-sector

• Appropriate additional costs for the selection and

technology companies.

purchase of the physical location and construction of

• Grant the USDSA the authority to accept outside funds

the infrastructure.

and gifts from individuals and corporations for startup,

maintenance, and infrastructure costs.

Phase Two

Phase Three

(Years 3-5)

(Years 6-7)

• Begin classes with an initial class of 500 students at

• Graduate the first class.

the beginning of year three.

• Ongoing improvement through accreditation

• Demonstrate compliance with all requirements and

assessments.

standards of the regional accrediting organization in

• Assess, and as appropriate, expand

order to be granted Membership status.

class sizes.

127

TECHNICAL TALENT IN GOVERNMENT

“Digitally talented people should

be able to reasonably expect

to spend a career performing

meaningful work focused on their

field of expertise in government.”

Employ.

Digitally talented people should be able to reasonably expect to spend a career

performing meaningful work focused on their field of expertise in government. Without

such an expectation, they are unlikely to join the government workforce, and without their

experience matching expectations, they are unlikely to stay for long. Aligning expectations

and experience for the digital workforce requires three changes:

• Opportunity for technologists to spend an entire career focused on the field they are

passionate about;

• Well-informed leaders, some of whom are digitally proficient themselves; and

• Access to tools, data sets, and infrastructure.

These changes are more tactical than those described above, but no less impactful.

Strategic initiatives succeed or fail at the tactical level, and many digital initiatives that

might otherwise have strategic impact are struggling or failing tactically in part because

the government does not employ its technologists effectively.

Establish new digital career fields. New career fields challenge an organization’s definition of

Recommendation

its necessary competencies and, potentially, the nature of its identity. If the military services

create career fields for software developers and data scientists, this will almost inevitably

change what it means to be a soldier, sailor, airman, or marine, much as the introduction

of aviation did generations ago. The government should create civilian occupational series

for software development, software engineering, knowledge management, data science,

and AI. The military services should create career fields in software development, data

science, and AI, with both management and specialist tracks. Digital corps will need

additional career fields as they develop, but these steps will establish a strong foundation.

128

TECHNICAL TALENT IN GOVERNMENT

Chapter 6 - Endnotes

¹ There are pockets of excellence in several parts of the government, such as in the United States

Digital Service, Kessel Run, the Army Artificial Intelligence Task Force, the USAF-MIT AI Accelerator,

components of the Intelligence Community, and the national labs--but there are too few, and they

have not spread widely enough across the government. Agencies’ requirements for the size and type

of AI workforce vary, but every agency NSCAI has engaged has expressed a need to expand its AI

workforce, and the recommendations here are broadly applicable.

² NSCAI staff discussions with the Defense Innovation Board and Defense Digital Service (May 2019).

³ William A. LaPlante, Owning the Technical Baseline, Defense AT&L at 18-20 (July-Aug. 2015), https://

apps.dtic.mil/dtic/tr/fulltext/u2/1016084.pdf.

⁴ For more information on the Intergovernmental Personnel Act, see Intergovernmental Personnel Act,

OPM (last accessed Feb. 1, 2021), https://www.opm.gov/policy-data-oversight/hiring-information/

intergovernment-personnel-act/.

⁵ Code.org (last accessed Jan. 11, 2021), https://code.org/promote. See also Oren Etzioni, What

Trump’s Executive Order on AI Is Missing: America Needs a Special Visa Program Aimed at Attracting

More AI Experts and Specialists, Wired (Feb. 13, 2019), https://www.wired.com/story/what-trumps-

executive-order-on-ai-is-missing/.

⁶ For the AI career field in particular, TensorFlow is one of the world’s most popular libraries for

training neural networks and other machine learning (ML) algorithms. PyTorch is another open-source

library that aids in transforming research prototypes to production-ready machine learning models.

130

CHAPTER 7

Artificial intelligence (AI) systems must be developed and fielded

with justified confidence.¹ If AI systems do not work as designed,

or are unpredictable in ways that can have significant negative

consequences, then leaders will not adopt them, operators will not

use them, Congress will not fund them, and the American people will

not support them.

“If AI systems ... are

unpredictable in ways that

can have significant negative

consequences, then leaders will

not adopt them, operators will

not use them, Congress will not

fund them, and the American

people will not support them.”

Achieving acceptable AI performance often is linked to the decision to accept some level

of risk. No technology works perfectly under all conditions. Risk calculus changes with

circumstances. The variables and considerations that inform judgments to rely on AI will

vary significantly across military, intelligence, homeland security, and law enforcement

missions. In a high-threat environment like combat, in some cases it may be reasonable

to employ a system offering some immediate military advantage, while recognizing that

it might fail; in other cases, however, a reasonable commander might want the highest

assurances of AI reliability before fielding when lives are at risk.

133

ESTABLISHING JUSTIFIED CONFIDENCE IN AI SYSTEMS

“As departments and agencies

rely more heavily on machines,

a central guiding principle

across national security scenarios

is the continued centrality of

human judgment.”

As departments and agencies rely more heavily on machines, a central guiding principle

across national security scenarios is the continued centrality of human judgment. Those

charged with utilizing AI need an informed understanding of risks, opportunities, and

tradeoffs. They need awareness of the possibilities and limitations in a system’s expected

performance. Ultimately, they need to formulate an educated answer to this question: In the

given circumstance, how much confidence in the machine is enough confidence? These

issues bear on the full lifecycle of an AI system--from acquisition or system development

and the thresholds for justified confidence to deploy a specific AI-intensive system to the

performance of the system in the field.

While there is no absolute assurance of perfection, there are policies and best practices

that support making these decisions responsibly. Agencies are broadly aware of the

principal challenges in employing AI systems and the necessity of incorporating best

practices in the engineering and management of AI systems.

The Commission has produced a detailed framework to guide the responsible

development and fielding of AI across the national security community (see the Appendix

on Key Considerations for Responsible Development and Fielding of AI). It contains key

considerations for policymakers and technical practitioners covering the full breadth

of the AI lifecycle. The framework includes recommended practices that should be

integrated and updated as the technology advances. The Commission is heartened that

some departments have already taken actions to integrate recommendations from our

framework.²

To assist agencies in meeting baseline criteria for Responsible AI, we highlight the main

challenges and key recommendations in our framework across five issue areas.

134

CHAPTER 7

1. Robust and Reliable AI.

Current AI systems, such as those used for perception and classification, have different

kinds of failure--characterized as rates of false positives and false negatives. They are

often brittle when operating at the edges of their performance competence, and it is

difficult to anticipate their competence boundaries.³ They are also vulnerable to attack,

and they can exhibit unwanted bias in operation. For national security missions, these can

be serious problems. U.S. government agencies should:

Focus more federal R&D investments on advancing AI security and robustness. These

Recommendation

investments should also advance the interpretability and explainability of AI systems, so

users can better understand whether the systems are operating as intended.

Consult interdisciplinary groups of experts to conduct risk assessments, improve documentation

Recommendation

practices, and build overall system architectures to limit the consequences of system failure.⁴

Such architectures should securely monitor component performance and handle errors

when anomalies are detected⁵; contain AI components that are self-protecting (validating

input data) and self-checking (validating data passed to the rest of the system); and include

aggressive stress testing.

“The government needs

AI systems that augment

and complement human

understanding and

decision-making so that the

complementary strengths of

humans and AI can be leveraged

as an optimal team. Achieving

this remains a challenge.”

135

ESTABLISHING JUSTIFIED CONFIDENCE IN AI SYSTEMS

2. Human-AI Interaction and Teaming.

The government needs AI systems that augment and complement human understanding

and decision-making so that the complementary strengths of humans and AI can be

leveraged as an optimal team. Achieving this remains a challenge. For instance, humans

are prone both to over-trusting and to under-trusting machines depending on context.

Challenges also exist for measuring the performance of human-AI teams, conveying

enough information while avoiding cognitive overload, enabling humans and machines

to understand the circumstances in which they should pass control between each other,

and maintaining appropriate human engagement to preserve situational awareness and

meaningfully take action when needed. Agencies will also need to determine machine

performance standards and expectations as compared with humans. The government

should:

Pursue a sustained, multidisciplinary initiative through national security research labs to enhance

Recommendation

human-AI teaming. This initiative should focus on maximizing the benefits of human-AI

interaction; better measuring human performance and capabilities when working with

AI systems, including testing through continuous contact and experimentation with end

users; and helping AI systems better understand contextual nuances of a situation.

Clarify policies on human roles and functions, develop designs that optimize human-machine

Recommendation

interaction, and provide ongoing and organization-wide AI training.

DoD AI Total RDT&E Investments by Research Area, FY 2015-2025

Source: Govini

DoD AI

total RDT&E

Investments.

This figure displays estimated DoD spending levels across five major research categories devised by NSCAI

commissioners, indicating that investments in human-AI interaction lags behind other research categories.

Note the spending levels presented in figure represent estimates based on an analysis of DoD RDT&E budget

documents for FY2021-FY2025. See Analysis of DoD RDT&E Investments in AI, NSCAI (on final with the

Commission). Due to inherent quality issues in the source data, estimates presented contain significant, difficult

to estimate margins of error.

136

CHAPTER 7

3. Testing and Evaluation, Verification and Validation (TEVV).

Having justified confidence in AI systems requires assurances that they will perform

as intended, including when interacting with humans and other systems. The TEVV of

traditional legacy systems is not sufficient at providing these assurances. As a result,

agencies lack common metrics to assess trustworthiness that AI systems will perform as

intended. To minimize performance problems and unanticipated outcomes, an entirely

new type of TEVV will be needed. This is a priority task, and a challenging one. The federal

government will need to increase R&D investments to improve our understanding of how to

conduct AI and software-related TEVV. Toward this end:

Recommendation

DoD should tailor and develop TEVV policies and capabilities to meet the changes needed

for AI as AI-enabled systems grow in number, scope, and complexity in the Department. This

should include establishing a TEVV framework and culture that integrates continuous

testing; making TEVV tools and capabilities more readily available across DoD; updating or

creating live, virtual, and constructive test ranges for AI-enabled systems; and restructuring

the processes that underlie requirements for system design, development, and testing.⁶

National Institute of Standards and Technology (NIST) should provide and regularly refresh a

Recommendation

set of standards, performance metrics, and tools for qualified confidence in AI models, data,

and training environments, and predicted outcomes. NIST should lead the AI community

in establishing these resources, closely engaging with experts and users from industry,

academia, and government to ensure their efficacy.

4. Leadership.

Responsible development and fielding of AI requires end users and senior leaders to be

aware of system capabilities and limitations so that they are not misused. It also requires

subject-matter experts to support training, acquisition, risk assessment, and adoption of

best practices as they evolve. Today, only the DoD has a dedicated lead for Responsible

AI; employees in national security agencies taking on these roles typically do so on a

voluntary, part-time basis. Without full-time dedicated staff, agencies will not succeed in

fully adopting and implementing these recommended practices. The government should:

Appoint a full-time, senior-level Responsible AI lead in each department or agency critical

Recommendation

to national security and each branch of the armed services. Such an official should drive

Responsible AI training, provide expertise on Responsible AI policies and practices, lead

interagency coordination, and shape procurement policies.

Create a standing body of multidisciplinary experts in the National AI Initiative Office. The

Recommendation

standing body would provide advice to agencies as needed on responsible AI issues. The

group should include people with expertise at the intersection of AI and other fields such

as ethics, law, policy, economics, cognitive science, and technology, including adversarial

AI techniques.

137

ESTABLISHING JUSTIFIED CONFIDENCE IN AI SYSTEMS

5. Accountability and Governance.

Congress and the public need to see that the government is equipped to catch and

fix critical flaws in systems in time to prevent inadvertent disasters and hold humans

accountable, including for misuse. Agencies need the ability to monitor AI performance

as systems run (to assess if they are performing as intended) and to build systems with

the necessary instrumentation to do so.⁷ Departments and agencies critical to national

security and oversight entities have all expressed challenges with having visibility into their

systems, while vendors are calling for clarity on instrumentation/auditability requirements.

Government agencies should:

Adapt and extend existing accountability policies to cover the full lifecycle of AI systems and

Recommendation

their components.

Establish policies that allow individuals to raise concerns about irresponsible AI development

Recommendation

and institute comprehensive oversight and enforcement practices. These should include

auditing and reporting requirements, a review mechanism for the most sensitive or high-

risk AI systems, and appeals and grievance processes for those affected by the actions

of AI systems.

138

ESTABLISHING JUSTIFIED CONFIDENCE IN AI SYSTEMS

Chapter 7 - Endnotes

¹ The term “justified confidence,” taken from a widely used international standard, uses a specific

definition of assurance as being “grounds for justified confidence.” It notes that “stakeholders need

grounds for justifiable confidence prior to depending on a system” and that “the greater the degree

of dependence, the greater the need for strong grounds for confidence.” ISO/IEC/IEEE International

Standard - Systems and Software Engineering - Systems and Software Assurance, IEEE/ISO/IEC

15026-1 (2019), https://standards.ieee.org/standard/15026-1_Revision-2019.html.

² The Department of Defense’s Joint Artificial Intelligence Center (JAIC) Subcommittees on

Responsible AI and Test & Evaluation have both conducted substantial mapping exercises to

determine which existing practices correspond to recommendations found in the Key Considerations.

Recommendations from the Key Considerations are also reinforced by inclusions in the Department of

Homeland Security (DHS)’s AI Strategy. See Department of Homeland Security Artificial Intelligence

Strategy, DHS (Dec. 2020), https://www.dhs.gov/publication/us-department-homeland-security-

artificial-intelligence-strategy?topic=intelligence-and-analysis.

³ Like other intelligent systems, including software and humans, AI systems have competency

limitations. However, we have less science to understand the performance limitations of AI systems

including why, when, and how they fail.

⁴ Such interdisciplinary teams should explore the possibility of documentation/labels specifying the

narrow task/mission for which a system was designed and tested. As noted in the Appendix on Key

Considerations for Responsible Development & Fielding of AI, documentation of the AI lifecycle

should include information about the data used to train and test a model and the methods used to test

a model, both based on the context in which it will be used. It also should include requirements for re-

testing, retraining, and tuning when a system is used in a different scenario or setting.

⁵ Monitoring can add a layer of robustness, but must itself also be guarded to prevent new openings

for external espionage or tampering with AI systems.

⁶ Upgrades to digital infrastructure, as outlined in Chapter 2 of this report, will be required to augment

physical test ranges to create digital testing environments that can leverage digital twins.

⁷ Cases in which new sensors and instrumentation are added can also introduce new vulnerabilities.

It is especially important to ensure that the overall architecture of such systems is secure against

external espionage and tampering.

140

CHAPTER 8

The basic purpose of the American government is to protect the

security and liberty of the American people. Americans have a

long tradition of debating how best to achieve these twin goals

when tensions arise between them. The two decades following

9/11 saw intensive efforts to calibrate the government’s powers to

stop another terrorist attack with its obligations to respect individual

rights and liberties. Artificial intelligence (AI) is ushering in the next

era of this debate because new technologies offer government

agencies more powerful ways to collect and process information,

track individuals’ behavior and movements, and act on the basis

of computer-generated analyses.

In addition to supporting military and intelligence missions abroad, these tools are promising

for national security purposes closer to home—whether to examine foreign intelligence to

find signs of danger to the United States, to screen for threats at the borders, to protect

against cyber attacks and information operations, or to identify domestic terrorism plots.

Americans have concerns that AI applications used for these security and public safety

purposes—especially those involving biometric technologies or the analysis of aggregated

personal data—will invade their privacy, restrict their freedoms of speech and assembly,

and reinforce bias and discrimination. At the same time, if applied effectively, AI can help

improve protections for privacy and civil liberties. Machine analysis could be more precise,

and AI systems potentially could enhance oversight through real-time monitoring.

For the United States, as for other democratic countries, use of AI by officials must comport

with principles of limited government and individual liberty. These principles do not uphold

themselves. In a democratic society, any empowerment of the state must be accompanied

by wise restraints to make that power legitimate in the eyes of its citizens.

As this report argues, the promise of emerging AI technologies to enhance national

security is real and significant. The ability of U.S. intelligence, homeland security, and

law enforcement agencies to develop and use them for national security purposes must

be preserved. To do so, however, the government must ensure that their use is effective,

legitimate, and lawful. Public trust will hinge on justified assurance about compliance with

privacy, civil liberties, and civil rights.

Democratic AI Governance and Novel Challenges for Privacy,

Civil Liberties, and Civil Rights.

With new models of techno-authoritarian governance gaining traction abroad, the United

States must continue to serve as a beacon of democratic values. The democratic model

143

UPHOLDING DEMOCRATIC VALUES: PRIVACY, CIVIL LIBERTIES, AND CIVIL RIGHTS IN USES OF AI FOR NATIONAL SECURITY

must prove its resilience in the face of emerging technological changes that could

challenge it. Fundamentally, we are confident that the American system—and the rules,

norms, and institutions that uphold it—can adapt to uphold the dual imperatives of security

and liberty in the AI era.

For the Intelligence Community (IC), core features of that system include laws, rules, and

procedures to minimize the collection, retention, and dissemination of U.S. persons’ data,

as well as oversight from all three branches of government.¹ Homeland security and law

enforcement agencies likewise operate within frameworks of policy, oversight, and judicial

review that guide border protection and criminal investigations. Ultimately, the actions of

all federal agencies are subject to the Constitution’s guarantees.

Within this context, the advent of modern AI—and the novel capabilities it can bring

to intelligence, homeland security, and law enforcement missions—raises a number of

concerns and difficult questions and challenges with respect to the privacy, civil liberties,

and civil rights of U.S. persons. For example:

•

AI-powered analytics can help officials process and make sense of huge amounts of

information, which can be aggregated to form a revealing “mosaic” picture of a person’s

activities, whereabouts, and patterns of behavior.² Combining disparate data streams

involving geolocation, web browsing, financial transactions, and other data sources

creates the possibility of new insights for analysts or investigators. This could be highly

useful to identify threats, but it has also raised questions about the proper scope and

authorization for border or law enforcement searches.³

•

Much of this personal information is held by private companies. This fact of modern

digital life has raised constitutional questions about whether and when individuals

should have a “reasonable expectation of privacy” in the information they provide to

third parties like technology firms—and questions about the circumstances in which

that information may be accessed and utilized by intelligence, homeland security, or law

enforcement agencies for a legitimate national security purpose.⁴

•

AI can help automate aspects of data collection and analysis. Such methods can

augment the ability of analysts or investigators to sif through and triage masses of

information to establish patterns or pinpoint threats. But they also raise questions about

the proper roles of machine and human analysis in these processes, including for

making predictive judgments. To the extent that an AI system’s functions are opaque,

it may be difficult to trace and justify the computational process that led the system to

make a recommendation. Determining when and how to rely on algorithms is especially

pertinent to minimization and querying procedures in the IC and to building cases for

law enforcement action.⁵

•

AI models can evolve based on changing data and interaction with other models,

leading to unexpected outcomes. As a result, AI systems require more continuous

testing and evaluation than prior generations of technology.

•

Unintended bias can be introduced during many stages of the machine learning (ML)

process, which can lead to disparate impacts in American society, a problem that has

been documented in law enforcement contexts.⁶

144

CHAPTER 8

Tenets for Managing AI Challenges.

This Commission will not endeavor to draw all of the lines for what may be permissible

or wise in particular circumstances. However, important principles to follow in different

national security contexts include the following:

Foreign Intelligence Collection and Analysis: The Office of the Director of National

Intelligence (ODNI) AI Ethics Guidance to the Intelligence Community is an encouraging

step, as it places strong emphasis on utilizing AI for foreign intelligence missions in

ways that uphold the privacy and civil liberties of Americans.⁷ As these guidelines are

implemented, it will be important to pay close attention to ensuring that data minimization,

retention, and querying procedures are adequate and rigorously enforced.

Border Security: AI surveillance and analysis capabilities can make the government’s

operations more efficient and effective at the borders and ports of entry. But to sustain

public support for these uses, the Department of Homeland Security (DHS) must take care

to ensure that automated screening processes lead agents only to the information they

need and are authorized to access, and do not impermissibly single out individuals based

on characteristics such as race or religion.

Domestic Security and Public Safety: Rapid advances in AI-enabled technologies for

law enforcement purposes, including biometric surveillance techniques such as facial

recognition, may be outpacing rules for their proper use. The government must exercise

special caution in managing risks to bedrock constitutional principles including equal

protection, due process, freedom from unreasonable searches and seizures, and freedoms

of speech and assembly.⁸

In carrying out these missions, it will be important to maintain clear distinctions between

appropriate authorities in these different national security contexts. It is also important to

gain greater public confidence by enhancing transparency, improving the performance

and reliability of AI technologies, ensuring due process, and strengthening oversight. With

these tenets in mind, the government should take the following steps.

145

UPHOLDING DEMOCRATIC VALUES: PRIVACY, CIVIL LIBERTIES, AND CIVIL RIGHTS IN USES OF AI FOR NATIONAL SECURITY

“Agencies should assess near-

term opportunities and research

gaps in applications of AI to

address privacy and civil liberties

challenges ...”

Recommendations.

Invest in and adopt AI tools to enhance oversight and auditing in support of privacy and

Recommendation

civil liberties. Agencies should assess near-term opportunities and research gaps in

applications of AI to address privacy and civil liberties challenges, such as ML techniques

for classification, recommendation, anomaly detection, and other applications.⁹ Examples

of advances in AI to improve auditing include tools that support financial audits and

model risk management. Agencies should examine the utility of these and other current or

emerging practices.¹⁰

Improve public transparency about how the government uses AI. There is a lack of

Recommendation

transparency into agency policies and procedures and into the accuracy of AI systems

that may impact civil liberties.¹¹ The “black box” nature of some ML systems only adds to

this opacity.¹² More transparency could help to ease public concerns. Of course, in certain

operational contexts, especially for intelligence and law enforcement agencies, secrecy

is essential to the mission. However, existing transparency mechanisms could be utilized

more effectively, and, in some cases, revised. New agency reporting requirements would

also be beneficial.

• For AI systems that impact U.S. persons, Congress should require AI Risk Assessment

Reports and AI Impact Assessments from the Intelligence Community, the Department

of Homeland Security, and the Federal Bureau of Investigation. These should assess

the privacy, civil liberties and civil rights implications for each qualifying AI system or

significant system refresh.¹³

• DHS and the FBI should also improve practices for issuing system of records notices

and privacy impact assessments to provide a more holistic view of the role of AI systems

before they are fielded.

146

CHAPTER 8

Develop and test systems with the goal of advancing privacy preservation and fairness. ML

Recommendation

systems in particular require ongoing assessments of privacy and fairness assurances,

including the specific definition of fairness being assumed. Although an ML system may

meet requirements at a static point in time, ongoing compliance is not a given once

the system is operational. This is in large part due to changing data, the introduction of

unintended bias, and potential re-identification of anonymized data.¹⁴ This is a complex

technical area, and continued work in the technical, legal, and policy domains is required

to find greater consensus on technical approaches to preserving privacy, civil liberties,

and civil rights.¹⁵ Meanwhile, agencies should take several steps:

• Assess and mitigate risks in the design, development, and testing of AI systems.

In addition to conducting risk assessments for the privacy, civil liberties and civil rights

of U.S. persons, IC elements, DHS, and the FBI should take measures to mitigate those

risks, and document remaining risks that are accepted. In doing so, they should adopt

practices from the Key Considerations, including using privacy protections such as

robust anonymization, and when possible, privacy-preserving technology; taking steps

to mitigate bias in development and testing; and assessing model performance on an

ongoing basis.¹⁶

• Identify an office, committee, or team in each agency that will conduct a pre-

deployment review of AI technologies that will impact privacy, civil liberties,

and civil rights. This should include review in advance of their deployment and for

compliance over the lifespan of the system. An office in each IC element, DHS, and

the FBI should be equipped to assess data, model, and system documentation, and to

assess the testing results of systems with respect to their intended use.

• Establish third-party testing centers for national security-related AI systems that

could impact U.S. persons. Such independent, third-party testing could be done by

a national laboratory, a University Affiliated Research Center, or a Federally Funded

Research and Development Center. Such testing should be mandatory for high-stakes

systems but otherwise voluntary.¹⁷ It would provide agencies with additional expertise to

help overcome in-house limitations.

Recommendation

Strengthen the ability of those impacted by government actions involving AI to seek redress and

have due process. AI systems will make errors.¹⁸ Agencies have to accept non-zero false

positive and false negative rates in order to deploy any AI system. It is important to ensure

opportunities for redress, consistent with the constitutional principle of due process—

for example, when a system error leads to a benefit being denied (e.g., visa approval);

restrictions on movement (e.g., being placed on a no-fly list); or an arrest.¹⁹ There are also

due process concerns in cases in which AI contributes to building a case to press criminal

charges.²⁰ We recommend two steps to start addressing these issues:

• Review DHS and FBI policies and practices that may impact due process and the

ability to seek redress. DHS and the FBI should review agency policies and practices to

ensure that parties aggrieved by government action involving AI technology, including

through system actions or misuse, can seek redress and clearly know how to do so. This

review should include whether adequate notice of AI use in decision-making is provided

to impacted parties, as well as the degree to which AI systems can be audited to trace

the process by which a system arrived at a recommendation, if it is contested.

147

UPHOLDING DEMOCRATIC VALUES: PRIVACY, CIVIL LIBERTIES, AND CIVIL RIGHTS IN USES OF AI FOR NATIONAL SECURITY

• Issue Attorney General guidance on AI and due process. The guidance should

describe how relevant agencies should safeguard the due process rights of U.S. persons

when AI use may lead to a deprivation of life or liberty.

Strengthen oversight mechanisms to address current and evolving concerns. The advancement

Recommendation

of AI requires a forward-looking approach to oversight that anticipates the continued

evolution and adoption of new technologies and better positions the government to

responsibly manage their employment well into the future.

“The advancement of AI requires

a forward-looking approach

to oversight that anticipates

the continued evolution and

adoption of new technologies,

and better positions the

government to responsibly

manage their employment well

into the future.”

The government should:

• Establish a task force to assess the privacy and civil liberties implications of AI and

emerging technologies. The goal of the task force would be to identify gaps and make

recommendations to ensure that uses of AI and associated data in U.S. government

operations comport with U.S. law and values, and to study organizational reforms that

would support this goal. Specifically, it should assess existing policy and legal gaps for

current AI applications and emerging technologies, and make recommendations for:

o legislative and regulatory reforms on the development and use of AI and

emerging technologies and associated data²¹; and

o institutional changes to ensure sustained assessment and recurring guidance

on privacy and civil liberties implications of AI applications and emerging

technologies.

148

CHAPTER 8

• Strengthen the ability of the Privacy and Civil Liberties Oversight Board (PCLOB)

to provide meaningful oversight and advice on AI use for national security. Since

its creation in 2007, following a recommendation of the 9/11 Commission, PCLOB has

had an especially important role in overseeing, and advising the government on, U.S.

counterterrorism missions. In recent years, it has started turning attention to the use of

new technologies in foreign intelligence collection and analysis.²² The board should

be given visibility into AI systems before they are fielded, including at a more granular

technical level, and should be resourced and staffed to fulfill the more technically

sophisticated mission that the AI era now requires.²³

• Empower DHS Offices of Privacy and Civil Rights and Civil Liberties (CRCL). The

Chief CRCL Officer, in coordination with the Privacy Officer, must play an integral role in

the legal and approval processes for the procurement and use of AI-enabled systems,

including for associated data used in DHS ML systems.

• Require stronger coordination and alignment among federal oversight and

audit organizations. Compliance by agencies with AI documentation and testing

requirements should be supported by rigorous, technically informed oversight. To

achieve this and overcome current audit and oversight impediments, a standing body

should align and coordinate to enhance AI oversight and audit with respect to privacy,

civil liberties, and civil rights.²⁴

149

CHAPTER 8

Chapter 8 - Endnotes

¹ For a compilation of Attorney General guidelines from the IC components, see Status of Attorney

General Approved U.S. Person Procedure under E.O. 12333, ODNI (July 14, 2016), https://www.

dni.gov/files/documents/Table_of_EO12333_AG_Guidelines%20for%20PCLOB_%20Updated%20

July_2016.pdf. Elements of the IC oversight system include counsels and privacy officials within

intelligence agencies, the Department of Justice, independent bodies such as the Privacy and Civil

Liberties Oversight Board, Federal courts including the Foreign Intelligence Surveillance Court, and

the House and Senate intelligence committees.

² On the mosaic concept, see, e.g., Steven M. Bellovin, et al., When Enough Is Enough: Location

Tracking, Mosaic Theory, and Machine Learning, NYU Journal of Law & Liberty, Vol. 8 (Sept. 3, 2013),

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2320019.

³ For an informative discussion of evolving debates over Fourth Amendment regulation of government

searches in the context of AI, see James E. Baker, The Centaur’s Dilemma: National Security Law for

the Coming AI Revolution, Ch. 6 (Brookings, 2020).

⁴ Congress and the Judiciary will need to assess the adequacy of current legal constraints over

the federal government’s obtainment and use of third-party data, including data acquired from

data brokers. Either through evolving case law or legislation, agencies would benefit from clarity

surrounding the Fourth Amendment’s application with respect to third-party data. On third-party

doctrine, see Richard M. Thompson II, The Fourth Amendment Third-Party Doctrine, Congressional

Research Service (June 5, 2014), https://fas.org/sgp/crs/misc/R43586.pdf.

⁵ For a discussion and different views on the implications of human and machine analysis in the

intelligence context, see Robert Litt, The Fourth Amendment in the Information Age, Yale Law Journal

(April 27, 2016), https://www.yalelawjournal.org/forum/fourth-amendment-information-age; Cindy

Cohn, Protecting the Fourth Amendment in the Information Age: A Response to Robert Litt, Yale Law

Journal (July 27, 2016), https://www.yalelawjournal.org/forum/protecting-the-fourth-amendment-in-the-

information-age.

⁶ Concerns about algorithmic error rates and disparate performance across age, skin tones,

and genders are especially pronounced for facial recognition. See Patrick Grother, et al., Face

Recognition Vendor Test, Part 3: Demographic Effects, NIST (Dec. 2019), https://nvlpubs.nist.gov/

nistpubs/ir/2019/NIST.IR.8280.pdf. The Gender Shades Project found that various facial recognition

systems were very accurate for white men, but they were significantly less accurate for women and

people of color (and worst for women of color). See Gender Shades (last accessed Jan. 11, 2021),

http://gendershades.org/.

⁷ See Principles of Artificial Intelligence Ethics for the Intelligence Community, ODNI (last accessed

Jan. 11, 2021), https://www.dni.gov/index.php/features/2763-principles-of-artificial-intelligence-

ethics-for-the-intelligence-community.

⁸ Some observers have found a “chilling effect” that impacts the degree to which individuals exercise

freedoms of expression, association, and assembly. See, e.g., Rachel Levinson-Waldman, Hiding in

Plain Sight: A Fourth Amendment Framework for Analyzing Government Surveillance in Public, Emory

Law Journal Vol. 66 (2017), https://scholarlycommons.law.emory.edu/elj/vol66/iss3/4/.

⁹ Xuning (Mike) Tang & Yihua Astle, The Impact of Deep Learning on Anomaly Detection, Law.com

(Aug. 10, 2020), https://www.law.com/legaltechnews/2020/08/10/the-impact-of-deep-learning-on-

anomaly-detection/.

¹⁰ See, e.g., Bernhard Babel, et al., Derisking Machine Learning and Artificial Intelligence, McKinsey &

Company (Feb. 19, 2019), https://www.mckinsey.com/business-functions/risk/our-insights/derisking-

machine-learning-and-artificial-intelligence; Saqib Aziz & Michael Dowling, Machine Learning and

AI for Risk Management, Disrupting Finance at 33-50 (Dec. 7, 2018), https://link.springer.com/

chapter/10.1007/978-3-030-02330-0_3.

151

UPHOLDING DEMOCRATIC VALUES: PRIVACY, CIVIL LIBERTIES, AND CIVIL RIGHTS IN USES OF AI FOR NATIONAL SECURITY

¹¹ For instance, disclosure by U.S. Customs and Border Protection (CBP) when using facial recognition

at airports has been inconsistent, and claims exist that the FBI failed to provide information about its

Next Generation Identification database and use of facial recognition as required by law. In 2020, the

U.S. Government Accountability Office (GAO) found that “CPB’s privacy notices—which inform the

public about its use of this technology—were not always current or available [at airports] where this

technology is being used or on CBP’s website.” Facial Recognition: CBP and TSA Are Taking Steps to

Implement Programs, but CBP Should Address Privacy and System Performance Issues, GAO (Sept.

2, 2020), https://www.gao.gov/products/GAO-20-568; see also The Perpetual Line-Up: Unregulated

Police Face Recognition in America, Georgetown Law Center on Privacy & Technology (Oct. 18,

2016), https://www.perpetuallineup.org/.

¹² In a 2018 report, GAO has raised concerns about lack of transparency from tech companies

that build algorithms and “limited testing on the systems for accuracy.” Artificial Intelligence:

Emerging Opportunities, Challenges, and Implications, GAO (March 2018), https://www.gao.gov/

assets/700/690910.pdf.

¹³ The Commission proposes that the task force described in this chapter, and in the accompanying

Blueprint for Action, should provide binding guidance on two issues: first, when the IC, DHS, and FBI

should prepare and publish AI Risk Assessment Reports and AI Impact Statements; and second, what

should constitute a “qualifying AI system or significant system refresh.”

¹⁴ For example, pseudonymized data can be linked with other data to uncover a cell phone

owner’s identity. See Byron Tau & Michelle Hackman, Federal Agencies Use Cellphone

Location Data for Immigration Enforcement, Wall Street Journal (Feb. 7, 2020), https://www.

wsj.com/articles/federal-agencies-use-cellphone-location-data-for-immigration-enforcement-

11581078600?mod=breakingnews.

¹⁵ See the Appendix of this report containing the abridged version of NSCAI’s Key Considerations for

Responsible Development & Fielding of AI. For further discussion of recommendations to: (1) employ

technologies and operational policies that align with privacy preservation and mitigate unwanted

bias and (2) to continuously monitor and evaluate AI system performance, see sections “Aligning

Systems and Uses with American Values and the Rule of Law” and “System Performance” in Key

Considerations for Responsible Development & Fielding of Artificial Intelligence: Extended Version,

NSCAI (2021) (on file with the Commission).

¹⁶ Key Considerations for Responsible Development & Fielding of Artificial Intelligence, NSCAI (July

2020), https://www.nscai.gov/previous-reports/.

¹⁷ To provide agencies guidance on when such a test mechanism should be leveraged, an

organization should establish guidance on thresholds by which agencies would be required to

conduct third-party testing. This should include criteria for when an AI system may pose high enough

risk for privacy, civil liberties, and civil rights that it would trigger a testing requirement by a third-

party auditor.

¹⁸ See, e.g., Kashmir Hill, Wrongfully Accused by an Algorithm, New York Times (June 24, 2020),

https://www.nytimes.com/2020/06/24/technology/facial-recognition-arrest.html.

¹⁹ An individual’s right to due process, including notice, is grounded in the Constitution and the case

law expounding on that right. Our recommendation’s aim is to help ensure that the government does

its part to uphold these long-standing rights notwithstanding its use of AI.

152

CHAPTER 8

Chapter 8 - Endnotes

²⁰ Due process rights require that individuals have the ability to meaningfully challenge a decision

made against them. In federal criminal trials, this includes the government explaining how an

unfavorable decision was reached, so it can be contested. In cases where AI-assisted or AI-

enabled decisions are made, certain AI techniques will be less conducive to due process. See

Danielle Keats Citron, Technological Due Process, Washington University Law Review (2008), https://

openscholarship.wustl.edu/cgi/viewcontent.cgi?article=1166&context=law_lawreview; Ryan Calo &

Danielle Keats Citron, The Automated Administrative State: A Crisis of Legitimacy, Emory Law Journal

(March 9, 2020), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3553590. Early cases in

which an AI system’s predictions, classifications, or recommendations have been challenged in

court illustrate that defendants encounter substantial impediments in seeking to exercise their rights.

See Litigating Algorithms: Challenging Government Use of Algorithmic Decision Systems, AI Now

Institute (Sept. 2018), https://ainowinstitute.org/litigatingalgorithms.pdf. There are also open questions

including federal rules of evidence and criminal procedure as they relate to AI. For instance,

evidentiary standards for admitting AI evidence in court have yet to be developed.

²¹ Examples include baseline AI standards and policy guidance for biometric identification

technologies, for government procurement of commercial AI products, and for federal data privacy

standards.

153

UPHOLDING DEMOCRATIC VALUES: PRIVACY, CIVIL LIBERTIES, AND CIVIL RIGHTS IN USES OF AI FOR NATIONAL SECURITY

²² See Projects, PCLOB (last accessed Jan. 9, 2021), https://www.pclob.gov/Projects.

²³ PCLOB works alongside multiple oversight organizations to conduct oversight. It will also

be important for PCLOB and such organizations to better align and coordinate to conduct

complementary AI oversight and auditing with respect to privacy, civil liberties, and civil rights.

²⁴ For examples of impediments, see Taka Ariga & Stephen Sanford, A is for Accountability: Oversight

in the Age of Artificial Intelligence, ECA Journal at 88-91 (Jan. 2020), https://www.eca.europa.

eu/Lists/ECADocuments/JOURNAL20_01/JOURNAL20_01.pdf; see also Press Release, Office

of the Inspector General of the Intelligence Community, The Inspector General of the Intelligence

Community Issues Statement on Artificial Intelligence (May 30, 2019), https://www.dni.gov/files/

ICIG/Documents/News/ICIG%20News/2019/May%2030%20-%20AI/Press%20Release%20-%20AI.

pdf; Michael K. Atkinson, Semiannual Report: October 2018-March 2019, Office of the Inspector

General of the Intelligence Community (2019), https://www.oversight.gov/sites/default/files/oig-sa-

reports/20190430_ICIG-SAR_Oct18-Mar19.pdf.

154

CHAPTER 9

The impact of artificial intelligence (AI) on the world will extend far

beyond narrow national security applications. The development of

AI constitutes a new pillar of strategic competition, and it heightens

the competition in existing pillars. The nation with the most resilient

and productive economic base will be best positioned to seize

the mantle of world leadership. That base increasingly depends

on the strength of the innovation economy, which in turn will

depend on AI. AI technologies will drive waves of advancement

in critical infrastructure, commerce, transportation, health,

education, financial markets, food production, and environmental

sustainability.

The race to research, develop, and deploy AI and associated technologies is already

intensifying strategic competition. The U.S. government must embrace the AI competition

and organize to win it. The American approach to innovation, which has served the country

well for decades, must be recalibrated to account for the centrality of the competition

involving AI and associated technologies to the emerging U.S.-China rivalry. To retain

its innovation leadership and position in the world, the United States needs a stronger

government-led technology strategy that integrates promotion and protection policies and

links investments in AI to a larger constellation of related emerging technologies.¹

“The race to research,

develop and deploy AI and

associated technologies is

already intensifying strategic

competition. The U.S.

government must embrace the

AI competition and organize to

win it.”

159

A STRATEGY FOR COMPETITION AND COOPERATION

This chapter articulates the nature of the AI competition and the two prerequisites for

winning it: organizing for technology competition under White House leadership and

establishing the principles for continued cooperation with competitors. The following

chapters (10-16) enumerate the core elements of an integrated strategy and prescribe

actions to ensure the United States wins the AI competition and sets the foundation to win

the broader technology competition. It is foremost an affirmative agenda for growing and

recruiting talent, promoting a diverse AI innovation ecosystem, and investing in the R&D

to harness AI and associated technologies to build a healthier, more prosperous, and

secure society. Protection of research, intellectual property (IP), and investments will be

necessary to complement the effort to invigorate AI competitiveness at home and build a

coalition of like-minded partners in the world.

The United States Government must understand and define the

technology competition, organize for it, and set the terms to engage

with China.

Organizing for the

Understanding the

Competition

Competition.

Establish a White House

National Technology

Competitiveness Council:

Empower a single entity

in the White House to set

strategic direction and

oversee a coordinated

approach to technology

competition.

Made in China 2025

and AI World Leader

2030:

Managing the Competition

China has already

Begin a U.S.-China

developed a strategy

Comprehensive S&T

for technology

Dialogue: Establish a high-

leadership, picked key

level dialogue with China

technology sectors,

to discuss challenges and

started high-tech

manage tensions related to

projects within key

emerging technologies (e.g.

sectors, and delegated

AI, quantum, biotech).

authority across

individual government

agencies.

Winning the Competition

Develop a National

Technology Strategy: The

Technology Competitiveness

Council should develop a

national strategy to guide

U.S. policy across the

constellation of emerging

technologies starting with AI.

160

CHAPTER 9

The U.S.-China AI Competition Is Serious and Complex.

The leading indexes that measure progress in AI development generally place the United

States ahead of China.² However, the gap is closing quickly. China stands a reasonable chance

of overtaking the United States as the leading center of AI innovation in the coming decade.³

In recent years, technology firms in China have produced pathfinding advances in natural

language processing,⁴ facial recognition technology,⁵ and other AI-enabled domains. China’s

businesses, investors, technologists, and academics are integral to global AI development.

China’s social media and e-commerce companies compete for users around the world. Its

telecoms build global 5G infrastructure. Its venture capitalists and large technology firms

invest huge sums in new startups.⁶ Its leading AI companies have research labs in the United

States⁷ and elsewhere.⁸ Its researchers produce a trove of respected AI papers that advance

the field.⁹ None of this would concern us from a national security perspective, except for the

fact that China is led by a single-party authoritarian regime that threatens American interests.

“China stands a reasonable

chance of overtaking the

United States as the leading

center of AI innovation in the

coming decade.”

China has moved more quickly and with more determination than the United States, guided

by a constellation of AI plans for government ministries, universities, and companies.¹⁰These

strategic documents reflect Beijing’s view that advances in AI will fundamentally reshape

military and economic competition in the coming decades.¹¹ China has backed up its

strategic plans with significant state subsidies to technology firms and academic institutions

that engage in cutting-edge AI research.¹² China preserves its capital by taking advantage of

basic research done by the West so that it can focus more on applications. It pours significant

sums of money into research and talent in relevant fields,¹³ and it promotes “national champion”

companies to win markets abroad.¹⁴ Through its military-civil fusion programs, China has

sought to integrate advances in AI from the commercial and academic worlds into military

power.¹⁵ Using espionage, technology transfer programs, and targeted investment, Beijing

seeks to acquire sensitive IP and data from the United States and other countries.¹⁶

161

A STRATEGY FOR COMPETITION AND COOPERATION

The U.S.-China competition is complicated by the complex web of supply chains, research

partnerships, and business relationships that link the world’s two AI leaders. Dramatic steps

to sever these ties could be costly for Americans and reverberate across the world. The

relationships between American and Chinese academics, innovators, and markets are

deep, often mutually beneficial, and help advance the field of AI.¹⁷ Moreover, it remains in

the U.S. national interest to leverage formal diplomatic dialogue about AI and other emerging

technologies and to explore areas for cooperative AI projects that will benefit humanity.

“The United States can

compete against China

without ending collaborative

AI research and severing all

technology commerce.”

The United States can compete against China without ending collaborative AI research and

severing all technology commerce. Broad-based technological decoupling with China could

deprive U.S. universities and companies of scarce AI and science, technology, engineering,

and mathematics (STEM) talent,¹⁸ sever American companies’ efficient supply chains,¹⁹ and

cut off access to markets and capital for innovative firms.²⁰ Instead, the United States should

conceive of targeted disentanglement as just one element of its overall approach, which, if

applied judiciously to key sectors, can help build U.S. technological resilience, reduce threats

from illicit technology transfer, and protect national security-critical sectors.

The Policy Challenges.

China’s competitive approach should not define the U.S. approach to innovation, but it does

present an alternative model of AI development, frame the stakes of competition, and expose the

sheer breadth of public policy choices the U.S. government must make to preserve American

advantages.

The United States will need to reexamine its immigration policies to ensure that America

wins the competition for AI talent. It will need to consider AI and broader STEM education

initiatives through the lens of global competitiveness. It will have to consider how to

diversify the AI research agenda and expand access to the data and tools necessary to

162

CHAPTER 9

conduct AI research in the face of costs for compute and data that are consolidating AI in

fewer locations and shifting the balance from universities to the private sector. The United

States will have to consider whether the long-standing approaches to IP are best suited to

an era in which IP theft is pervasive and the U.S. IP regime has not yet fully accounted for

AI and other emerging technologies. The United States will need to protect its leadership

in the design of microelectronics, which may include encouraging the domestic reshoring

of critical manufacturing on which our national security depends. And the United States

will have to ensure that its tools and policies designed to prevent illicit technology transfer

are postured to address the national security challenges presented by dual-use emerging

technologies.

These AI-specific challenges,

•

How to draw on the best global

talent without enabling damaging

in turn, expose even more

technology and knowledge transfer to

fundamental questions

competitors.

spanning the technology,

•

How to foster an open collaborative

economic, and national security

research environment while closing

spheres:

licit and illicit loopholes exploited by

foreign competitors.

• How to compete with a rival without

compromising U.S. values—including

•

How to sustain long-term strategies for

free-market principles, individual

R&D that are nevertheless responsive

liberty, and limited government.

to rapidly shifing geopolitical and

technology developments.

• How to ensure the proper balance

between defense and economic

•

How to ensure the free flow of

priorities.

investment/capital without allowing

strategic competitors to buy strategic

• How to preserve hardware advantages

advantage.

without suffocating the domestic

designers and producers that rely on

•

How to engage with our allies

foreign competitors’ markets.

and other partners to reduce their

dependence on China’s digital

• How to capitalize on and shape

technologies, build more resilient

private-sector developments for

supply chains, and develop

national security ends without stifling

technology standards and norms that

private sector-led and free-market

reflect democratic values.

innovation.

The Need for a Stronger Government Role in Technology Strategy.

The Commission is not calling for a state-directed economy, a five-year plan, or China-

style “military-civil fusion.” It is instead urging a government-led process to restore a

more balanced equilibrium between government, industry, and academia that ensures a

163

A STRATEGY FOR COMPETITION AND COOPERATION

diverse research environment, competitive economy, and the sustainment of a research

agenda that supports the needs of the nation. The government has a long history of

mobilizing industry and academia and making huge investments when the United States

is challenged.²¹ Against the backdrop of a declared and committed competitor like China,

and given AI’s transformative potential, the United States is confronting such a moment.

“... the U.S. government

champions AI leadership in

speeches and memorandums,

but deploys few resources

relative to commercial

investment and historic

funding benchmarks ...”

Today, the U.S. government champions AI leadership in speeches and memorandums,

but it deploys few resources relative to commercial investment and historic funding

benchmarks and relies on a decentralized governance structure for achieving it.²² There

is talk of a global talent competition, but in recent years the United States has tightened

restrictions on visas for highly skilled workers,²³ and U.S. students at the kindergarten to

12th grade (K-12) level have lagged behind East Asian and European competitors in exams

designed to measure competency in STEM fields.²⁴ Tech leaders and government officials

talk about the importance of “public-private partnership,” but there is little action in either

direction to deepen it in concrete ways. U.S. experts warn of the danger of AI being used

for techno-authoritarian ends,²⁵ but Washington has not led any new enduring coalition

to create democratic alternatives. Current policies amount to a compilation of disparate

AI-related activities underway in the federal government. Nowhere can one find a strategy

coupled with the organization and resources to win an AI competition and preserve the

United States’ AI leadership.

The government will have to orchestrate policies to promote innovation; protect industries

and sectors critical to national security; recruit and train talent; incentivize domestic

research, development, and production across a range of technologies deemed essential

for national security and economic prosperity; and marshal coalitions of allies and

164

CHAPTER 9

partners to support democratic norms. Some elements of a national strategy will need

to be coordinated and replicated at the state level, through state-specific strategies to

support AI research, commerce, and education. This will require a complex sequencing

of promotion and protection actions to minimize costs and risks of punitive actions; ensure

basic and applied research agendas are mutually reinforcing; coordinate approaches with

international partners; and align executive priorities with legislative powers. It will require

identifying technology trends and assessing the relative strengths of the United States and

its competitors. It will require, above all, strong and consistent White House leadership.

The Case for White House Leadership.

The government will require a center of power that can exert gravitational pull on domestic

economic, national security, and science and technology policies. We have no such

organization today. Several separate Executive Office of the President (EOP) entities

possess some responsibility and capacity to fulfill the basic organizational requirements:

the National Security Council (NSC),²⁶ the Office of Science and Technology Policy (OSTP)

²⁷ and its associated National Science and Technology Council (NSTC),²⁸ and the National

Economic Council (NEC).²⁹ The Domestic Policy Council (DPC) also has critical related

responsibilities and a similar mandate with leadership in the realm of immigration policy,

education policy, and regulatory policy.³⁰ An additional entity—the Office of Management

and Budget (OMB)—oversees related budgets and government reform efforts.

In the absence of an overarching structure, it is left to the President and Vice President

to identify, adjudicate, and reconcile the positions that emerge from parallel interagency

processes, while leaving endless room for gadflies to run the gaps and influence the

President. The President needs a tool to help decide and drive a new technology strategy

down through the necessary but not sufficient existing councils and into the rest of the

government. The White House should:

Technology

Competitive

Council.

165

A STRATEGY FOR COMPETITION AND COOPERATION

“The government will require

a center of power that can

exert gravitational pull on

domestic economic, national

security, and science and

technology policies. We have

no such organization today.”

Create a Technology Competitiveness Council. The United States must strengthen executive

Recommendation

leadership in technology policy in the White House by empowering a single entity to

implement a comprehensive technology strategy. The Commission proposes creating a new

Technology Competitiveness Council (TCC), which would include the same amalgamation

of EOP leaders and Cabinet secretaries as other White House forums for convening the

interagency, and be chaired by the Vice President with a newly appointed Assistant to

the President for Technology Competitiveness serving as the day-to-day leader. The TCC

would ensure that the gaps between NEC, OSTP, and NSC responsibilities are filled and

linked to OMB. It would not replace the NSC, NEC, or OSTP-led NSTC structures, but would

provide a forum for reconciling competing security, economic, and scientific priorities and

elevate technology policy and concerns from a technical to a strategic level. To coordinate

the council’s work, it is necessary to create a new principal, the Assistant to the President

for Technology Competitiveness, responsible for ensuring policies pertaining to emerging

technologies receive sufficient Presidential-level attention.

Develop a National Technology Strategy. The TCC should create a National Technology

Recommendation

Strategy, building on the elements we present here, which can guide U.S. policy across all

key emerging technologies starting with AI. The goal of the National Technology Strategy

should be to ensure long-term, overall U.S. leadership in technology, particularly emerging

technologies critical to national security and competitiveness. The strategy should weigh

the difficult tradeoffs between competing policy interests and priorities, identify critical

technologies where competitors have sought to match or overtake U.S. leadership, and

facilitate an integrated policy approach to emerging technologies. As a starting point,

the strategy should build on the following pillars: 1) winning the AI talent competition; 2)

promoting American AI innovation; 3) protecting U.S. AI advantages; and 4) leading a

favorable international AI order.

166

A STRATEGY FOR COMPETITION AND COOPERATION

Chapter 9 - Endnotes

¹ While the U.S. government has released a number of documents emphasizing the importance

of AI research and development—see, for example, President Trump’s executive order on AI—the

U.S. lacks a comprehensive, whole-of-government plan to guide policymakers, researchers, and

businesses toward a more secure U.S. future. Artificial Intelligence for the American People, The

White House (last accessed Jan. 28, 2021), https://trumpwhitehouse.archives.gov/ai/.

² See, e.g., Alexandra Mousavizadeh, et al., The Global AI Index, Tortoise Media (Dec. 3, 2019),

https://www.tortoisemedia.com/2019/12/03/global-ai-index/; Jean François Gagné, et al., Global AI

Talent Report 2020 (last accessed Dec. 29, 2020), https://jfgagne.ai/global-ai-talent-report-2020/;

The Global AI Talent Tracker, MacroPolo (last accessed Dec. 29, 2020), https://macropolo.org/

digital-projects/the-global-ai-talent-tracker/; Jeffrey Ding, et al., MERICS Web Seminar: China as an

AI Superpower? Quantifying China’s AI Progress Against the US and Europe, MERICS (July 1, 2020),

https://merics.org/en/video/merics-web-seminar-china-ai-superpower-quantifying-chinas-ai-progress-

against-us-and-europe.

³ Audrey Cher, ‘Superpower Marathon’: U.S. May Lead China in Tech Right Now—But Beijing Has the

Strength to Catch Up, CNBC (May 17, 2020), https://www.cnbc.com/2020/05/18/us-china-tech-race-

beijing-has-strength-to-catch-up-with-us-lead.html; Graham Allison & Eric Schmidt, Is China Beating

the U.S. to AI Supremacy?, Belfer Center for Science and International Affairs (Aug. 2020), https://

www.belfercenter.org/publication/china-beating-us-ai-supremacy; Will Knight, China May Overtake

the US with the Best AI Research in Just Two Years, MIT Technology Review (March 13, 2019),

https://www.technologyreview.com/2019/03/13/136642/china-may-overtake-the-us-with-the-best-ai-

research-in-just-two-years/.

⁴ Karen Hao, Three Charts Show How China’s AI Industry Is Propped Up by Three Companies, MIT

Technology Review (Jan. 22, 2019), https://www.technologyreview.com/2019/01/22/137760/the-future-

of-chinas-ai-industry-is-in-the-hands-of-just-three-companies/.

⁵ James Kynge & Nian Liu, From AI to Facial Recognition: How China Is Setting the Rules in New Tech,

Financial Times (Oct. 7, 2020), https://www.ft.com/content/188d86df-6e82-47eb-a134-2e1e45c777b6.

⁶ Yusho Chao, Chinese Venture Capitalists Take a Shine to Startups Again, Nikkei (Sept. 13, 2020),

https://asia.nikkei.com/Business/Finance/Chinese-venture-capitalists-take-a-shine-to-startups-again.

See also, Visualizing Chinese Tech Giants Billion-dollar Acquisitions, CB Insights (May 28, 2020),

https://www.cbinsights.com/research/bat-billion-dollar-acquisitions-infographic/.

⁷ See, e.g., A Chinese Tech Giant Is Setting Up an A.I. Research Lab on Amazon’s Home Turf, CNBC

(May 2, 2017), https://www.cnbc.com/2017/05/02/tencent-ai-research-lab-seattle.html.

⁸ See, e.g., Saheli Roy Choudhury, Alibaba Sets Up Joint A.I. Research Center Outside China to Focus

on AI, CNBC (Feb. 28, 2018), https://www.cnbc.com/2018/02/28/alibaba-sets-up-joint-a-i-research-

lab-in-singapore.html.

⁹ In 2019, China had the largest number of submitted and accepted papers to the Association for

the Advancement of AI (AAAI), one of the longest-running AI conferences. See Artificial Intelligence

Index: 2019 Annual Report, Stanford Institute for Human-Centered AI at 41 (2019), https://hai.stanford.

edu/sites/default/files/ai_index_2019_report.pdf. The Allen Institute for AI also predicts that China is

poised to overtake the U.S. in the share of top-cited, breakthrough papers in AI by 2025. See Field

Cady & Oren Etzioni, China May Overtake US in AI Research, Allen Institute for Artificial Intelligence

(March 13, 2019), https://medium.com/ai2-blog/china-to-overtake-us-in-ai-research-8b6b1fe30595.

¹⁰ For a selection of such strategic documents, see AI Policy—China, Future of Life Institute (last

accessed Dec. 30, 2020), https://futureoflife.org/ai-policy-china/; Graham Webster, et al., Full

Translation: China’s ‘New Generation Artificial Intelligence Development Plan,’ New America (Aug. 1,

2017), https://www.newamerica.org/cybersecurity-initiative/digichina/blog/full-translation-chinas-new-

generation-artificial-intelligence-development-plan-2017/ (translating China’s State Council Notice on

the Issuance of the New Generation Artificial Intelligence Development Plan, dated July 20, 2017).

¹¹ Gregory C. Allen, Understanding China’s AI Strategy, Center for a New American Security (Feb. 6,

2019), https://www.cnas.org/publications/reports/understanding-chinas-ai-strategy.

168

CHAPTER 9

Chapter 9 - Endnotes

¹² Ashwin Acharya & Zachary Arnold, Chinese Public AI R&D Spending: Provisional Findings, Center

for Security and Emerging Technology (Dec. 2019), https://cset.georgetown.edu/wp-content/uploads/

Chinese-Public-AI-RD-Spending-Provisional-Findings-1.pdf; see also Emily Weinstein, Mapping

China’s Sprawling Efforts to Recruit Scientists, Defense One (Nov. 30, 2020), https://www.defenseone.

com/ideas/2020/11/mapping-chinas-sprawling-efforts-recruit-scientists/170373/; David Cyranoski,

China Joins the Battle for AI Talent, Nature (Jan. 17, 2018), https://www.nature.com/articles/d41586-

018-00604-6.

¹³ Id.

¹⁴ U.S.-China Economic and Security Review Commission, Hearing on Technology, Trade, and

Military-Civil Fusion: China’s Pursuit of Artificial Intelligence, New Materials, and New Energy at 46,

115-116 (June 7, 2019), https://www.uscc.gov/sites/default/files/2019 -10/June%207,%202019%20

Hearing%20Transcript.pdf.

¹⁵ U.S.-China Economic and Security Review Commission, Hearing on Technology, Trade, and

Military-Civil Fusion: China’s Pursuit of Artificial Intelligence, New Materials, and New Energy (June

7, 2019), https://www.uscc.gov/sites/default/files/2019 -10/June%207,%202019%20Hearing%20

Transcript.pdf.

¹⁶ Id.

¹⁷ As Eric Schmidt noted in Building a New Technological Relationship and Rivalry. See Hal Brands &

Francis J. Gavin, COVID-19 and World Order: The Future of Conflict, Competition, and Cooperation,

Johns Hopkins University Press at 406-418 (Aug. 31, 2020), https://muse.jhu.edu/chapter/2696578.

¹⁸ Ishan Banerjee & Matt Sheehan, America’s Got AI Talent: US’ Big Lead in AI Research Is Built on

Importing Researchers, MacroPolo (June 9, 2020), https://macropolo.org/americas-got-ai-talent-us-

big-lead-in-ai-research-is-built-on-importing-researchers/?rp=m.

¹⁹ U.S. Dependence on China’s Rare Earth: Trade War Vulnerability, Reuters (June 27, 2019), https://

www.reuters.com/article/us-usa-trade-china-rareearth-explainer/u-s-dependence-on-chinas-rare-

earth-trade-war-vulnerability-idUSKCN1TS3AQ.

²⁰ Member Survey, US-China Business Council (Aug. 2019), https://www.uschina.org/sites/default/

files/member_survey_2019_-_en_0.pdf.

²¹ For instance, adjusted for inflation, the Manhattan Project cost an estimated $27 billion and the

Apollo program totaled roughly $121 billion. Deborah Stine, The Manhattan Project, the Apollo

Program, and Federal Energy Technology R&D Programs: A Comparative Analysis, Congressional

Research Service (June 30, 2009), https://fas.org/sgp/crs/misc/RL34645.pdf (conversion into 2020

dollars was calculated using the U.S. Bureau of Labor Statistics’ CPI Inflation Calculator, available at

https://www.bls.gov/data/inflation_calculator.htm ).

169

A STRATEGY FOR COMPETITION AND COOPERATION

²² In 2018, U.S. federal R&D funding amounted to 0.7% of GDP, down from its peak at above 2% in

the 1970s. See James Manyika & William H. McRaven, Innovation and National Security: Keeping

Our Edge, Council on Foreign Relations (Sept. 2019), https://www.cfr.org/report/keeping-our-edge/

recommendations/.

²³ Zolan Kanno-Youngs & Miriam Jordan, Trump Moves to Tighten Visa Access for High-Skilled Foreign

Workers, New York Times (Oct. 6, 2020), https://www.nytimes.com/2020/10/06/us/politics/h1b-visas-

foreign-workers-trump.html.

²⁴ Moriah Balingit & Andrew Van Dam, U.S. Students Continue to Lag Behind Peers in East Asia

and Europe in Reading, Math and Science, Exams Show, Washington Post (Dec. 3, 2019), https://

www.washingtonpost.com/local/education/us-students-continue-to-lag-behind-peers-in-east-asia-

and-europe-in-reading-math-and-science-exams-show/2019/12/02/e9e3b37c-153d-11ea-9110-

3b34ce1d92b1_story.html.

²⁵ Alina Polyakova & Chris Meserole, Exporting Digital Authoritarianism, Brookings (Aug. 2019), https://

www.brookings.edu/research/exporting-digital-authoritarianism/.

²⁶ The National Security Council has a statutory mandate to “advise the President with respect to the

integration of domestic, foreign, and military policies relating to the national security so as to enable

the Armed Forces and the other departments and agencies of the United States Government to

cooperate more effectively in matters involving the national security.” 50 U.S.C. § 3021(b)(1).

²⁷ See Pub. L. 94-282, National Science and Technology Policy, Organization, and Priorities Act of

1976, 90 Stat. 459 (1976), https://obamawhitehouse.archives.gov/sites/default/files/microsites/ostp/

ostp_organic_statute.pdf.

²⁸ The function of the NSTC under the supervision of the Director of OSTP is: “(1) to coordinate

the science and technology policy-making process; (2) to ensure science and technology policy

decisions and programs are consistent with the President’s stated goals; (3) to help integrate the

President’s science and technology policy agenda across the Federal Government; (4) to ensure

science and technology are considered in development and implementation of Federal policies and

programs; and (5) to further international cooperation in science and technology. The Assistant may

take such actions, including drafting a Charter, as may be necessary or appropriate to implement

such functions.” William J. Clinton, Executive Order 12881: Establishment of the National Science and

Technology Council (Nov. 23, 1993), https://www.govinfo.gov/content/pkg/ WCPD-1993-11-29/pdf/

WCPD-1993-11-29-Pg2450.pdf.

²⁹ William J. Clinton, Executive Order 12835: Establishment of the National Economic Council (Jan. 25,

1993), https://www.govinfo.gov/content/pkg/WCPD-1993-02-01/pdf/WCPD-1993-02-01-Pg95.pdf.

³⁰ William J. Clinton, Executive Order 12859: Establishment of the Domestic Policy Council (Aug. 16,

1993), https://www.archives.gov/files/federal-register/executive-orders/pdf/12859.pdf.

170

CHAPTER 10

The United States is in a global competition for scarce AI talent.¹

The Commission is very concerned with current talent trends. The

number of domestic-born students participating in AI doctorate

programs has not increased since 1990, and competition for

international students has accelerated, endangering the United

States’ ability to retain international students.² For the first time

in our lifetime, the United States risks losing the competition for

talent on the scientific frontiers. Cultivating more potential talent at

home and recruiting and retaining more existing talent from foreign

countries are the only two options to sustain the U.S. lead.

“For the first time in our lifetime,

the United States risks losing

the competition for talent on the

scientific frontiers.”

Competitors and allies recognize the importance of implementing AI talent strategies.

Between 2000 and 2014, China’s university system increased its number of science,

technology, engineering, and mathematics (STEM) graduates by 360%, producing 1.7

million in 2014 alone.³ The number of STEM graduates in the United States’ university

system rose by 54% during the same time period, and many were international students.⁴

China’s researchers now represent roughly 29% of top-tier deep learning talent in the

world.⁵ China and other states have also taken steps to attract international talent with

flexible immigration policies and incentives for tech talent.⁶

The United States needs to invest in all AI talent pipelines in order to remain at the forefront

of AI now and into the future. A passive strategy will not work in the face of the AI talent

competition.

173

THE TALENT COMPETITION

To achieve dominance in AI, the U.S. needs to train four archetypes to propel AI in

America: researchers, implementers, end users, and informed consumers.

Researchers

AI research engineers will focus on R&D of technologies that enable and

advance semi- and fully-autonomous systems. They serve as algorithm experts

with up-to-date knowledge of modern AI research and may be involved in the

inception of ideas and drive the development cycles from research to testing of

prototypes for a major project or component of a major project.

Implementers

They will be responsible for data cleaning, feature extraction and selection,

and analysis; model training and tuning; partnerships with domain knowledge

experts and end users; and the discovery of local opportunities for exploitation.

Developers require less training and education than AI experts, and will have

training, education, and/or experience that is roughly equivalent to an associate

or bachelor’s degree; and that includes relevant ethics and bias mitigation in

data processing and model training.

End Users

They will have their daily business augmented/enabled by AI. Use of AI will

strongly resemble the use of currently available software in that it will require

some system-specific training, but, with the exception of some positions that

manage data, little to no AI specific expertise.

Informed Consumers

This group of people needs the ability to make better consumer choices when

purchasing technology and understand the importance of their actions in the market.

The Promise and Limits of Expanding STEM.

Investments in STEM education are a necessary part of increasing American national

power and improving national security. The United States ranks well overall on international

measures of talent because of our ability to attract international talent, in spite of our

uneven kindergarten to 12th grade (K-12) education system.⁷ It is critical that the United

States invest significantly in STEM education as an engine to drive the growth of AI talent in

America. Investments in STEM education alone, however, will not be enough for the United

States to win the international competition for AI and STEM talent. China is producing

large numbers of computer scientists, engineers, and other STEM graduates.⁸ For the

foreseeable future, the United States’ STEM education system does not have the capacity

nor the quality to produce sufficient STEM or AI talent to supply the United States’ markets

or national security enterprise.⁹ To compete, the United States must reform its education

system to produce both a higher quality and quantity of graduates.

174

CHAPTER 10

Pass a National Defense Education Act II. Motivated by fear that America had fallen behind

Recommendation

in education and innovation after the Soviets launched Sputnik in 1957, Congress passed

the National Defense Education Act (NDEA) the following year. The NDEA promoted the

importance of science, mathematics, and foreign languages for students, authorizing more

than $1 billion toward decreasing student loans, funding for education at all levels, and

funding for graduate fellowships. Many students were able to attend college because of

this legislation. In 1960, 3.6 million students attended college; by 1970 it was 7.5 million.¹⁰

This act helped America win the Space Race, helped power the microelectronics industry,

accelerated the U.S. capacity to innovate, and, ultimately, played an important role in

America’s victory in the Cold War.

The Commission believes the time is right for a second NDEA, one that mirrors the first

legislation, but with important distinctions. NDEA II would focus on funding students

acquiring digital skills, like mathematics, computer science, information science,

data science, and statistics. NDEA II should include K-12 education and reskilling

programs that address deficiencies across the spectrum of the American educational

system, purposefully targeting under-resourced school districts. The Commission also

recommends investments in university-level STEM programs with 25,000 undergraduate,

5,000 graduate, and 500 PhD-level scholarships. Undergraduate scholarships should

include credit hours at community colleges to ensure more Americans have access to

affordable STEM education. Ultimately, the goal of NDEA II is to widen the digital talent

pool by incentivizing programs for underrepresented Americans.

“The Commission believes the

time is right for a second NDEA ...”

Strengthen AI talent through immigration. Immigration reform is a national security imperative.

Recommendation

Nations that can successfully attract and retain highly skilled individuals gain strategic

and economic advantages over competitors. Human capital advantages are particularly

significant in the field of AI, where demand for talent far exceeds supply.¹¹ Highly skilled

immigrants accelerate American innovation, improve entrepreneurship, and create jobs.¹²

The United States benefits far more from the immigration of highly skilled foreign workers

than other countries. In 2013, the United States had 15 times as many immigrant inventors

as there were American inventors living abroad.¹³ By contrast, Canada, Germany, and

the U.K. all maintain a net negative inventor immigration rate.¹⁴ Compared with other U.S.

advantages in the AI competition—such as financial resources or hardware capacity—this

immigration advantage is harder for other countries to replicate.

175

THE TALENT COMPETITION

“Nations that can successfully

attract and retain highly

skilled individuals gain strategic

and economic advantages

over competitors.”

Unfortunately, international students in the United States are increasingly choosing to

study in other countries or return home.¹⁵ One reason is the growing backlog of green card

petitions.¹⁶ Indian immigrants face a particularly long wait. Many will spend decades on

constrictive work visas waiting to receive their green cards, hindering both the technology

sector’s ability to recruit talent and Indian immigrants’ quality of life. At the same time, other

countries are increasing their efforts to attract and retain AI talent, including immigrants in

Silicon Valley.¹⁷

“Restrictions harm U.S. innovation

and economic growth and only

help our competitors by enabling

their human capital to grow. They

also incentivize U.S. technology

companies to move to where

talent resides, whether right

across our borders or overseas.”

176

CHAPTER 10

While immigration benefits the United States, policymakers must also bear in mind the

threat of unwanted technology transfer. However, restricting immigration is far too blunt a

tool to solve this problem.¹⁸ Restrictions harm U.S. innovation and economic growth and

only help our competitors by enabling their human capital to grow. They also incentivize

U.S. technology companies to move to where talent resides, whether right across our

borders or overseas.¹⁹ Technology transfer will only get worse if significant components

of the U.S. technology sector move their research and development to China or other

countries that are more vulnerable than the United States to technology transfer efforts.²⁰

A more effective strategic approach would pair actions to improve the United States’ ability

to attract top global talent with targeted efforts to combat technology transfer vectors.

NSCAI addresses technology transfer in detail in Chapter 14 of this report. Changes to

immigration policies should be paired with those recommendations.

Immigration policy can also slow China’s progress. China’s government takes the threat of

brain drain seriously, noting that the United States’ ability to attract and retain China’s talent

is an obstacle to the Chinese Communist Party’s (CCP) ambitions.²¹ Increasing China’s

brain drain will create a dilemma for the CCP—which will be forced to choose between

losing even more human capital, further slowing their economic growth and threatening

their advancement in AI, or denying Chinese nationals opportunities to study and work in

the United States. At the same time, the United States should be cautious about potential

adverse effects on talent pools in partner nations.

“Increasing China’s brain drain will

create a dilemma for the CCP ...”

Recommendation

Broaden the scope of “extraordinary” talent to make the O-1 visa more accessible and

emphasize AI talent. The O-1 temporary worker visa is for people with extraordinary ability

or achievement. Currently adjudicators determine an applicant’s eligibility through a

subjective assessment. For the sciences and technology, this aligns largely with academic

criteria such as publications in major outlets and is not well suited for people who excel in

industry.

Recommendation

Implement and advertise the international entrepreneur rule. The International Entrepreneur

Rule (IER) allows U.S. Citizenship and Immigration Services (USCIS) to grant a period

of authorized stay to international entrepreneurs who demonstrate that “their stay in the

United States would provide a significant public benefit through their business venture.”²²

An executive action could announce the administration’s intention to use the IER to boost

immigrant entrepreneurship, job creation for Americans, and economic growth. USCIS

177

THE TALENT COMPETITION

could also be directed to announce that it will give priority to entrepreneurs active in

high-priority STEM fields such as AI, or in fields that use AI for other applications, such

as agriculture. Entrepreneurs’ ability to attract investors should be used as a screening

criterion for entrepreneurs.

Expand and clarify job portability for highly skilled workers. The criteria for workers with H-1B,

Recommendation

O-1, and other temporary work visas to obtain open market work permits for a one-year

renewable period are too limited and ambiguous. Changes should clarify when highly

skilled, nonimmigrant workers are permitted to change jobs or employers, increase job

flexibility when an employer either withdraws their petition or goes out of business, and

increase flexibility for H-1B workers seeking other H-1B employment.

Recapture green cards lost to bureaucratic error. Federal agencies generally issue fewer green

Recommendation

cards than they are allowed. As of 2009, the federal government had failed to issue more

than 326,000 green cards based on cumulative bureaucratic error.²³ The Departments of

State and Homeland Security (DHS) should publish an up-to-date report on the number of

green cards lost due to bureaucratic error. Using available authorities, both should grant

lost green cards to applicants waiting in line. Congress should support the Departments of

State and Homeland Security by passing legislation to recapture lost green cards.²⁴

Grant green cards to students graduating with STEM PhDs from accredited American universities.

Recommendation

Congress should amend the Immigration and Nationality Act²⁵ to grant lawful permanent

residence to any vetted (not posing a national security risk) foreign national who graduates

from an accredited United States institution of higher education with a doctoral degree in

a STEM-related field in a residential or mixed residential and distance program and has

a job offer in a field related to science, technology, engineering, or mathematics. They

should not be counted toward permanent residency caps.

Double the number of employment-based green cards. Under the current system,

Recommendation

employment-based green cards are unduly scarce: 140,000 per year, fewer than half of

which go to the principal worker.²⁶ This leaves many highly skilled workers unable to gain

permanent residency and unable to transfer jobs or negotiate with employers as effectively

as domestic workers. This decreases the appeal of joining the American workforce. To

reduce the backlog of highly skilled workers, the United States should double the number

of employment-based green cards, with an emphasis on permanent residency for STEM

and AI-related fields.

Create an entrepreneur visa. International doctoral students are more likely than their native

Recommendation

peers to want to found a company or become an employee at a startup, but they are less

likely to pursue those paths.²⁷ One reason is the constraints of the H-1B visa system.²⁸

Similarly, immigrant entrepreneurs without the capital to use the EB-5 route to permanent

residency are forced to use other visas that are designed for academics and workers

in existing companies, not entrepreneurs.²⁹ All of these issues make the United States

178

CHAPTER 10

less attractive for international talent, and, perhaps as important, reduce the ability of

startups and other small companies—the main source of new jobs for Americans—to hire

highly skilled immigrants, who have been shown to improve the odds that the business

will succeed. Congress should create an entrepreneur visa for those who would provide

a “significant public benefit” to the United States if allowed to stay in the country for a

limited trial period to grow their companies.³⁰ This visa should serve as an alternative

to employee-sponsored, investor, or student visas and should instead target promising

potential founders.

Create an emerging and disruptive technology visa. The National Science Foundation (NSF)

Recommendation

should identify critical emerging technologies every three years. DHS would then allow

students, researchers, entrepreneurs, and technologists in applicable fields to apply for

emerging and disruptive technology visas. This would provide much-needed talent R&D

and strengthen our economy.³¹

179

THE TALENT COMPETITION

Chapter 10 - Endnotes

¹ Estimates on the gap of talent necessary to fill AI slots vary greatly, but it is agreed upon that the

gap in talent currently is and will continue to be significant as nations compete for scarce resources.

See Remco Zwetsloot, et al., Strengthening the U.S. AI Workforce: A Policy and Research Agenda,

Center for Security and Emerging Technology at 2 (Sept. 2019), https://cset.georgetown.edu/wp-

content/uploads/CSET-Strengthening-the-U.S.-AI-Workforce.pdf (“The Research Institute at Tencent,

a major Chinese technology company, asserts there are roughly 300,000 AI researchers and

practitioners worldwide, with market demand for millions of roles. Element AI, a leading Canadian

AI company, estimated in 2018 that there are roughly 22,000 PhD-educated researchers globally

who are able to work on AI research, with only about 25 percent of those ‘well-versed enough in the

technology to work with teams to take it from research to application.’ AI firm Diffbot estimates that

there are over 700,000 people skilled in machine learning worldwide.”).

² Remco Zwetsloot, et al., Keeping Top AI Talent in the United States, Center for Security and

Emerging Technology at iii-vi (Dec. 2019), https://cset.georgetown.edu/wp-content/uploads/Keeping-

Top-AI-Talent-in-the-United-States.pdf.

³ The Rise of China in Science and Engineering, NSF National Science Board (2018), https://www.nsf.

gov/nsb/sei/one-pagers/China-2018.pdf (China also passed the United States in the global share of

peer-reviewed S&E articles).

⁴ Science & Engineering Indicators 2018, NSF National Science Board (2018), https://www.nsf.gov/

statistics/2018/nsb20181/assets/561/higher-education-in-science-and-engineering.pdf.

⁵ For these purposes “top tier” talent was defined by accepted papers at the prestigious AI deep

learning conference Neural Information Processing Systems in 2019. This reflected approximately the

top 20% of researchers in the field. The Global AI Talent Tracker, MacroPolo (last accessed Dec. 28,

2020), https://macropolo.org/digital-projects/the-global-ai-talent-tracker/. China has placed a strong

emphasis on deep learning, just one of the important components of AI.

⁶ For example, China’s Thousand Talents Plan is part of a state-organized blueprint to be a global

leader in science and technology by 2050. Staff Report, Threats to the U.S. Research Enterprise:

China’s Talent Recruitment Plans, U.S. Senate Permanent Subcommittee on Investigations at 14 (Nov.

2019), https://www.hsgac.senate.gov/imo/media/doc/2019-11-18%20PSI%20Staff%20Report%20-%20

China’s%20Talent%20Recruitment%20Plans.pdf.

⁷ The Global AI Talent Tracker, MacroPolo (last accessed Dec. 28, 2020), https://macropolo.org/

digital-projects/the-global-ai-talent-tracker/. See also Gordon Hanson & Matthew Slaughter, High-

Skilled Immigration and the Rise of STEM Occupations in U.S. Employment, National Bureau of

Economic Research at 1 (Sept. 2016), https://www.nber.org/system/files/working_papers/w22623/

w22623.pdf.

⁸ The Rise of China in Science and Engineering, NSF National Science Board (2018), https://www.nsf.

gov/nsb/sei/one-pagers/China-2018.pdf.

⁹ As noted in Chapter 6 of this report, there were 433,116 open computer science jobs in the United

States in 2019, while only 71,226 new computer scientists graduated from American universities in

2019. Code.org (last accessed Jan. 11, 2021), https://code.org/promote. See also Oren Etzioni, What

Trump’s Executive Order on AI Is Missing: America Needs a Special Visa Program Aimed at Attracting

More AI Experts and Specialists, Wired (Feb. 13, 2019), https://www.wired.com/story/what-trumps-

executive-order-on-ai-is-missing/.

¹⁰ Sputnik Spurs Passage of the National Defense Education Act, U.S. Senate (last accessed Dec. 28,

2020), https://www.senate.gov/artandhistory/history/minute/Sputnik_Spurs_Passage_of_National_

Defense_Education_Act.htm#:~:text=The%20National%20Defense%20Education%20Act%20of%20

1958%20became%20one%20of,and%20private%20colleges%20and%20universities.

¹¹ According to one report, job listings for AI on one popular job website “increased more than five-

fold between 2015 and 2017 and demand for ‘deep learning’ skills increased by a factor of more than

30,” while the number of people looking for jobs in the field grew much more slowly. This mismatch is

slowing the adoption of AI. Most firms report that skills gaps are one of the top obstacles preventing

them from adopting AI. Remco Zwetsloot, et al., Strengthening the U.S. AI Workforce: A Policy and

Research Agenda, Center for Security and Emerging Technology at 1 (Sept. 2019), https://cset.

georgetown.edu/wp-content/uploads/CSET-Strengthening-the-U.S.-AI-Workforce.pdf.

180

National Security Commission on Artificial Intelligence. Final Report - page 3