The United States, with much of the world’s AI-enabling infrastructure, has positioned itself as the global leader in AI innovation. That might not be the case for much longer. China continues to develop state-of-the-art AI systems, despite an extensive network of U.S. and allied export controls aimed at curtailing China’s technical progress. Beijing is expanding and accelerating new investments in AI and investing significant government resources into supporting its emerging domestic technology base. The result: global AI competition has now become synonymous with U.S.-China competition.
As the United States fights to maintain technological primacy, policymakers are enacting and considering new export controls to limit China’s access to the necessary components for building advanced AI systems, such as American-developed state-of-the-art models, semiconductors, and computational power.
Export Controls in US-China AI Competition
Getting export controls right is notoriously difficult. Not long after the United States passed its expansive export controls targeting China’s access to AI-relevant semiconductors in October 2022, cracks in the effectiveness of these controls appeared. In response, China increased investment in its domestic semiconductor manufacturing capabilities intending to achieve self-sufficiency. Markets, too, responded to these controls with successful smuggling efforts, pouring controlled chips into the PRC. Firms like NVIDIA diligently retooled products to route sales around legal strictures. For tangible goods such as semiconductors, the success of export controls has proved challenging.
Now, U.S. policymakers are eyeing new export controls on AI at the software level. The Biden administration reportedly is considering enacting new restrictions for “closed models” — AI systems without publicly available source code — and the pending Enforce Act would extend and clarify the applicability of existing export regulations to both AI software and support services.
Controlling physical goods is already difficult. New initiatives to restrict or control the export of dual-use software, like AI, will be even more so due to its “invisibility” and “intangibility.” Unlike physical goods, software is borderless, easy to hide and obfuscate. This allows software to slip through controls via a range of means and methods.
Even if stringent controls on software were put into place, without being given access to source code or expansive system access it can be difficult to distinguish a covered AI system from other types of software. In response to this challenge, policymakers involved in early AI regulation discussions were supportive of stringent controls at the earliest stages of AI development, such as licensing for AI development or regulating compute, on the grounds that export controls would be more effective if intangible goods could be contained at their source.
The AI ecosystem of today, however, has evolved. Instead of being concentrated in a handful of AI development firms, numerous state-of-the-art open-source solutions are now available with little performance difference from closed-source models. As this capability gap between models closes, the long-term effectiveness of export controls specifically targeting AI software is likely to decrease further.
Not only will these export controls be increasingly difficult to implement, but they would also unlikely be in the best interests of the United States. Indeed, the current trajectory of export policies risks unintended consequences for little long-term strategic benefit. These include a decline in the competitiveness of the United States, a decoupling from U.S.-developed technology, and uncertainty for the domestic tech industry, amongst other risks.
A Better Way Forward
For the United States to maintain its global AI leadership, it must focus on competition and outcompeting its geopolitical rivals in the development, implementation, and diffusion of AI-based systems domestically and internationally instead of an expert-control-first approach. Defending against the rise of digital authoritarianism requires embracing competition and openness, enabling effective market access, and supporting the diffusion of U.S. AI-enabled technology and governance standards.
There are several steps U.S. policymakers can take to move in this direction:
First, adopt a “do no harm” and common sense approach to AI export controls, relying on existing authorities where possible. The Bureau of Industry and Security’s power to restrict certain application-specific AI systems is one clear example, as are “catch-all controls” that allow for any good that could be used to aid in the production of nuclear or biological weapons to be controlled. If an AI system was found to be able to enable or support the creation of these threats, existing authorities could likely be applied to stop its export.
Second, work with international partners to develop standards that facilitate the international adoption and diffusion of American-built AI systems throughout the world. It is essential for the United States to take a leading role in the development of internationally accepted, objective, and science-backed standards. This is a mission that the National Institute of Standards and Technology has historically excelled in, but where further resources and support clearly needed.
Third, acknowledge that much of the world’s future AI users reside outside of U.S. borders. The United States has much to gain for its global competitiveness if it makes the necessary investments in tailoring new AI systems to diverse international communities by supporting under-resourced languages and aligning with local customs and norms.
If the United States fails to change course from an AI strategy based on control, not geopolitical competition, it will quickly find itself at a strategic disadvantage – with adversarial nations filling the void. By contrast, embracing competition as a guiding component of its AI strategy will allow the United States to strengthen its market position and technological leadership while unleashing AI’s domestic and global potential.