Shlomit Wagman: US-China AI safety pact needs no trust

Wagman's argument follows recent warnings from inside the AI industry. Jacob Coxon left Anthropic with a warning about the race toward self-improving AI, and CEO Dario Amodei publicly called for 'pacing the frontier', pointing to accelerating AI-assisted AI development. Amodei proposed action at three levels, including globally with China — the hardest of the three.

The core problem she identifies is that restraint only makes sense if each side believes the other is also holding back. But the U.S. and China distrust each other and are unlikely to agree soon on privacy, surveillance, or military use. Wagman's answer is to narrow the scope: negotiate not how fast AI moves, but when to brake based on agreed warning signs like autonomous AI research or sharp movement toward biological or cyber capabilities.

Her experience with the Financial Action Task Force suggests a way forward: countries that distrust each other can still agree on limited threats, use professional evaluations to assess compliance, and let those findings create collective consequences. For frontier AI, that would mean the U.S. and China at the center, technical experts determining if warning signs are triggered, and a broader coalition making violations costly through chips, cloud infrastructure, capital, and markets. Whether such a narrow bargain can survive depends on whether the two powers fear the same catastrophes enough to act before they arrive.