How to Make the U.S.-China AI Race Less Dangerous

The timing could hardly be worse. Even as Treasury Secretary Scott Bessent prepares to meet Chinese Vice Premier He Lifeng this weekend to discuss AI, President Donald Trump has dismissed public warnings about runaway AI as a “hoax.” Meanwhile, an editorial in the state-owned China Daily accused America’s leading AI companies of using calls for a slowdown to “blunt China’s AI advance.”

Neither the United States nor China is ready to slow down. Neither trusts the other. And rapidly advancing AI systems will not wait for diplomacy to catch up. But these realities do not mean Trump and Xi can ignore the mounting calls for the U.S. and China to talk about AI. Any such dialogue must avoid the two most common pitfalls for U.S.-China diplomacy: setting the ambitions so low that the talks become meaningless or so high that agreement becomes impossible.

Aim too low, and the dialogue becomes a photo op followed by an anodyne communiqué about promoting innovation and reducing risks. Aim too high, chasing sweeping restrictions on frontier development that neither side has the political trust or the capacity to verify and enforce, and the meeting is set up for failure. The former path fails to meet the stakes of the moment; the latter fails to recognize its constraints.

The path to AI safety

A better near-term goal would be for the U.S. and China to pursue AI safety in parallel and demonstrate their progress in practice. That effort should be anchored in technical exchanges that make each country’s AI systems safer on their own terms. The two countries should also establish a protocol to communicate about emerging AI incidents before they spiral into international crises. These steps would represent meaningful progress, far more valuable than lofty statements or unverifiable commitments.

The Trump Administration should push for technical exchanges on how to actually test models for dangerous capabilities and behaviors that neither country wants to see released into the wild. These include AI systems capable of helping amateursdevelop bioweapons, or evading human oversight and control. American AI companies have been studying these risks for years, and they have also begun to appear in Chinese technical standards related to safety testing. Exchanging best practices and new insights about how to test for and mitigate these risks could render the systems built in both countries meaningfully safer.

These proposed technical exchanges would not depend on “trust” or even strict “reciprocity” between the two superpowers, nor would they require identical conceptions of AI risk. They would be driven by self-interest. As the most advanced AI models grow even more powerful, both countries have an interest in ensuring that these systems remain controllable, regardless of where they are built or deployed. That will require both countries to improve their model testing, monitoring and safeguards, and share promising findings that could enhance safety without surrendering a competitive edge or undermining national security.

This diplomacy is less like negotiating a sweeping arms-control treaty than establishing shared safety practices and procedures for a technology neither government fully understands or controls. The premise is not that Washington should trust Beijing, or vice versa, but that neither country benefits if the other loses control of a powerful system.

For this approach to work, it must be accompanied by a government-to-government crisis-communication protocol. What if an American AI model escapes its company’s control and hacks the networks of a Chinese company or the Chinese government? What if a Chinese AI-powered drone attacks a U.S. Navy vessel, and neither side can immediately determine whether the incident was accidental or deliberate? What if Washington detects a Chinese model conducting a dangerous cyber operation but cannot tell whether Beijing authorized it or whether a company has lost control of it?

These are not the kind of existential risks President Trump has dismissed, but they are close cousins of incidents already unfolding in real time. The two countries need to lay the basic groundwork for dealing with them. Whether the communication channel succeeds or fails will depend on its operational design. What constitutes a notifiable incident? Who are the designated contacts? How do the two sides signal that an incident was unauthorized without revealing sensitive intelligence? Serious, technically-informed diplomacy will be required to reach even provisional answers to these questions.

One essential feature of the channel should be that information about an incident, at least initially, be transmitted in writing rather than solely by phone. Real-time calls are often poor vehicles for conveying technical information with a Chinese system that runs on committees and documents, increasing the risk of miscommunication and misunderstanding.

Sending a message in writing gives the relevant committees in China the ability to disseminate and analyze it and formulate a response that no individual official in China would feel empowered to give. It also reduces the recipient’s ability to shade or distort the message. Even this approach will be challenging, given China’s spotty record of suspending crisis-communication channels—or simply refusing to use them—precisely when tensions are highest. But it would eliminate one of the stumbling blocks: the refusal to pick up the phone.

How to talk to Beijing

The talks should begin without preconditions or concessions in the broader technology competition. If Beijing seeks to make any progress on AI safety conditional on relief from U.S. export controls on advanced chips and chipmaking equipment, Washington should keep the two issues separate. The goal of AI safety diplomacy is not to turn the United States and China from intense competitors into cooperative partners. It is to keep their competition from producing outcomes neither side wants.

The two sides are nowhere near the political conditions necessary for a sweeping agreement. Trump has rejected calls to slow American AI development, while Beijing increasingly interprets American pacing proposals through the lens of technological containment. Neither side seems prepared to accept broad restrictions on frontier development, and the United States has barely begun to regulate its own AI sector, which is a far more important step toward addressing AI risk than what diplomatic engagement can deliver. That is precisely why the two countries should begin with something narrower.

Starting modestly does not mean succumbing to fatalism. The politics of AI have shifted dramatically in just a few weeks, and advancing capabilities will continue to alter the calculations in both capitals. The upcoming U.S.-China talks will not be the last word. A limited start could build momentum over time, creating channels, expertise, and habits of interaction that will become more valuable if the trajectory of AI development increases Washington’s and Beijing’s levels of concern about severe risks.

Someday, as unlikely as it may seem now, the two countries may realize that they do not have to agree on who should win the AI race to agree that neither side can afford to lose control of it.