Export Controls in the Age of AI
What does technological leadership look like in an era of artificial intelligence? The United States, like other countries, is in the midst of grappling with this question against a backdrop of the rise of China and the growing realization that “business as usual” will no longer suffice for America to maintain its technological advantage. Washington has begun to take some important steps to translate this realization into action. In February, President Donald Trump launched the American AI Initiative in recognition that “American leadership in AI is of paramount importance to maintaining the economic and national security of the United States.” In a less constructive fashion, two months later Sen. Josh Hawley (R-Mo.) introduced the China Technology Transfer Control Act of 2019 that would “make it harder for American companies to export major emerging technologies to China.” Clearly, AI is on the agenda.
Unfortunately, Washington appears to be defaulting to traditional, 20th-century policy tools to address a 21st-century problem. Last November, the Department of Commerce published its advanced notice of proposed rulemaking to signal to the public that it is considering a regulatory change to the export of several emerging technologies, including AI. The notice establishes how the public can participate in the rulemaking process via the submission of comments. By the January deadline, 245 public comments had been submitted to the Bureau of Industry and Security.
Export controls are a web of regulations that prohibit the transfer of certain commodities or information, motivated by national security concerns or trade objectives, or both. During the Cold War, the United States and its allies developed an American-led system of export controls to sustain the West’s advantage by denying or delaying the Soviet Union’s acquisition of militarily relevant goods and technologies. Today, those who favor the proposed export controls hope that these measures can curb what they regard as China’s aggressive efforts to appropriate Western intellectual property, particularly in high-technology domains. Critics, on the other hand, warn that export controls risk intensifying the technology race between the United States and China and crippling the U.S. commercial AI industry. Moreover, these opponents argue that new export controls on emerging technologies — designed to address national security issues — could exacerbate broader trade tensions between the United States and China.
The pursuit of AI is becoming a key feature of great power rivalry, but this is crucially unlike previous periods of competition. For the United States, securing technological advantage on the world stage will require moving beyond traditional tools such as export controls. This is a time for new, globally-oriented policies. These should be focused on creating an innovative and open environment for research and development, as well as balancing the interests of multinational tech corporations, allied countries, and scientists. Critically, the United States must also bear in mind the stakes — the opportunity to realize the immense bounty of AI as a transformative general-purpose technology, and the risk of these gains being squandered by geopolitical competition. Washington should therefore be wary of stoking potential hostility — for instance, through the use of exclusionary policies such as export controls — and instead focus on aligning U.S. national strategy on AI with a vision to serve the common, global good.
Great Power Contestation in the Era of AI
America’s pursuit of technological advantage used to be more straightforward. Consider the strategy implicit at the beginning of the Cold War. In its mission to contain the Soviet Union, the U.S. government would actively support strategically relevant industries at home, creating loyal firms that were dependent on the state for investment and support. Simultaneously, the United States and its allies would devise mechanisms to prevent or delay the loss of technology to the adversary, using tools such as export controls to limit the movement of specific goods and technical information across its borders. These included, for example, radar equipment and aero-engines. This was achievable and effective at least for some time because American firms were the leading producers of these technologies, and the United States could reliably access them for military purposes.
Today, however, a state-led national technology strategy is inadequate. Cutting-edge developments in high-technology domains such as AI are spearheaded by powerful multinational technology companies with distinctly global interests and perspectives, rather than by the government. These companies are not dependent on government funding, but instead reinvest a substantial share of their internationally accrued revenues in research and development. They are focused on their 21st-century missions and on international markets, and are composed of influential employees who often share these internationalist values.
Meanwhile, the U.S. government has come to depend on these multinational companies to project the nation’s influence on the international stage and strengthen America’s industrial base. Thus, the successful pursuit of power now requires cooperating with private global tech companies. Further, for export control policies to work, they will need to be supported by both the U.S. public and America’s allies in order to gain the necessary political legitimacy and economic leverage. Notably, both the American people and the country’s allies are presently less convinced than those inside the Beltway of the threat from China.
Export controls made sense during the Cold War with the Soviet Union. However, the Cold War is over and the Soviet Union no longer exists. Today, the effectiveness of export control policies has been challenged by a shift towards global technology markets, as well as private sector leadership in the development of strategically important technologies. Export control policies risk interfering with the economic aims of global tech companies, obstructing the scientific aspirations of researchers, and offending allies by making the United States seem overly nationalistic and aggressive. To succeed, then, export controls will need to be well-motivated and carefully designed — a substantial challenge in today’s context.
Will Export Controls Aid U.S. Technological Leadership?
Export control policy must weigh the envisioned national security benefits against any unintended economic, scientific, and political costs. The primary aim is to deny rival countries or malicious actors access to strategically important technologies. Nations are consequently obliged to develop these technologies domestically, rather than shortcutting the technological development process by acquiring U.S. technologies.
Whether the national security benefit of restricting the diffusion of strategic technologies can be achieved depends on several features of both the technology and the international context in which it operates. First, export controls are most effective when the actor implementing them — a single country like the United States, or the members of an export control agreement like the Wassenaar Arrangement — has sufficient market leverage over the targeted product such that barring its export has a meaningful impact on the product’s availability on the international market. In the case of AI, however, there are few components over which the United States has a clear monopoly, or where allies who can produce the target components are willing to harmonize their export control systems with that of the United States. This limits the effectiveness of export controls in achieving a comprehensive restriction on the availability of AI technologies on the international market. A notable exception is in the equipment and expertise required for the design and production of microprocessors for AI systems, in which a small number of democratic countries, including the United States, still have an edge and thus could plausibly implement export controls effectively.
Second, the restricted technology should be feasible to control, which is more likely to be the case when it is tangible, large, can only be exported through certain ports, and so forth. Many of the key components of AI, however, do not have those properties. Algorithms and software are intangible or small (the size of a data-stream or thumb drive) and can be exported through any internet-connected device. The fact that export controls failed to prevent the proliferation of cryptographic products illustrates this.
Valuable information is not always digitized. Moreover, much of it is intangible, such as the technical skills required to integrate software and hardware components into a product, or the intuition that comes with deep research experience for where the next technological breakthrough may lie. Thus, policymakers may be tempted to try to control the movement of and conversations between skilled individuals. Indeed, the recent tightening of visa restrictions for Chinese students and the introduction of the Protect Our Universities Act indicate a building policy momentum in this direction. Such policies conflict with liberal and scientific norms about the free flow of ideas and people, and have been heavily criticized by national and global scientific communities. Further, export controls which would prohibit foreign nationals in the United States from working on technologies covered by new export control regulations risk undermining the U.S. technology base — more than half of the top AI talent working in the country are foreign nationals, with the largest group coming from China. If these experts are forced to leave, or are never allowed to enter, AI export controls could inadvertently strengthen the technology base of U.S. competitors while weakening its own.
Finally, export control policies risk imposing substantial costs on the private sector. The case of satellite technology demonstrates the chilling effect of export controls on the commercial satellite industry. The Strom Thurmond Act of 1999 placed communications satellites and all of their associated components and research on the U.S. Munitions List. This contributed to the decrease in U.S. market share of the global satellite components market from 90 percent in 1995 to 56 percent in 1999. Similarly, export controls on cryptographic products in the 1990s caused technology companies to lose hundreds of millions of dollars in sales to foreign competitors. Hampering the performance of commercial firms has substantial negative second-order effects, from harming the health of the domestic economy to undermining American economic and technological leadership and weakening the national industrial base.
In sum, history suggests that export controls, if not wielded carefully, are a poor tool for today’s emerging dual-use technologies. At best, they are one tool in the policymakers’ toolbox, and a niche one at that. Encouragingly, at least a few recent remarks suggest that the White House recognizes some of these concerns.
A Positive Agenda for Technological Leadership
If the goal is U.S. technological leadership, policymakers should look beyond export controls and develop a comprehensive, positive strategy for U.S. scientific and economic innovation. The Trump administration’s executive order on AI signals a step in this direction, emphasizing the development of domestic AI talent, the promotion of an international R&D environment, and investment of federal resources into AI research and commercialization. However, much needs to be done to clarify the strategy, make resources available, and implement this plan.
The first important element of a positive agenda is improved STEM education at American high schools and universities. This is critical not only to nurture new AI talent, which is currently in scarce supply globally, but also to retain this talent within the United States. Training and maintaining a highly skilled domestic AI workforce is an important cornerstone for strengthening the industrial base, maintaining a leading position in global markets for AI, and bolstering overall American technological leadership. STEM education has long been neglected in the United States relative to other countries such as China.
Second, given its high dependency on foreign researchers, the United States should adapt its immigration policy so as to better attract and retain AI talent from abroad. Promisingly, as of 2019, the United States is in a leading position globally in terms of attracting and retaining AI researchers. One step is to streamline the visa application process for highly qualified workers. On average, the process currently takes several months, whereas the procedure in countries such as Canada only takes two weeks. Changes in visa regulations could also make it easier for foreign STEM students at U.S. universities to stay after graduating.
Third, to effectively shape the development of AI internationally, the United States will need to articulate a persuasive global vision for AI and establish sufficient domestic support for that vision. The vision should acknowledge the necessity of addressing the social and economic risks from AI given the strong concerns held by the American public. Further, the United States should commit to investing in global governance efforts to address the social, ethical, and economic dimensions of AI development. There have been promising efforts in this vein spearheaded by the recently established Joint Artificial Intelligence Center and the Defense Innovation Board. Further, the United States should prioritize close cooperation with allies, and acknowledge that strategic cooperation with rivals — specifically China — in areas such as AI safety research and the development of international standards may be necessary for Washington to be able to maintain and project its leadership.
The Global Stakes
Export control conversations focus on techno-military superiority, but the stakes are even higher. The game of technological preeminence shapes world order — whose values are built into future institutions, which actors become powerful or irrelevant, and how transformative technologies like AI will impact the structure of the economy and the functioning of societies. From this global perspective, the consequences of a techno-military race could be terribly destructive. Indeed, the competition over AI is but one of several broader trends in global politics that risk a reversion to rivalrous techno-nationalist economic blocs. This could, in turn, promote competition between these blocs and consequently undermine globalization, innovation, and geopolitical stability. This then risks weakening our ability to coordinate transnationally to solve other global problems like global health, poverty, climate change, and environmental sustainability.
Critically, this means that Western technological leadership needs to be pursued with caution, moderated by the knowledge that national strategies are now, more than ever, dependent on the health of the international economy and strategic cooperation with competitors. A robust national strategy for AI should be grounded in both the national interest and a commitment to build the necessary institutions and norms to govern AI for the global good, for generations to come.
Jade Leung is Head of Research and Partnerships at the Centre for the Governance of AI, housed at the Future of Humanity Institute at the University of Oxford. Her research centers on the strategic politics between firms and states with respect to emerging transformative technologies.
Sophie-Charlotte Fischer is a PhD Candidate at ETH Zurich and Research Affiliate with the Centre for the Governance of AI. Her research focuses on business-government relations in the defense space and the diffusion of emerging dual-use technologies.
Allan Dafoe is Associate Professor (Oxford) and Director of the Centre for the Governance of AI. He works on the governance of transformative AI and AI grand strategy.
For helpful comments we thank Emefa Agawu, Markus Anderljung, Jeffrey Ding, Carrick Flynn, Ben Garfinkel, and Remco Zwetsloot.
Image: Wiki Commons
