By Robert O. Work
We find ourselves in the midst of a technological tsunami that is inexorably reshaping all aspects of our lives. Whether it be in agriculture, finance, commerce, health care, or diplomatic and military activities, rapid technological advancements in fields like advanced computing, quantum science, AI, synthetic biology, 5G, miniaturization, and additive manufacturing are changing the old ways of doing business. And AI—the technologies that simulate intelligent behavior in machines—will perhaps have the most wide-ranging impact of them all.
This judgment is shared by many countries. China, Russia, members of the European Union, Japan, and South Korea all are increasing AI research, development, and training. China in particular sees advances in AI as a key means to surpass the United States in both economic and military power. China has stated its intent to be the world leader in AI by 2030 and is making major investments to achieve that goal.
The United States needs to respond to this technological challenge in the same way it responded to prior technology competitions, such as the space race. U.S. leadership in AI is critical not only because technology is a key enabler of political, economic, and military power, but also because the United States can shape how AI is used around the world. As this report explains, while AI can be used for incredible good by societies, it already is being abused by authoritarian states to surveil and repress their populations. And advances in AI technology are enabling future malign uses, such as launching sophisticated influence attacks against democratic nations. The United States must make sure it leads in AI technologies and shapes global norms for usage in ways that are consistent with democratic values and respect for human rights.
This CNAS report offers sensible ways to ensure U.S. leadership in the coming “AI century.” If you are seeking a primer on AI or a long argument about its import, look elsewhere.1 This report has a bias for action. It is built around a list of concrete recommendations, segmented into seven sections, with concise rationale and explanation for each. Together, the recommendations provide the framework for a national strategy for AI leadership.
U.S. leaders would do well to read, consider, and implement them.
Executive Summary
The United States excels when it pursues big ideas. It is one of the few countries in the world that can rally its resources and its human capital to achieve the most ambitious of goals. The United States stands at the cusp of another such moment. Prudent policy decisions today will help to protect and cement America’s lead in AI for decades. Together these actions will help to ensure that the coming AI century is an American one, a new technological era where America’s national security—and that of U.S. allies and partners—is more secure, its economy is poised to flourish, and its norms and values underpin AI technologies worldwide.
American leaders in government and industry should commit to harnessing the country’s science and technology base through greater research and development (R&D) funding and by pursuing international collaboration with like-minded partners. These leaders also must dedicate resources to prepare future generations by building up America’s human capital. Educators will require new skills to teach the future American work force as part of updated K-12 curricula.
Human capital investments must go beyond those born in the United States. International talent is a cornerstone of American innovation. Immigrants and their children play an outsize role in the U.S. technology ecosystem, producing many of the United States’ leading scientific minds and founding many of the country’s most iconic companies. Congress and the executive branch must address outdated and constricting immigration laws to continue to encourage the world’s best AI talent to study, work, and stay in the United States.
America’s openness and opportunities are among its greatest attributes. Malign actors also use these qualities against it. Illicit technology transfer is a serious problem that erodes U.S. competitiveness and costs the U.S. economy hundreds of billions of dollars. Countering this widespread theft is imperative, as is the need to do so in a manner consistent with U.S. freedoms and values.
The United States must protect its technological competitiveness further by controlling exports and securing and diversifying supply chains of advanced AI-specific hardware. U.S. action also is needed to promote government readiness to assure that the country is prepared for the likely transformative impact of AI on U.S. national security, the U.S. economy, and American society. Finally, U.S. leadership in setting global AI norms, standards, and measurement is essential to promote AI ethics, safety, security, and transparency in accordance with U.S. interests.
The specific recommendations that flow from our analysis of these issues are detailed in this report. Together they form the strategic foundation needed to make the current U.S. vision for AI, and continued American leadership in AI, a reality.
Summary of Policy Recommendations
Research and Development
Boost yearly U.S. government funding of AI R&D to $25 billion by FY2025
Spending at this level is realistic and doable: $25 billion is equal to less than 19 percent of total federal R&D spending in the FY20 budget.
Basic research funding, which is foundational to game-changing technological achievements is under pressure; the U.S. government is the largest funder.
Incentivize private sector AI R&D with tax credits and easing access to government datasets
America’s corporations are a key comparative advantage in sustaining overall leadership in AI.
Data scarcity is a common barrier to entry for AI researchers at universities and startups.
Promote international R&D collaboration
Decades of experience show that joint work with foreign researchers can be done with great benefit and little detriment to U.S. economic and national security.
Human Talent
Increase public and private sector AI and science, technology, engineering, and math (STEM) education and skills training
To remain competitive, the United States needs a national human capital strategy for technology.
Increase funding opportunities for university researchers
Federal grants to academia decreased from their 2011 peak of $45.5 billion to $40.9 billion in 2017.
Raise the cap for H1-B visas; remove the cap for advanced-degree holders entirely
International talent remains a critical backbone of the country’s technological ecosystem.
U.S. technology firms currently rely heavily on temporary-hire foreign workers to fulfill critical shortages in STEM occupations.
Amend the Department of Labor Schedule A occupations list to include high-skilled AI technologists
Updating the Schedule A occupations list to include high-skilled AI technologists would streamline the permanent residency sponsorship process for employers.
Create a new program that couples visa grants to ten-year open-market work commitments
This approach would attract foreign students already highly predisposed to remain in the United States, target specific AI-related disciplines, and eliminate the cost and uncertainty of extending job offers to qualified foreign nationals by removing employer sponsorship requirements.
Illicit Technology Transfer
Provide more cyber defense support to small firms
Small and medium businesses in general are more vulnerable to cyber attacks.
Authorize consular officials to act on risk indicators for espionage to screen out high-risk individuals before they arrive
Broader screening is required beyond simply PLA-sponsored individuals or sensitive research projects because AI is highly dual-use.
Improve collaboration between U.S. counterintelligence experts and universities
Associations representing U.S. universities have expressed desire for better engagement with the national security community on counterintelligence threats, among other issues.
AI Hardware
Increase the availability of affordable compute resources
The high cost and limited availability of compute is often a barrier to entry for startups and researchers in academia.
Establish multilateral export controls on semiconductor manufacturing equipment (SME)
Limiting the diffusion of SME to China, which looks to indigenize advanced semiconductor fabrication, is essential to protecting America’s edge in semiconductors.
Multinational cooperation is necessary as other SME leaders are located in Japan, South Korea, Singapore, and the Netherlands.
Boost domestic semiconductor manufacturing with retooling incentives
Some fabs need to retool every 2-3 years to stay competitive and these costs are burdensome.
Secure semiconductor supply chains through public-private partnerships
The U.S. military and intelligence community have special needs for security that go above and beyond what is available in commercial facilities, yet they lack the scale of demand to make a purely government-dedicated foundry possible.
Diversify semiconductor fabrication by creating an international fab consortium with key allies
A consortium with allies should share the cost burden of building new semiconductor foundries to ensure a trusted and diverse supply chain.
AI Norms
Lead in establishing norms for appropriate AI use
The United States has unparalleled influence and authority on the global stage and is in a unique position to set an example for the world on how AI should and should not be used.
Collaborate with allies and partners on norms for AI use
Alliances and partnerships with like-minded countries will help to ensure that responsible stewardship of AI becomes the global norm.
Protect U.S. research from supporting human rights violations by modernizing export controls
Due to the dual-use nature of facial recognition and other biometrics-detection technology, U.S. organizations are at risk of indirectly contributing to human rights violations through research collaborations, technology exports, and investments.
Government Readiness
Prioritize talent management with hiring reforms and AI-related training
Talent management in an era of AI will require attracting and retaining top talent with technical AI expertise. Government officials require training to responsibly and effectively use AI applications, and to craft policy and inform acquisitions.
Allocate funding for federal agencies to implement AI
The American AI Initiative falls short in that it does not establish budget targets.
Increases in spending beyond what existing budgets can support is needed to ensure that agencies don’t lag on AI implementation.
Modernize IT processes
Consolidating data centers and standing up cloud services will be required to ensure the U.S. government can update systems; manage, leverage, and share data; have access to compute; and manage a technical work force.
Define what is AI
Setting a definition for AI for federal government purposes will improve standards setting, formulating measurements, appropriations of AI-related funding, and tracking AI spending across government.
AI Standards & Measurement
Establish an NSTC Subcommittee for AI Standards & Measurement
The central importance of standards and measurement to fostering AI technologies that are safe, secure, reliable, and comport with U.S. norms and values warrant a stand-alone subgroup dedicated to the issue.
Establish a permanent horizon scanning effort devoted to AI
Tracking global progress in AI is necessary to hedge against technology surprise.
Introduction
AI, the technologies that simulate intelligent behavior in machines, likely will create dramatic opportunities for the U.S. economy, national security, and our health and well-being. Unfortunately, without policy action, the United States is at risk of ceding its leadership in AI. China, European Union member states, Japan, South Korea, and Russia are increasing spending on AI R&D and training new researchers to leverage AI. Some—Russia, most notably—seek to develop autonomous robotic weapons to replace human soldiers on the battlefield. Most of these countries do not just have AI strategies but have begun implementation in a way that threatens America’s technological edge. This report recommends concrete actions to ensure that the United States remains the leader in AI, to promote the development of standards in line with U.S. interests and values, and to anticipate and prepare for security challenges.
These policy recommendations build on existing U.S. AI policy, including the Executive Order on Maintaining American Leadership in Artificial Intelligence (February 2019) and The National Artificial Intelligence R&D Strategic Plan: 2019 Update (June 2019). The executive order highlights the need for using AI to promote economic and national security, fostering technological breakthroughs, development of technical standards, skills training for workers and researchers, international collaboration, and investments in R&D. The 2019 AI R&D plan lays out eight high-level strategic concepts that identify federal R&D priorities. Both documents identify U.S. strengths, challenges, opportunities, and priorities.
The legislative branch also contributes to laying out a high-level vision. The U.S. Congress mandated the establishment of the National Security Commission on Artificial Intelligence (NSCAI) in the 2019 National Defense Authorization Act (NDAA), recognizing the need for a comprehensive national approach to AI. The commission was created as an independent body to “review advances in artificial intelligence, related machine learning developments, and associated technologies” in order to “comprehensively address the national security and defense needs of the United States.”2 NSCAI released its interim report on November 4, 2019, identifying five lines of effort on which the U.S. government should focus: R&D investments, national security applications of AI, training and recruiting AI talent, protecting and building upon U.S. technical advantages, and promoting global AI cooperation. Underpinning these areas of focus is the clear-eyed recognition that the United States is in a strategic competition, and that AI is at its center.
This report presents specific next steps to ensure the U.S. government’s vision for sustained AI leadership becomes reality. It offers policy recommendations to strengthen U.S. competitiveness in AI through increased R&D spending and international collaboration; building American human capital through education and immigration reforms; tackling academic and industrial espionage; improving U.S. government readiness for widespread AI adoption; investing in AI-specific hardware at home and restricting exports of semiconductor manufacturing equipment; promoting the development of safe, transparent, explainable, reliable, secure, and resilient AI; and keeping abreast of global AI developments to prevent technology surprise.
The report is divided into eights sections that feature analysis of the issue at hand and provide corresponding policy recommendations. An appendix provides an overview of ongoing AI initiatives across the U.S. government.
Create an American AI Future
The United States is approaching a consequential decision point in its history: how to address AI’s looming impact on the American economy and society writ large. AI technologies will fundamentally alter how people interact with machines, devices, and each other. AI will change how we learn, how we work, how we treat illness, how we wage war. AI promises great opportunities. It also poses tremendous risks. How the United States prepares for this new era will determine whether American society can reap the benefits while mitigating the threats.
The path of least resistance is that of the status quo. It is the laissez-faire approach that is popularly thought to have served America well. In that narrative, the United States became the wealthiest and most powerful country in the world based on a unique mix of freedom, ingenuity, hard work, optimism, and a bit of luck. There is a certain comfort in this American exceptionalism. The current U.S. vision for AI as expressed by the Trump administration reflects this sentiment with its faith in industry, innovative spirit, and its hands-off approach to guidance and funding.
This path is fraught with risk, however, because it takes for granted U.S. leadership by underplaying the effort, expense, and vision it took to get here. American technological leadership is not guaranteed. The United States of today is rooted in investments in education, science, R&D, and infrastructure made decades ago. On its current trajectory, with a shrinking share of global R&D spending, human capital shortfalls, and the rapid rise of a near-peer competitor, the United States cannot continue to coast. America’s ability to harness AI to the fullest extent possible is at stake. Falling short would squander economic and societal benefits and expose the United States to avoidable risks and challenges.
Instead, bold action is needed to reinvigorate America’s competitiveness and position the United States for success. That framework for action is detailed in the pages that follow. This pathway includes decisions that will require tradeoffs and political fortitude. Success also will require long-term commitments from policymakers, academia, and private industry. Achieving the American AI Century depends on a whole-of-society approach spanning decades.
Lead in Research & Development
Greater investments in AI R&D are essential to maintaining American leadership in AI. Throughout the 20th century, the federal government played a critical role in fueling technological innovation by funding pivotal basic research. Government funding was essential to developing the transistor, the Global Positioning System, and the Internet—inventions that transformed the world economy. Yet over the past several decades, federal government spending on R&D as a percentage of GDP declined from about 1.2 percent in 1976 to around 0.7 percent in 2018.3 This is a worrisome trend as the federal government remains the main funder of basic research.4 Government support again could be pivotal both in fostering new AI breakthroughs and ensuring that the U.S. government has access to those breakthroughs.
Government funding was essential to developing the transistor, the Global Positioning System, and the Internet—inventions that transformed the world economy
U.S. government entities already are pursuing important AI R&D initiatives. The National Science Foundation funds an array of basic research and partners with stakeholders across government, academia, and the private sector to foster advances in the field.5 The National Institutes of Health are incorporating deep learning to improve disease screening and natural language processing for information retrieval and discovery.6 In 2018, the Defense Advanced Research Projects Agency (DARPA), the Department of Defense organization charged with developing emerging technologies, launched a $2 billion multi-year campaign to incentivize the creation of a range of new AI capabilities and applications.7
Unclassified federal government spending on defense AI R&D in FY2020 will be about $4 billion, according to a Bloomberg analysis from March 2019.8 In September 2019, the White House announced an FY20 non-defense AI R&D budget request of nearly $1 billion.9
In contrast, the level of Chinese government spending on AI R&D is not clear. Complete annualized figures for Chinese government spending are not publicly available. Instead, only announcements of planned, multi-year spending offer a window into the scale of overall government R&D spending at the national, provincial, and local levels.10 For instance, two Chinese cities alone announced the creation of RMB 100 billion (approximately $15 billion) multi-year AI development funds while Beijing unveiled plans for a $2 billion AI research park in 2018.11
The United States enjoys robust private sector R&D funding. Precise figures are hard to discern because companies typically do not divulge details for R&D expenditures in their financial statements and privately owned firms do not have such reporting requirements. That said, looking at overall R&D expenditures by major AI-intensive companies gives a sense of the scale of private investments in AI R&D. The combined 2018 R&D expenditures by U.S. firms Alphabet, IBM, Facebook, Microsoft, and Amazon was $80.5 billion.12
China’s tech giants also report significant R&D investments, although they are considerably smaller than those of their U.S. counterparts. Leading Chinese AI firms Alibaba, Baidu, and Tencent collectively spent $9.1 billion on R&D in 2018.13 These firms are also major investors in Chinese AI startups.14
Min Wanli, Alibaba’s Chief AI Scientist, speaks with CNBC’s Arjun Kharpal on a panel at CNBC’s East Tech West conference. While the U.S. is still the leader in private sector R&D investment, Chinese firms, such as Alibaba, Baidu, and Tencent, are rapidly increasing their investments. (Dave Zhong/Stringer/Getty Images)
In contrast, Europe is a laggard. Combined R&D spending by the EU (national governments and private investments) is projected to be EUR 20 billion (approximately $22.1 billion) in 2020, up from about EUR 3.7 billion (approximately $4.1 billion) in 2016.15
The United States’ dominant position in startup funding, a key driver of technological innovation, is starting to erode. In 2017, the U.S. share of global AI startup funding was less than half of the world’s total—ceding the lead to China—for the first time ever.16 This happened despite venture capital funding of American AI startups growing at a 36 percent compound annual growth rate since 2013.17
Figures are clearer for all national R&D spending, beyond simply AI, and indicate worrisome trends. Other countries are outpacing the United States with faster growth of their national R&D budgets. Total U.S. national (public and private) R&D expenditures as a share of GDP have been mostly stagnant since 1996. China quadrupled its R&D expenses as a share of GDP over the same time frame, and countries like Israel and South Korea also significantly ramped up spending.18 As a result, the U.S. share of global R&D has declined over the past several decades, falling from 69 percent in 1960 to 28 percent in 2016. From 2000 to 2015, the United States accounted for 19 percent of global R&D growth, while China accounted for 31 percent.19 China is on track to top the United States in total R&D investments (in purchasing power parity-adjusted dollars) as soon as 2019.20
R&D is a key driver of long-term economic growth.21 The Congressional Budget Office reaffirmed in 2018 that federal R&D spending is a positive influence on private R&D spending and increases macroeconomic growth.22 Authors of a ten-year study of 28 EU economies concluded that a 1 percent increase in R&D expenditure as a percentage of GDP would cause an increase of real GDP growth rate of 2.2 percent.23
The United States gains further benefits from federal R&D spending through effective technology diffusion. Technology transfers from the public to the private sector are stipulated in several laws.24 This legislation gives ownership and title to federally funded research by universities and small businesses and has resulted in thousands of spin-off companies, increased technology transfer, and greater innovation.25 Under these laws, the U.S. receives government royalty-free access to the research.26
To strengthen U.S. competitiveness in AI, Congress and the White House should:
Boost yearly U.S. government funding of AI R&D to $25 billion by FY2025
Congress and the White House should work together to increase federal AI R&D spending to $25 billion in five years. This target represents a fivefold increase over FY2020 but is affordable. It would still represent less than 19 percent of the amount requested for all unclassified R&D in the president’s FY20 budget. A large jump in spending on a specific line item also has precedent: The president’s FY19 budget requested an $18.1 billion increase in defense R&D over FY18.27
Given the central role AI technologies likely will play in economic growth, geopolitics, and global security, and the sharp growth in global spending on AI, this is a modest sum in relative terms. In FY2020, the United States is poised to spend nearly $59 billion in unclassified defense R&D alone (including AI R&D).28 For historical perspective, the five-year Manhattan Project cost $23 billion in 2018 dollars.29 The 1960–1973 Apollo program cost $288.1 billion when adjusted for inflation, and NASA spent $490 billion in total over those 13 years, an average of $37.7 billion a year.30
Leading up to the Apollo 11 Saturn V launch on July 16, 1969, the U.S. government invested billions in R&D funding during the 13-year Apollo program, demonstrating the power of targeted government R&D investment for breakthroughs in technology advancement. The U.S. government will need to similarly increase its R&D investment to spur new breakthroughs in artificial intelligence. (NASA/Getty Images)
The priority should be to fund high-risk/high-reward basic science research—areas where private industry has little incentive to invest but that hold tremendous potential for valuable new knowledge. Breakthroughs in software, such as novel AI techniques that address the limitations of existing AI methods, and hardware, such as next-generation semiconductor technologies and superconducting artificial neurons, could be game-changers that provide the United States with a continuing technological edge.31
The federal government should adopt a phased approach to increasing funding levels, so that the resources are spent effectively and responsibly. The departments and agencies that receive federal R&D monies (primarily DOD, HHS, DOE, NASA, NSF, USDA, VA, DOT, DOI, DHS, EPA) will require time to plan for expanded research agendas and to formulate relevant metrics to measure progress and effectiveness.
Incentivize AI R&D in the private sector
America’s private sector is a key comparative advantage in sustaining overall AI leadership by the United States. Policymakers have a number of ways to stimulate further R&D activity by corporations while adhering to free market principles. First and foremost is maintaining the PATH Act of 2015, which permanently extended the federal R&D tax credit.32 It offers strong incentives to conduct and expand R&D by reducing tax liabilities.
Second is standardizing and making current and future government datasets more readily available to the private sector and academia to facilitate training of machine learning models, as the Trump administration’s AI executive order proposes. The government’s Project Open Data is a major step in making data discoverable and usable.33 Doing so will help to address data scarcity problems, especially for entities with significant resource constraints such as startups or some university researchers, by expanding the number of open-source high-quality datasets.34
Third is exploring additional stimulants of private sector R&D activity. A comprehensive survey of R&D incentives in use around the world provide additional options for policymakers to consider to enhance U.S. competitiveness:35
Accelerated depreciation of qualifying R&D assets
Allows greater deductions in the earlier years of an asset
Can be used to minimize taxable income
Purpose: encourage more frequent investments in, and upgrades to, R&D assets such as labs and equipment.
R&D expenses super-deduction tax incentive
Allows a taxpayer to deduct qualified R&D expenses from its net income
Can be used to minimize taxable income
Purpose: promote increased R&D spending by corporations.
Cash grants, low interest loans
Provide funds for qualified R&D activity through non-repayable funds or loans with favorable terms and conditions
Purpose: provide capital to entities pursuing high-risk/high-reward research, which often face barriers to obtaining funding.
Tax exemptions and reductions for qualified tech transfer
Can be used to minimize taxable income
Purpose: promote cross-industry technology diffusion and spin-off company creation to boost innovation.
Patent-related incentives such as reduced tax rates on income from intangible assets36
Lower tax rates on assets that do not derive their value from physical attributes, such as software and chemical formulas
Purpose: promote R&D of intangible assets, which often have a longer development time line and a higher risk of failure.
Promote international R&D collaboration
As an open democratic society with world-class universities, research institutes, and corporations, the United States makes for an attractive partner in joint R&D. Decades of experience show that joint work with foreign researchers can be done with great benefit and little detriment to our economic and national security. President Trump’s executive order is right to emphasize the importance of collaborating with international partners.
The benefits of international collaboration include cost sharing; aligning complementary knowledge, experience, and know-how; improved interoperability; developing norms and principles; and more efficient standards setting. The United States joining the Organization for Economic Cooperation and Development (OECD) in adopting global AI principles was an important step in the right direction because it shows U.S. support for international norms in developing trustworthy AI.37 This helps foster global cooperation and promotes values such as human rights.
Global AI issues—ensuring AI is safe, transparent, explainable, reliable, and resilient—are especially well suited to broad international research cooperation.
The United States is fortunate to have most of the world’s leading AI powers as allies and partners. The United Kingdom, France, Japan, Singapore, and South Korea, for example, have committed $100s of millions to AI R&D.38 Toronto is a global AI hub. Each of these locales, and numerous others, are prime candidates for mutually beneficial cooperation. Global AI issues—ensuring AI is safe, transparent, explainable, reliable, and resilient—are especially well suited to broad international research cooperation.
Mechanisms to promote multinational collaboration range from personnel exchanges to establishing cooperative international R&D centers at home and abroad. Such collaborative relationships can be encouraged by enhancing visa and work permit regimes, providing grants and loans, and organizing multinational innovation prize competitions. Such competitions could be modeled on DARPA’s series of Challenges and the XPRIZE competitions, which have successfully tackled some of the toughest science and engineering problems, including in AI.39
The finalist teams for DARPA’s 2016 Cyber Grand Challenge—one of several such events designed to address the nation’s toughest science and engineering problems—gather on stage. The U.S. government could host similar competitions with other nations to promote collaboration on AI research and innovation. (Cherly Pellerin/Department of Defense)
Harness America’s Talent Pipeline
American Talent
Homegrown talent is key for the U.S. AI ecosystem. U.S. leadership in AI begins with STEM education at the K-12 level. The U.S. government has taken positive, albeit belated, strides in improving K-12 STEM education and should continue to expand these efforts. In 2018, the National Science & Technology Council (NSTC) published its Strategy for STEM Education, which built on prior efforts to promote computational literacy like the recent follow-through of the 2016 Computer Science for All initiative.40
The initiative’s original funding request for $4 billion was never realized, but the federal government’s attention to the matter likely contributed to state and local efforts to improve computational literacy programs.41 In 2017, President Trump signed a presidential memorandum which directed the Department of Education to devote $200 million to STEM and computer science annually.42 Technology companies promised another $300 million to support the initiative, well short of the proposed target. Meanwhile, the Department of Education prioritized funding exclusively for computer science for the first time in 2019.43 Yet, while programs to promote computer science have slowly expanded, initiatives specific for building AI skills lag.44
The U.S. government will need to prioritize and invest more in K-12 STEM and computer science education in order to build a robust AI workforce for the future. (Ariel Skelley/Getty Images)
Building American AI talent also relies on having a robust teaching base at universities, but that teaching base is losing its faculty to private companies due to greater resources.45 Computer science faculty at universities who leave academia for private companies and the number of new PhDs who choose industry jobs rose from 38 percent to 57 percent in the last decade.46
This trend may be even steeper in the deep learning field.47 This exodus harms U.S. leadership in AI now and over the long term because it decreases the expert base available for training the next generation of AI talent. Faculty and advanced PhD students selecting into industry in ever greater numbers drains the community of expertise available to train the next generation of AI experts and occurs at the expense of longer-term research projects important to breakthrough innovation. Faculty leaving academia for industry also harms long-term innovation potential. A study published in 2019 concluded that when professors left their teaching positions for the private sector, their students became less likely to start a company and those who did raised less money.48
In addition to higher salaries, those leaving academia for the private sector cite access to compute, data, research funding, and high-impact projects as draws.49 Some companies are taking steps to preserve the faculty bases that build their talent pipelines, for example by establishing fellowships and consortiums or allowing professors to rotate between responsibilities.50 The U.S. government, in its National Artificial Intelligence Research and Development Strategic Plan, calls for long-term investments in AI research and the release of publicly available datasets to fill some of this resource gap.51
Corporate America can further assist in ensuring they will have access to future elite talent and promote U.S. AI leadership in the long term by helping to teach the next generation. Tech leaders including Microsoft, Google, and Amazon make employees available to teach computer science skills such as coding to high school students, particularly in underserved areas.52 Such initiatives help to nurture the future AI industry.
Apple hosts an “Hour of Code” workshop for third grade students in New York City in 2015. Some employees at technology companies like Apple have begun teaching computer science classes and workshops to K-12 students, which helps foster interest in STEM subjects from an early age. (Andrew Burton/Getty Images)
In addition to cultivating the experts responsible for future cutting-edge breakthroughs, the United States will need to facilitate the talent necessary for implementing and managing AI solutions. This part of the talent base will not necessarily require doctorates so much as relevant bachelor’s and master’s degrees for literacy in AI applications. These application-savvy coders will comprise the bulk of the workforce. AI leadership will require both programmers with bachelor’s and master’s degrees and smaller numbers of elite talent.
To ensure the United States has the requisite homegrown talent, the White House and Congress should:
Increase public and private sector AI and STEM education and skills training
The White House’s strategy for STEM education includes plans to cultivate the talent pipeline early with such measures as building computational thinking and teaching data science. It would benefit, though, from emphasis on increasing general AI literacy specifically.53
Congress should provide more funding for NSF to expand grant-giving to school districts to develop new AI-focused curricula and associated resources such as professional development of teachers. In addition, Congress should appropriate the requested $4 billion for the Computer Science for All initiative or a new analogous effort.
To remain competitive, the United States needs a national human capital strategy for technology and must invest in improved education in STEM. The NSTC should build on its 2018 Strategy for STEM Education by presenting a detailed plan to execute their recommendations, task specific government agencies, and fund reforms.54 Additionally, Congress should provide tax credits for companies that offer relevant STEM training to employees, students, and teachers, either internally or through third parties.55
Education is governed largely at the local and regional levels, but the federal government plays an important role in agenda-setting. It can shape progress in AI not just by providing resources but by empowering AI as a national priority and a focus area for education.56
Increase funding opportunities for university researchers
Congress should increase R&D funding for AI research at universities. Federal grants to academia decreased from their 2011 peak of $45.5 billion to $40.9 billion in 2017. The value of those grants dropped even more in real terms.57 Researchers have directly cited stagnant AI R&D funding as an incentive to move to private industry.58 Adequate funding is especially important for machine learning because it is resource intensive. Training a single model can cost tens of thousands of dollars for compute resources alone.59 Increasing AI R&D funding not only will keep more professors at universities; it will enable them to pursue longer-term research in important areas that may be less of a priority for industry research.
Researchers have directly cited stagnant AI R&D funding as an incentive to move to private industry.
International Talent
Immigrants long have been a source of innovation in the United States. Throughout the country’s history, high-skilled immigrants have contributed to some of the nation’s most transformative technologies. Today is no different. International talent remains a critical backbone of the country’s technological ecosystem. Immigrants founded one-quarter of the technology start-ups in the United States, and immigrants and their children founded nearly half of U.S. Fortune 500 companies, including Apple, Google, General Electric, and IBM.60
Sergey Brin, cofounder of Google, is an industry leader and immigrant from the Soviet Union. Brin illustrates the success and importance of international talent for America’s technological ecosystem. (Justin Sullivan/Getty Images)
High-skilled immigrants play an indispensable role in American AI. More than half of the country’s top AI talent base is composed of foreign nationals.61 With too few STEM-educated Americans and higher employment growth in STEM careers compared to the overall job market, U.S. technology firms currently rely heavily on temporary-hire foreign workers to fulfill critical shortages in STEM occupations.62
While labor market indicators suggest a shortage of AI talent, estimates of exact numbers vary considerably.63 More reliable tallies from the broader discipline of computer science and STEM-focused occupations generally hint at the scale of the issue. Within the United States alone, more than 300,000 cyber-related positions currently go unfilled, and this number is projected to skyrocket to an estimated 1.8 million unfilled positions by 2022.64 Within the STEM field overall, a 2018 study by Deloitte and the National Association of Manufacturing estimates a need for 3.5 million STEM jobs by 2025, with more than 2 million of those positions going unfilled due to a lack of skilled talent.65
Even if the United States undertakes a robust—and fully funded—STEM education program, immigrants will remain an invaluable component of the U.S. talent base. It will take a generation to develop a new cohort of American-born scientists and engineers, while high-skilled immigrants can be recruited immediately, resolving acute workforce shortages today. Additionally, immigration allows the United States to draw on the best and brightest from around the world. In a global competition for AI talent, the United States has a natural advantage in the fact that many want to come and work in the United States. Washington should capitalize on this advantage by maximizing opportunities to recruit high-skilled immigrants to work in the United States.
Despite the need for international talent in AI research and development, immigration mechanisms for working in the United States are insufficient and the process for entry is often cumbersome. Since the passage of the Immigration Act of 1990, the immigration cap has remained stagnant, while the U.S. labor force has grown by 30 percent to around 163 million people.66 For high-skilled immigrants, visas and green cards are scarce and difficult to acquire.
U.S. technology companies have relied most heavily on H-1B visas to recruit qualified immigrants. Established in 1990, the H-1B visa program was designed as a short-term solution to address labor shortages in particular areas, allowing employers to hire temporary employees with specialized knowledge.67 As the Internet Age unfolded, the demand for high-skilled labor continued to rise. Subsequently, the U.S. technology sector—with great need for computer science specialists—began using the H-1B visa program to hire international talent.68
Today, the available number of H-1B visas is capped at 85,000 per year, with 20,000 visas designated for those with graduate degrees.69 While this cap has remained at 85,000 since 2005, the number of H-1B applications has skyrocketed, peaking in 2017 at 236,000 applications, though declining to 199,000 in 2018.70 The denial rate for new applicants has grown from 6 percent in FY2015 to 32 percent in FY 2019. The denial rate for visa renewals also has increased since 2017.71
The recent decline in H1-B applications and the increase in denial rates may be a consequence of new policies enacted by the Trump administration. In 2017, President Trump signed the “Buy American and Hire American” executive order, which directed the Department of Homeland Security to award H-1B visas to the “most skilled or highest-paid” workers.72 This led the U.S. Citizens and Immigration Services to reevaluate the kinds of work and educational experiences that constitute a “specialty.” In past instances where temporary visa denial rates increased, employers have reported “time lost due to the increase in denials,” and that the impact of these denials have cost “millions of dollars in project delays and contract penalties.”74
During a visit to a manufacturing facility in Wisconsin President Trump signed the “Buy American and Hire American” Executive Order, which made changes to the H-1B visa program. U.S. technology companies rely heavily on H-1B visas to recruit international talent. (Scott Olson/Getty Images)
While the H-1B visa program is the dominant pathway to hire temporary employees, the Optional Practical Training (OPT), which allows F-1 student visa holders to work in the United States following graduation, is a critically important program for retaining international talent. Student visa holders studying in STEM fields are allowed to work in the United States for up to three years after graduation, and there is no cap on work permits granted under the OPT program.75
In recent years, various lawmakers have called for the limitation or elimination of the OPT program.76 This would be a mistake. The OPT program is the country’s largest source of temporary high-skilled immigrant talent.77 In 2016, 172,000 work permits were granted under the OPT program for F-1 visa holders studying STEM, up from 73,000 in 2014.78
The OPT program is the country’s largest source of temporary high-skilled immigrant talent.
Issues with the available immigration pathways shape the extent to which the United States is a beacon for highly talented individuals from around the world, including China. From 2005 to 2015, nearly 87 percent of Chinese doctoral students studying in the United States planned to remain following graduation.79 Today, while a large fraction of top-tier Chinese AI researchers stay to work at American institutions, the overall total number of Chinese graduates remaining after graduation is shrinking.80 In 2016 there was a nearly 57 percent growth in Chinese international students across all fields of study returning home compared to 2011 numbers.81 China’s evolving technological ecosystem and the numerous obstacles to obtaining a worker visa in the United States are likely major factors in this decline.82
Given the structure of the AI labor market and the demonstrated need signaled by employers, policies to attract and retain the next generation of top STEM researchers is essential to long-term U.S. competitiveness.
To ensure the United States attracts the best AI talent in the world, Congress should:
Reform the H-1B visa application process
The U.S. Congress should work to reform the H1-B visa process to make it more suitable for hiring and retaining international talent.
First, Congress should raise the overall cap of available H-1B visas and remove the cap for advanced-degree holders entirely. By raising the current cap and removing the limit on advanced-degree applicants, Congress would address, partially at least, the striking imbalance between H-1B petitions and available H-1B visas. The annual limit on H-1B visas was exceeded the past 16 years, and thus, by limiting the H-1B visa cap, the United States is arbitrarily restricting a major source of talent for U.S. companies.83 This is an unfortunate example of government intervention in the marketplace that constrains American innovation.
The exact shortage of AI technologists in the United States is difficult to quantify, but it is clear that while the number of AI job postings continues to increase, the number of job seekers has leveled off.84 While the motivation to keep the H-1B cap low is to protect American workers, this is unneeded and detrimental for the AI and computer science fields, where the demand in the marketplace far outweighs the available U.S.-born researchers.85 In order to meet demand, America’s AI talent base for the upcoming generation will need to draw heavily on foreign nationals who choose to live and work in the United States, and Congress should provide sufficient opportunity for U.S. companies to recruit talent from abroad.
Second, Congress should simplify the process of applying for an H-1B visa to make it easier for start-ups and smaller tech companies to hire necessary international talent. The H-1B application process is expensive and requires extensive documentation from the applicant’s potential employer. Consequently, technology giants with greater personnel and resources—Amazon, Microsoft, Intel, and Google—make up a significant percentage of approved H-1B petitions.86 Therefore, in addition to raising the cap on available H-1B visas and simplifying the process, Congress should earmark a percentage of these visas for smaller technology firms and start-ups.
Create new ways to recruit high-skilled immigrants
In addition to reforming the H-1B visa process, Congress and the White House should identify alternative mechanisms to recruit and retain international AI talent.
First, the Department of Labor (DOL) should amend its list of Schedule A occupations to include high-skilled AI technologists. Under Schedule A authorities, DOL has authority to determine whether there are insufficient numbers of American workers for a specific occupation and that the hiring of foreign nationals will not negatively impact U.S. workers.87 As the need for AI technologists far outweighs the number of available U.S. AI scientists, this would be an appropriate occupation for Schedule A designation. If AI specialists were added to the list of Schedule A occupations, employers seeking to sponsor a foreign national for a green card could forgo the first step in the permanent residency process and proceed directly with the I-140 filing process, saving both time and resources for the employer and employee.88
Second, Congress should create a new program to attract qualified international students and retain them for the American AI workforce. This proposal is different from the “staple a green card to a diploma” concept popularized by New York Times columnist Thomas Friedman in that it is more targeted, requires up-front commitment on the part of the program participant, and frees the applicant from requiring employer sponsorship.89
The program as envisioned consists of three phases. In the first phase, an international student applies for an “F-prime” dual intent student visa.90 To obtain such a visa, the student must be accepted into a pre-approved AI-related graduate-level academic program, be successfully screened and vetted by the FBI and the State Department, and commit to working in an AI-relevant field in the United States for a minimum of ten years upon graduation.91 The F-prime visa is guaranteed for the duration of the student’s program of study as long as the student meets certain academic criteria.
The second phase begins upon completion of graduate school. The program participant is provided a ten-year conditional open-market EB-1 green card, the so-called “genius visa” for immigrants with extraordinary skills in their field. Like status quo EB-1 green cards, this new subcategory would have no labor certification or employer sponsorship requirement and allow the individual to work for any U.S. employer, but unlike a typical EB-1, it would not last indefinitely.
In phase three, after nine years of employment in the United States, the participant is eligible to commence the petition to remove the conditions on residence (permanent green card) or apply for naturalization. Unconditional permanent residency or citizenship would be granted upon the successful completion of the ten-year employment period.
U.S. immigrants attend their naturalization ceremony to become American citizens. Immigrants are the bedrock of American ingenuity and innovation. The U.S. government should reform existing immigration pathways and create new ones to recruit additional high-skilled immigrants for STEM and computer science jobs. (John Moore/Getty Images)
The program would have three key benefits. First, through the considerable up-front commitment, it attracts the best and brightest foreign students who are already highly predisposed to live, study, work, and remain in the United States. Second, unlike the “stapled green card’” approach, it is targeted for specific AI-related disciplines. Third, by removing employer sponsorship requirements, you solve the “start-up visa” problem by eliminating the cost and uncertainty of extending job offers to qualified foreign nationals.
Counter Illicit Technology Transfer
The United States needs to double down on protecting American intellectual property. Historically, adversaries and allies alike have pursued U.S. technologies to improve their military or economic comparative advantage including Russia, Japan, France, Israel, and South Korea.92 China is the primary point of reference here, however, because of the sheer scale of its ongoing collection efforts. China is an increasingly capable espionage actor, and is actively collecting information from America’s government, corporations, nonprofits, and colleges and universities. It employs a range of methods, both legal and illegal, to appropriate U.S. technology including U.S. company insiders, employees of Chinese firms partnering with U.S. companies, cyber espionage, foreign direct investment, and academic solicitation.93
In July 2019 FBI Director Christopher Wray stated, “There is no country that poses a more severe counterintelligence threat to this country right now than China . . . and I don’t say it lightly.”94 He further noted that the Bureau had around 1,000 investigations involving attempted theft of U.S. intellectual property (IP). The White House, in its 2017 National Security Strategy, also highlighted the importance of the issue.95 One concrete action the administration recommended was to strengthen the Committee on Foreign Investment in the United States (CFIUS), which was accomplished in 2018 through the passage of the Foreign Investment Risk Review Modernization Act of 2018. The bill cited factors such as the national security risks of increasing control of certain assets by foreign entities, and to what extent a transaction would expose sensitive U.S. citizen data to foreign governments, as considerations.96
The United States Trade Representative calculates that the annual cost of “the theft of trade secrets could be as high as $600 billion.”
Private Sector
China’s efforts to illicitly obtain American technologies and know-how are vast and effective. One in five U.S. companies reported IP theft by China in 2018.97 The United States Trade Representative calculates that the annual cost of “the theft of trade secrets could be as high as $600 billion.”98 This figure does not incorporate the full cost of patent infringement, nor the estimated $400 billion per year lost to economic espionage via cyber attacks.99
To counter the theft of U.S. technology from American companies, Congress should:
Provide more cyber defense support to small firms
Congress should increase funding for efforts supporting cyber defense at smaller firms active in critical technology areas. Small and medium business in general are more vulnerable to cyber attacks.100 One effective initiative worth expanding is the Department of Homeland Security’s National Cybersecurity and Communications Integration Center and the Critical Infrastructure Cyber Community Voluntary Program.101 The Department of Homeland Security’s Inspector General found that the department is overall behind on assessing its cybersecurity workforce and forming a strategy to build that workforce.102 As it builds its strategy, it should build in measures specifically catered toward supporting the U.S. innovation base.
Academia
Countering Chinese espionage and other forms of technology transfer in academia poses vexing challenges. The United States needs to blunt Chinese collection efforts without undermining civil liberties or the free and open environments of American universities. FBI Director Wray called China’s efforts a “whole of society threat” to the dismay of many universities that stressed the importance of openness, inclusion, and the contributions of international students.103 Lee Bollinger, president of Columbia University, went so far as to say that monitoring foreign-born students “is antithetical to who we are.”104 Rather than generate friction between itself and universities, the U.S. government should build dialogue with universities to understand their concerns and look for mutually agreeable solutions. The open environment at universities is important because it is foundational to U.S. innovation. At the same time, this openness is being exploited by China. The United States needs solutions that are consistent with American values while guarding against espionage and illicit or otherwise unintended technology transfer.
The U.S. government can support universities by bringing awareness to specific methods and technologies of interest. Focusing on methods and targeted technologies can create avenues for action that avoid profiling Chinese national students. This approach also would mitigate collateral risks of racial profiling of the Chinese-American and larger Asian-American communities. More than half of the top AI talent in the United States is composed of foreign nationals, and Chinese nationals who study and decide to stay in the United States form an important part of this research community.105
General-purpose technologies such as AI, which are inherently dual-use, create an additional challenge to counterespionage efforts because they are less likely to be safeguarded by classification protocols and because espionage actors may have more claims to plausible deniability. Despite this, the U.S. government can take steps to help protect the technology innovation community, especially universities, against illicit technology transfer.
To address the threat posed by academic solicitation, the U.S. government should:
Authorize consular officials to act on risk indicators for espionage
The U.S. State Department, FBI, and intelligence community should collaborate to protect open research environments by screening out high-risk individuals before they arrive. Possible risk factors include whether an individual is funded by China’s government, including the People’s Liberation Army (PLA), or cites highly specific research interests relating to defense technologies.
Disallowing PLA researchers is one place to start. The Australian Strategic Policy Institute estimates that 500 military scientists from China have been sent to the United States since 2007 and that “research collaboration with the PLA . . . comes with significant security risks while offering unclear benefits.”106 The U.S. State Department should work with the intelligence community to identify other risk factors, and Congress should legislate to authorize visa denials accordingly.
Legislators recently have proposed both actor-based and technology-based approaches to improve visa screening. The People’s Liberation Army (PLA) Visa Security Act would prohibit F or J visas for PLA employed, funded, or sponsored individuals, and the Protect Our Universities Act of 2019 would mandate background screening of students seeking to work on “sensitive research projects.”107 Both proposals are sensible measures and should be implemented, but broader screening also is required beyond simply PLA-sponsored individuals or sensitive research projects. AI is highly dual-use and has both commercial and military applications. Data-informed policies should help to protect universities against individual academic espionage risks while mitigating potential negative effects to open campus environments.
Improve collaboration with colleges and universities
The FBI should increase collaboration with universities and information sharing on academic espionage threats. Universities have a strong interest in preventing countries such as China from unfairly exploiting research by their faculty. Universities and the FBI both already are implementing measures to address security concerns on campus and combat academic espionage.108 Greater dialogue is urgently needed between investigators and academics to better understand the scope of the problem and solutions. Positive action is under way: In September 2019, the White House OSTP published an open letter to the U.S. research community to highlight the issue of research security. OSTP announced its plan to hold meetings at academic institutions across the country to discuss lines of efforts such as coordinating outreach and engagement and assessing and managing risk.109 The Securing American Science and Technology Act of 2019 supports this approach and calls for OSTP to convene meetings on best practices.110 Universities are taking their own steps to address the issue. The Association of American Universities (AAU) and Association of Public & Land-Grant Universities (APLU) sent a letter to universities in April 2019 on best practices for protecting against intellectual property theft and academic espionage.111 More should be done, however, to ensure greater coordination between universities and the national security community on this important topic.
In February 2018, the FBI disbanded the National Security Higher Education Advisory Board (NSHEAB), which had existed since 2005 to establish lines of communication between universities and the national security community on counterintelligence threats, among other issues. Since then, members of Congress and American education leaders have publicly expressed concern about the FBI’s decision to disband the NSHEAB and the need for a reconstituted organization to perform its functions.112 In April 2018, the American Council on Education (ACE) sent a letter to the FBI director on behalf of 15 different American education associations and councils requesting engagement with the national security community in a forum similar to the NSHEAB.113 Similarly, in April 2018 the ACE, AAU, APLU, and Council on Government Relations released a joint statement expressing support for the NSHEAB and a desire for a similar forum.114 U.S. government officials also have expressed support for such a mechanism of collaboration between universities and the national security community. The Commissioner of the U.S.-China Economic and Security Review Commission said the board was “vital and should be reinstated,” and FBI officials signaled in 2018 they were exploring creating a similar group.115
The FBI should reconstitute the NSHEAB or a similar body to increase awareness and cooperation on countering espionage in U.S. academia.116 Many universities would support this reconstitution and it would provide a valuable mechanism for coordination between universities and the national security community on sharing information on threats and actions to counter academic espionage.
Other agencies that work on technology transfer should build dialogue with universities as well. The Department of Commerce Bureau of Industry of Security already engages in training with colleges and universities to raise awareness of export control laws. Although export controls may not be suitable for AI software, the Department of State’s Bureau of International Security and Nonproliferation can use its expertise in dual-use technologies.117 Working off the Department of Commerce trainings as a model, the Department of State should work with researchers to identify red flags for dual-use technologies. As one PLA researcher said, foreigners “can be asked to develop an algorithm but not briefed on the details of how the algorithm would be used.”118 AI can be misused for malign ends but government experts can train university researchers to conduct due diligence of possible end uses. U.S. government collaboration with universities should not be limited strictly to security measures but should include knowledge-sharing on technologies themselves.
AI can be misused for malign ends but government experts can train university researchers to conduct due diligence of possible end uses.
Protect America’s Edge in Hardware
AI systems require computational power (“compute”) to run. Access to compute can make or break an AI project. Computational power rests in hardware, and like any other specialized physical object, hardware has its own design, fabrication, and supply chain considerations.119 Availability of compute and supply chain dynamics therefore will drive national AI adoption potential, including the ability of a country to develop more advanced systems. U.S. policymakers should regard hardware as an equal part of the talent-data-hardware triad and work to boost American leadership and security in this space.
The U.S. government has ongoing efforts to create compute infrastructure for research, to increase hardware supply chain security, and to develop next generation AI-optimized chips. Sound policies in all three areas are required to protect America’s technological leadership.
To support research, the White House’s American AI Initiative directs federal agencies to allocate compute resources for AI-applications and R&D.120 A number of U.S. government supercomputers built for AI applications, like the Department of Energy’s Center for Accelerated Application Readiness, have opened applications to the public for research projects.121 For hardware development, the National Science Foundation has jointly funded projects, for instance with DARPA, to develop next-generation chips.122
To address supply chain security, the Department of Defense has a Trusted Foundry program. It sought to accredit microelectronics sources as “trusted” if they could assure the integrity of persons and processes for chips’ chain of custody.123 As of 2016, the program boasted about 20 foundries, with IBM running the facilities responsible for the vast majority of leading-edge custom-made chips.124 These facilities lost their eligibility, however, when GlobalFoundries, a U.S.-based but United Arab Emirates–owned company, purchased them.125 For lack of alternative options, the Department of Defense signed a seven-year contract with GlobalFoundries to continue purchasing microchips.126 The Government Accountability Office (GAO) says that the Department of Defense is still seeking “new approaches to retain trustable, leading-edge capabilities.”127
U.S. government efforts to build and deploy hardware are outlined in the American AI Initiative, the White House’s National Artificial Intelligence Research and Development Strategic Plan, and some agency-specific projects.128 Congress and the administration should do more to make compute available, to mitigate supply chain risks, to build the next generation of AI-optimized chips, and to protect America’s edge by limiting the diffusion of advanced semiconductor manufacturing equipment.
To preserve and promote America’s advantage in AI hardware, Congress and the White House should:
Increase the availability of affordable compute resources
The U.S. government has a long history of building high-performance computers and awarding grants for researchers who require compute for their projects. It should supplement these existing efforts by using innovative approaches to facilitate access to affordable compute resources. While most researchers will not require exascale computing like that offered by Oak Ridge National Laboratory’s next supercomputer, compute even for smaller projects is expensive.129 It often poses a barrier to entry for start-ups, universities, and colleges. Smaller institutions like high schools doing the basic training of the next generation of AI leaders often only can afford the bare essentials.130 For example, the cloud computing cost for AlphaGo Zero—the improved version of DeepMind’s AlphaGo—by itself cost around $35 million.131 AlphaStar, DeepMind’s AI program that beat a top professional human player at the Starcraft II strategy game, cost as much as $100 million.132
The U.S. Department of Energy’s Oak Ridge National Laboratory (ORNL) unveils its most powerful scientific supercomputer. Supercomputers provide the U.S. government with massive amounts of compute, and the government is working to give researchers in academia increased access to compute. (ORNL/Carlos Jones/Flickr)
To help address the compute access gap, the National Science Foundation is building relationships with cloud computing providers through the “Enabling Access to Cloud Computing Resources for CISE (Computer and Information Science and Engineering) Research and Education” program and the “Exploring Clouds for Acceleration of Science (E-CAS)” project.133 NSF should expand these projects, and Congress should increase funding for them.
Congress also can incentivize companies to donate compute resources to universities. The compute used in the largest training runs doubles every three and a half months, and demand will likely increase.134 The U.S. government operates a number of compute facilities that are open to university researchers, but increasing capacity will be necessary to satisfy growing demand. In Massachusetts, a number of top universities partnered with the state government and the private sector to establish the Massachusetts Green High Performance Computing Center, which houses a number of high-end computer systems available to researchers.135 State and federal agencies partnered to build the infrastructure necessary to support the center. Additionally, the project benefited from the federal New Markets Tax Credit Program that promotes private capital investment in low-income regions.136 Tax incentives for donated or discounted compute resources could increase universities’ access to commercial compute resources, especially at a time when more companies seek relationships with academia.
Establish multilateral export controls on semiconductor manufacturing equipment (SME) and increase federal R&D funding for next-generation hardware
The United States has a major global lead in semiconductor design and should enact multilateral export controls, in concert with allies and partners, to protect its competitive edge in hardware. China is currently heavily dependent on imports of foreign-manufactured semiconductors to meet internal demand. As part of its Made in China 2025 plan, China is looking to reduce its reliance on foreign chips by ramping up domestic semiconductor production.137 Yet this desire to indigenize production is a major source of strategic leverage for the United States.
To accomplish this goal, China needs foreign imports of semiconductor manufacturing equipment (SME), which are the equipment and tools needed to establish a chip fabrication facility, or foundry.
The global SME market is highly centralized, with the United States, Japan, and the Netherlands accounting for 90 percent of global SME market share.138 In key areas the market is even more concentrated. A single Dutch company is the sole supplier of extreme ultraviolet lithography machines required to make the latest generation of semiconductors.139 Nearly the entire global supply of photoresists, chemicals essential to the production of semiconductors, is produced by a handful of companies based in the United States, Germany, Japan, and South Korea.140
The Commerce Department and State Department should work with key allies and partners (the Netherlands, Japan, South Korea, and Singapore) to establish multilateral export controls on SME, restricting sales to China. While export controls on semiconductors themselves should be rare and targeted, such as the action against Huawei and a handful of other companies linked to the Chinese military, the United States should enact broad restrictions on sales of SME to China, working in concert with allies and partners, in order to sustain the U.S. advantage in hardware.
The United States should enact broad restrictions on sales of SME to China, working in concert with allies and partners, in order to sustain the U.S. advantage in hardware.
One risk to SME export controls is that they deprive U.S. companies of profits they currently use to invest in R&D. Chinese investments in domestic chip fabrication have buoyed the SME industry, accounting for much of the recent market growth.141 U.S. companies must continue to invest in next-generation techniques to remain global leaders. As chip designs approach the atomic limit of silicon and Moore’s Law comes to an end, chip companies are searching for the next breakthrough that will lead to a new generation of computing hardware. In order to ensure continued U.S. leadership in semiconductors, the federal government should increase R&D in next-generation chip design, fabrication, and packaging. Enacting the following two recommendations also would offset SME revenue losses.
Boost domestic semiconductor manufacturing with retooling incentives
Reinvigorating the Trusted Foundry program would be a positive step for U.S. leadership in AI hardware and help to ensure a secure supply chain. Building plants for semiconductors is expensive—an advanced fabrication facility can cost up to $20 billion.142
A feasible alternative to all-new facilities to revive this part of the Trusted Foundry Program is to support retooling of existing U.S.-based foundries to facilitate leading-edge hardware. Some fabs need to retool every two to three years to stay competitive, and these costs are burdensome.143 Several companies already have requested funding to upgrade fabrication facilities in order to supply the government.144 The United States should identify the U.S.-owned domestic facilities most suitable for retooling and develop a long-term funding plan to rebuild a trusted leading-edge semiconductor manufacturing capacity.
Secure semiconductor supply chains
The United States should work with U.S. industry leaders to explore novel public-private partnerships to ensure trusted semiconductor supply chains and work with key allies to establish an international fab consortium to diversify semiconductor fabrication. The United States is a global leader in semiconductor design, with U.S.-headquartered firms accounting for roughly half of the global market, but most fabrication occurs overseas.145 This heavy reliance on overseas production presents risks of disruption or vulnerabilities introduced into the supply chain.146 For example, chip production at Taiwan Semiconductor Manufacturing Company was disrupted briefly in 2018 when a computer virus made its way onto fabrication equipment.147 Taiwan in particular is a major locale for semiconductor fabrication, accounting for over 70 percent of fabrication in “pure play foundries.”148 Taiwan is a major target of Chinese hackers and potential insider threats and a potential flashpoint for conflict.149
Employees at SK HYNIX Inc., South Korea’s largest semiconductor company, work on a production line at a plant in Icheon, South Korea. Semiconductor supply chains are increasingly global, and their chains of custody pose challenges for security verification, which is of growing concern for the U.S. Department of Defense. (Pool/Getty Images)
The costs of establishing a new foundry are significant, on the order of $10–20 billion, making on-shoring costly even with potential government subsidies.150 Additionally, the U.S. military and intelligence community have special needs for security that go above and beyond what is available in commercial facilities, yet they lack the scale of demand to make a purely government-dedicated foundry profitable.151 The DOD and intelligence community should explore novel approaches for public-private partnerships with U.S. companies to build the capability for trusted design, fabrication, packaging, and testing. Additionally, the United States should establish an international fab consortium with allies to share the cost burden of building new semiconductor foundries to ensure a trusted and diverse supply chain. Member nations should include the global leaders in semiconductor manufacturing equipment: the United States, Japan, and the Netherlands.
Shape Global Norms for AI Use
The applications of AI can do incredible good for societies. It can optimize city systems, study employment patterns to give insights to policymakers, and revolutionize biotechnology. Technology in general can be used to make human life easier but only if it is subjected to informed policy and good governance. Increasingly, AI-enabled technologies are being abused by authoritarian states to exert control over their populations. In some cases, U.S. research and academic and private institutions have been complicit in enabling these abuses.152
Myriad potential risks arise with AI technologies. AI’s general purpose and dual-use nature increase the risks of misuse and accidents. Risks of misuse involve the possibility of individuals or groups using AI systems in an unethical way, while accident risks are the potential harms that stem from AI systems behaving in unexpected ways.153 AI also poses structural risks, meaning that AI technologies have the capacity to shape the political, economic, or social environment in disruptive or harmful ways.154 Conversely, factors derived from structure, such as competitive advantages, also could influence how actors use AI, including by creating perverse incentives, such as actors racing to develop the technology first and taking shortcuts on safety.
As AI increasingly is incorporated into technologies that impact Americans’ daily lives, the potential for misuse, accidents, and structural risks increases. Facial, image, and speech recognition algorithms are becoming more accurate, ubiquitous, and affordable every day. While they strengthen law enforcement capabilities, these technologies also raise concerns about civil liberties. Speech generation and synthetic media can be used in art, but also to spread disinformation and undermine the public’s trust in truth and facts.
American leadership, alongside other democratic nations, in shaping international AI norms is essential to ensuring these technologies are developed and used in ways that align with democratic interests and values. The administration and Congress should:
No comments:
Post a Comment