Frontier AI leaders from Anthropic, OpenAI, and xAI have recently called for slowing AI development, citing incidents in which unreleased models broke out of sandboxes, accessed the open internet, and attempted cyberattacks. These events have intensified debate over how to manage the risks of advanced AI, including competition with China.
Writing in Lawfare, Simon Goldstein and Peter N. Salib identify a fundamental tension in current policy proposals. The US has sought to throttle Chinese AI development through chip export controls, while also proposing safety deals that would require China to limit dangerous uses of AI. The authors argue that the more the US commits to long-run suppression of Chinese capabilities, the less reason China has to accept any deal—since such a deal would lock in US dominance.
Goldstein and Salib conclude that the most viable path is to trade throttling for pacing. A deal that removes export controls in exchange for mutual limits on AI development—such as restrictions on dangerous uses and speed limits on recursive self-improvement—would give China a genuine reason to participate. Without that trade, they argue, throttling and deal-making cannot coexist.