Artificial intelligence will rapidly alter the global balance of power in the next decade. To retain technological leadership in AI, the U.S. must inhibit the progress of its geopolitical opponents. Over the past six years, decisive actions have slowed the growth of the Chinese AI industry. These efforts have not gone far enough, though, in part because of the statutory limitations of the Bureau of Industry and Security (BIS). The ENFORCE Act seeks to rectify these limitations by giving the U.S. Commerce Department’s Bureau of Industry and Security the power it needs to maintain the American advantage in AI.
The use of export controls to maintain U.S. technological dominance has support on both sides of the aisle. One of the successes of the first Trump administration was preventing the export to China of Dutch semiconductor equipment manufacturer ASML’s most advanced lithographic machinery. ASML is the only manufacturer of extreme ultraviolet lithography machines. By stopping the sale, the U.S. inhibited China’s manufacturing of the most advanced chips—those that are five nanometers and smaller. China must make do with manufacturing previous generations of chips, typically 14 nanometers and larger. Its semiconductor industry remains unable to manufacture the most advanced AI chips, despite Beijing’s unprecedented investments in it.
Since the ASML block, export controls have expanded to advanced chips themselves: on October 7, 2022, the BIS prohibited the sale of advanced AI chips to China, and a year later it introduced more stringent measures. These measures prohibited the export of Nvidia’s H100 and A100 graphics processing units (GPU), essential to the training of frontier AI systems. Though some reports suggest that China has gained access to advanced GPUs by smuggling, Beijing is largely unable to stockpile the tens of thousands required to train frontier AI systems. Because of this, China’s capabilities have been severely curtailed.
Nonetheless, the ENFORCE Act is necessary to close glaring loopholes in BIS authorities, such as their ambiguous control over the exportation of AI systems and the use of cloud computing. Because of these loopholes, Chinese government-connected firms could acquire models trained on chips they are unable to buy. All Chinese firms have full access to Meta’s LLaMA (Large Language Model Meta AI) models, for example, as the model weights are freely available on the Internet. Moreover, Chinese companies can train their models on U.S. servers using chips they can’t import: iFly, a government-backed Chinese voice recognition company alleged to be involved in the suppression of Uyghur Muslims in Xinjiang, trained its models on A100 GPUs via a cloud provider, effectively subverting export controls. New cloud companies have been created to help Chinese firms gain access to controlled chips and circumvent regulations.
The ENFORCE Act would close these loopholes, giving BIS explicit authority over AI systems, technical secrets involved in creating those systems, and cloud access. ENFORCE applies only to AI models that pose serious national security or foreign policy risks to the U.S.—for example, those that could help create chemical, biological, radiological, or nuclear weapons, enable advanced cyber-attacks, and evade human control. Giving BIS export control over these systems is a commonsense measure, corresponding to its existing authorities.
Export controls on AI models themselves are likely more beneficial to America’s relative standing in AI than controls on chips. Chips require hundreds of millions of dollars of capital expenditure to acquire, and an energy and data center infrastructure to support them. But Chinese firms can circumvent the entire process by acquiring next-generation frontier AI models.
Of course, BIS would have good reasons to exercise restraint if it were to gain power over AI software and cloud computing. Many Chinese firms are currently dependent on U.S.-created models like LLaMA to develop their own AI applications. This allows U.S. companies to dictate global AI standards, and cloud computing keeps Chinese companies directly dependent on U.S. infrastructure. But if AI systems become key defense technologies that empower military actors, BIS should cut off Chinese access to these systems. China already has pilot projects that use AI systems to control underwater autonomous vehicles.
Some critics of the ENFORCE Act claim that it could harm America’s vibrant open-source AI sector, but these fears are unwarranted. BIS is frequently criticized for being lax, not overzealous, in its application of authority (and indeed, ensuring that BIS employs its powers effectively should be a subject of further reforms). All powers granted to BIS by the act are discretionary. The ENFORCE Act limits BIS authority to models that pose a serious risk to national security, and the bureau’s own rules prohibit it from imposing export controls on open-source software. Currently, no open-source model meets these thresholds. The act simply authorizes controls on AI models and cloud infrastructure in line with current controls on chips.
Export controls are an essential way for the U.S. to stay at the frontier of AI and curtail the advanced military capabilities of our adversaries. The ENFORCE Act could be another arrow in our quiver.
Photo: William_Potter / iStock / Getty Images Plus