US-China AI talks unlikely to yield binding agreements, Campbell warns
In brief
- Trust deficit blocks enforceable AI agreements between US and China, Campbell says
- Risk notification mechanisms replace formal restrictions in recent US-China discussions
- Treasury Secretary Bessent proposed AI dialogue framework in September 2026
- Semiconductor export controls excluded from US-China AI proposal
- US and China pursue fundamentally different AI governance models
The notification framework approach
Rather than pursuing formal restrictions, recent discussions have centered on risk notification mechanisms. Treasury Secretary Scott Bessent led discussions in New York on September 20, proposing the creation of a US-China AI dialogue framework. The framework would establish a system where each country notifies the other about national security-level AI incidents.
Notably, the proposal steered clear of export controls on advanced semiconductors. This omission reflects the complexity of negotiating AI governance when semiconductor supply chains remain deeply entangled—and when trust is already fractured.
Divergent governance models
Prior dialogues included 2024 talks in Geneva focused on nuclear command risks, and a communication channel was established after the Trump-Xi summit in Beijing in May 2026. Yet structural differences in how each nation approaches AI remain unresolved.
Washington generally favors a model where private companies lead development under light regulatory oversight, with targeted export controls to limit adversary access. Major US AI firms including OpenAI and Nvidia participated in the discussions. Nvidia's high-end GPUs have been subject to US export restrictions targeting China since 2022.
Beijing prefers state-directed AI development with fewer constraints on government surveillance applications but tighter control over private sector autonomy. These competing philosophies make binding agreements difficult—each side views the other's model as a security threat.
Why verification matters
Campbell's core argument cuts to the heart of the impasse. Verification is impossible when the two systems operate on incompatible principles. A US company can report its AI capabilities (or claim to); a state-directed program in Beijing may not disclose anything at all. Risk notification sidesteps this problem by avoiding commitments that can't be monitored.
The result is pragmatic but limited. Notification channels don't restrict development. They don't slow proliferation. They simply create a line of communication when things go wrong.
