U.S.-China cooperation on artificial-intelligence governance can produce meaningful reductions in cross-border AI risks

What's this about?

People disagree about whether U.S.-China teamwork on AI can lower dangers that cross borders. AI means computer systems that can learn and make choices.

What supporters say

• Crisis phone lines and clear ways to talk could stop a mistake from growing into a bigger fight.
• Shared safety tests could show both countries the same dangers, failures, and warning signs.
• Common safety ideas could shrink gaps between their rules, even if their laws stay different.

What critics say

• The two countries want different things from AI rules, so shared rules may stay limited.
• Their rivalry makes them less willing to share key facts about AI risks and failures.
• So far, no clear proof shows that teamwork has lowered real cross-border AI harm.

How to read this

A longer list does not prove a side right; check how strong the proof behind each point is.

The bottom line

The idea could work in small areas, such as talks, shared tests, and safety rules. However, broad success remains unproven.

The critics have stronger support: rivalry and different goals make large gains hard to show.

The claim is that cooperation between the United States and China could reduce some of the risks posed by advanced AI across national borders. The current record suggests that this is plausible in limited areas, but not yet demonstrated on a broad scale.

The case for

Shared testing and technical standards could make AI risks easier to identify and manage. Joint or coordinated testing could help both countries compare how systems behave, share information about failures and incidents, and develop common methods for evaluating safety. The International Scientific Report highlights knowledge sharing, coordinated testing and incident response as useful ways to deal with uncertainty. The Bletchley Declaration and Seoul Leaders’ Statement also support risk assessment, interoperability and safety evaluation. (see Figure 2)

Common frameworks could help even if the two countries do not adopt identical laws. For example, the voluntary U.S. National Institute of Standards and Technology AI Risk Management Framework offers a shared vocabulary for identifying, measuring and managing risks. Such tools could make it easier for companies and regulators in different countries to understand one another’s safety practices. 1

Direct communication could reduce the risk of dangerous misunderstandings. AI-related accidents or security incidents could quickly affect more than one country, particularly when systems are linked to critical infrastructure or military decision-making. Research finds that U.S. and Chinese officials share concerns about some extreme AI risks, and the countries began a working-level dialogue in 2024. Those developments make limited confidence-building measures possible, even though they do not prove that the channels have worked. 2

High-level agreements may also reduce gaps between national rules. The OECD AI Principles, the Bletchley Declaration and a United Nations General Assembly resolution show broad support for ideas such as safety, transparency, accountability, cooperation and capacity building. This suggests that some common ground is possible despite major political differences. 3

The case against

The strongest objection is that there is no public evidence yet that cooperation has caused measurable reductions in cross-border AI harm. Existing agreements mainly establish commitments and communication channels. They provide limited monitoring, enforcement or compliance requirements. The bilateral dialogue has not publicly produced binding controls or demonstrated reductions in incidents, misuse or wider systemic risk. 4

Strategic rivalry makes meaningful information sharing especially difficult. Export controls, investment restrictions, competition over supply chains and national-security concerns have reduced trust and limited technical exchanges. These barriers are important because effective joint safety work may require sharing information about models, capabilities and failures. Dialogue by itself does not overcome those pressures. 5

Regulatory differences create another obstacle. China’s generative-AI rules impose obligations involving security, data, content and risk management, while comparative legal research describes growing fragmentation and competition among major powers. Different goals and requirements can produce loopholes or inconsistent safeguards when common principles are not backed by similar implementation. 6

Cooperation therefore appears more realistic in areas such as incident terminology, evaluation methods, knowledge sharing and emergency communication. It is much harder where sovereignty and military advantage are directly involved, including frontier-model access, military applications, export controls and binding limits. As AI capabilities and computing power grow, the value of cooperation may rise—but rapidly changing technology will also make rules harder to update and verify. (see Figure 1)

The evidence itself has important limits. Advanced-AI risks are uncertain, fast-moving and often underreported, so the absence of known incidents cannot show that cooperation is working or failing. There is no systematic bilateral dataset linking particular cooperative activities to changes in incident frequency, severity, reporting, model behavior or cross-border effects. Political incentives and possible conflicts of interest may also affect what governments report and how outcomes are judged.

The bottom line

The evidence favours a narrow version of the claim, but only weakly. U.S.-China cooperation could plausibly reduce some specific cross-border risks, especially through shared testing, common terminology and crisis communication. But the record is much stronger on proposed mechanisms, public declarations and institutional contacts than on real-world results.

Strategic rivalry and different governance objectives are better-supported obstacles than cooperation is a proven solution. Overall confidence is low: meaningful reductions may be possible, but broad or measurable reductions in AI risk have not yet been demonstrated.