Can the United States and China Compete on AI While Managing Shared Risks?

World leaders will soon gather in New York for the high-level meetings of the United Nations General Assembly’s 81st session. The Iran war, Russia’s war in Ukraine, and their global economic ramifications are set to dominate the agenda. But the risks posed by increasingly powerful artificial intelligence, which have captured global headlines, will likely also factor into the debate. Many of Silicon Valley’s most influential executives and some members of the US Congress are now calling for stronger oversight, though they are unlikely to look to the UN for solutions.

Middle powers are also seeking to shape the response. Canadian Prime Minister Mark Carney is pursuing closer ties with Germany and a deeper partnership with the European Union, where proposed associate membership remains undefined. AI is an important part of this agenda. Meanwhile, the United Kingdom’s King Charles III and the Ditchley Foundation this week brought technology leaders together in Scotland to consider how AI can serve humanity safely.

But can initiatives that operate largely without the leadership of world’s largest powers—the United States and China—really move the needle? There is reason to be skeptical, and also to look to Washington more than New York to determine the future direction of travel. US President Donald Trump will address the UN General Assembly, but the next day he and Chinese President Xi Jinping are expected to meet in Washington, with AI likely to feature prominently.

The political incentives around AI safety are complicated. Shared exposure to risk does not automatically produce a shared political response. Trump has stressed the importance of maintaining America’s technological edge. At the same time, many skeptics assume the seemingly sudden turn from Silicon Valley toward advocating for safety measures is driven less by a genuine concern for threat than by a desire to lock in gains.

With the midterms approaching, however, Americans are increasingly anxious about AI’s consequences. Polling last year by the Chicago Council on Global Affairs found a majority see job loss due to advanced technologies like AI or quantum computing as a critical (45%) or important but not critical (32%) threat. More recently, Pew Research Center finds that 71 percent expect AI to reduce the number of American jobs over the next 20 years. Concern crosses party lines: 75 percent of Democrats and 68 percent of Republicans anticipate fewer jobs. Data centers, and their demands on local electricity supplies, add a more immediate dimension.

Congress is paying attention, but in a polarized political class, concern does not imply agreement. House Speaker Mike Johnson has rejected a moratorium on AI development, warning that it would hand China a competitive advantage, while expressing support for independent auditors and greater transparency. This leaves room for oversight, even as restrictions on development remain deeply contested.

A few proposals are on the table.

Anthropic’s Dario Amodei has proposed embedding outside evaluators inside leading AI companies, giving them sustained access to assess safety practices and report incidents. He also calls for coordination among companies in democratic countries, along with efforts to secure wider international cooperation. Possible agreements range from prohibiting particularly dangerous uses of AI, such as assisting biological weapons development, to common testing requirements before models are released. But limits on the speed of AI development would be substantially harder to negotiate and verify.

A narrower agenda concerns military risk. American and Chinese experts participating in the track II dialogue led by the Brookings Institution and Tsinghua University’s Center for International Security and Strategy have proposed practical measures to reduce the danger of unintended escalation, including procedures for communicating about AI incidents and a pledge against using AI to interfere with nuclear command systems. These remain expert proposals. Their appeal is that both countries could benefit from reducing the chance that an accident is mistaken for an attack. But this still leaves wide open the problem of enforcement.

Middle powers are also trying something different: investing in the design of AI itself. Canada and Germany have announced plans to invest C$150 million and €100 million respectively in LawZero, the nonprofit developing the Scientist AI. The ambition is to build systems that reason from evidence without pursuing goals of their own, with safety incorporated from the outset. This remains a research program, and Germany’s funding is subject to European Commission notification.

The initiative also seeks to reduce dependence on foreign-controlled technology. For middle powers, having greater capacity to develop AI could provide more influence over the terms under which it is used. A successful alternative might eventually affect company behavior through competition. But its international significance will depend on whether the technology works and whether others adopt it, particularly the United States and China.

Even agreement on the word “safety” can obscure important differences. For executives warning about frontier AI, the central concern is often whether increasingly autonomous systems will remain under human control. For many Americans, safety also encompasses economic security: whether their job survives and whether their community bears the costs of new infrastructure. An agreement addressing catastrophic technical risks would leave many of these everyday concerns unresolved.

And could cooperation survive disagreements over access to advanced technology? Amodei’s own proposal illustrates the tension: He advocates cooperation with China while supporting tighter restrictions on its access to advanced chips. Beijing could interpret rules proposed under these conditions as an effort to preserve American advantage. A limited agreement on dangerous uses may therefore be more achievable than a bargain over the pace of development.

This is where the distinction between an announcement and an arrangement becomes consequential. A statement following the upcoming US-China summit expressing concern would tell us little about what happens next. Commitments require enforcement mechanisms. Pledges will not be enough. More meaningful progress would identify which institutions are responsible, establish mechanisms for coordinating among them, and establish a timetable for action. Company commitments would need scrutiny that reaches beyond information companies choose to disclose. Governments would need procedures that officials can actually use when an incident occurs.

The test would be whether either side changes a deployment decision, shares information it would otherwise withhold, or responds differently during a crisis. There would also need to be a process for addressing suspected violations without abandoning cooperation altogether.

Public concern, executive warnings, and royal convening may help create political space. Middle powers can contribute research and build support for common standards. Whether this produces durable restraint will depend heavily on Washington and Beijing. As world leaders gather in New York, the question is actually whether Trump and Xi, meeting in Washington, can translate a shared interest in avoiding catastrophe into commitments they are prepared to honor.