US and China Set for AI Talks Amid Tensions Over Competition and Safety

Treasury Secretary Scott Bessent is scheduled to meet with Chinese Vice Premier He Lifeng this weekend to discuss artificial intelligence, even as former President Donald Trump has dismissed public warnings about advanced AI as a "hoax." The timing coincides with heightened tensions, as China, through state-owned media, has accused leading American AI companies of using calls for a slowdown to "blunt China’s AI advance."

Neither the United States nor China appears ready to decelerate their AI development, and a lack of mutual trust complicates diplomatic efforts. However, experts suggest that rapidly advancing AI systems necessitate dialogue, urging leaders to avoid setting overly ambitious or overly modest goals for such discussions.

A more effective near-term objective for the U.S. and China, according to analysts, would be to pursue AI safety in parallel and demonstrate progress through practical technical exchanges. This approach could be anchored in sharing insights on how to make AI systems safer, irrespective of national development. Establishing a protocol for communicating about emerging AI incidents before they escalate into international crises is also proposed as a critical step.

The U.S. government is encouraged to promote technical exchanges focused on testing AI models for dangerous capabilities. These include systems that could assist in developing bioweapons or evade human control. American AI companies have been studying these risks, and China has begun incorporating safety testing into its technical standards. Exchanging best practices for testing and mitigating these risks could lead to safer AI systems in both countries.

This proposed diplomacy would not rely on trust or strict reciprocity but would be driven by mutual self-interest. As AI models become more powerful, both nations have a vested interest in ensuring these systems remain controllable. This necessitates improvements in model testing, monitoring, and safeguards, along with sharing findings that enhance safety without compromising competitive advantages or national security.

Crisis Communication Protocol

The proposed dialogue should also establish a government-to-government crisis-communication protocol. Such a framework would address scenarios like an AI model escaping control and hacking critical networks, or an AI-powered drone causing an incident that could be misinterpreted as deliberate. Establishing clear communication channels is crucial for determining whether incidents are accidental or state-sanctioned.

Designing an effective communication channel requires careful consideration. Key elements include defining what constitutes a notifiable incident, identifying designated contacts, and determining how to signal an unauthorized event without revealing sensitive intelligence. Technically informed diplomacy will be necessary to establish even provisional answers to these questions.

Experts suggest that initial incident information should be transmitted in writing, rather than solely by phone. This approach allows Chinese committees time to analyze information and formulate responses, reducing the risk of miscommunication inherent in real-time calls, especially given China's committee-based decision-making processes. While China has a history of suspending crisis communication channels during tense periods, a written protocol could bypass the issue of refusing phone calls.

Navigating the Competition

Discussions should commence without preconditions or concessions related to the broader technology competition. If Beijing attempts to link AI safety progress to relief from U.S. export controls on advanced chips, Washington should maintain the separation of these issues. The objective of AI safety diplomacy is not to transform the U.S. and China from competitors into partners, but to prevent their competition from resulting in undesirable outcomes for either side.

Currently, the political conditions for a sweeping agreement are not present. Trump's rejection of calls to slow AI development and China's interpretation of U.S. pacing proposals as technological containment highlight these challenges. Neither nation appears ready for broad restrictions on frontier AI development, and the U.S. has only begun to regulate its own AI sector.

Starting with a narrower focus does not imply fatalism. The political landscape surrounding AI is rapidly evolving, and advancements in capabilities will continue to influence calculations in both Washington and Beijing. The upcoming talks are unlikely to be definitive, but a limited beginning could foster momentum, building channels, expertise, and interaction habits that may prove increasingly valuable as AI development progresses and concerns about severe risks grow.