Can the United States and China simultaneously cooperate and compete on AI? That is the question driving debates in Washington about the wisdom of pursuing talks on the subject with Beijing. The Biden administration paired sweeping technology restrictions with a narrow dialogue on risk, and the Trump administration has kept a channel open even as it has downplayed safety. Some experts go further. They hope that cooperation on safety can be siloed off from competition entirely, a belief as appealing as it is fanciful. Any dialogue on AI risk takes place within the broader strategic competition between the United States and China. Beijing has long appreciated this dynamic. Much of Washington has not. Whereas the United States has shifted its approach to AI competition and safety across and within presidential administrations, Beijing has kept its AI diplomacy focused on gaining competitive advantage over the United States. China has taken part in international AI safety discussions, but it has used them, along with talks with the United States, to paint U. S. export controls as the main obstacle to cooperation. It has cast itself as the champion of the developing world, promising a Chinese-led AI ecosystem without technological or financial barriers as an alternative to a restrictive U. S.-led AI order. The pitch is working. A growing number of states are embracing China’s message, and Chinese open-weight models are gaining users across the developing world. If the United States wants to maintain its lead and make meaningful progress on safety, it needs to treat the dialogue as part of the competition with China. To do so, it must ramp up pressure on China outside the talks and offer countries still choosing sides a more compelling vision for AI adoption and governance. And when U. S. and Chinese officials next meet, in November, Washington should test Beijing’s seriousness about safety. In short, the United States should approach AI diplomacy with the same strategic clarity China has. CHIPPING AWAYThe Biden administration’s first round of advanced semiconductor export controls in October 2022 and the release of ChatGPT a month later signaled that the United States had taken an early lead in the AI competition and would act to weaken China’s AI industry. Washington and its allies also seized the diplomatic initiative, taking the first steps toward global AI governance. In May 2023, the G-7 began intense discussions on AI governance that would eventually produce a code of conduct for developers. Weeks later, the United Kingdom announced it would host the first global AI safety summit at Bletchley Park. China responded by rolling out a diplomatic script it has been honing ever since. Its core tenets first appeared in the Global AI Governance Initiative, published in October 2023. It called for countries to “put equal emphasis on development and security” (though it left security to be defined by individual states) and contrasted China’s focus on cooperation and development with U. S. export controls, attacking Washington for “creating barriers and disrupting the global AI supply chain through technological monopolies and unilateral coercive measures.” Despite Beijing’s growing body of AI regulations at home, the initiative offered little more than slogans on the technology’s risks. That basic formula—cooperation and development, set against U. S. restrictions, with little to say about risk—has guided China’s AI diplomacy for the last three years. In March 2024, the UN General Assembly adopted a U. S.-led resolution on safe, secure, and trustworthy AI intended to appeal to both developed and developing countries. China supported the resolution, then introduced another resolution centered on help for poorer countries and “emphasizing that member states should enjoy equal opportunities” to develop and use AI. At a UN Security Council meeting on AI chaired by the United States in December that year, China’s UN ambassador, Fu Cong, once again counterposed Beijing’s stated interest in cooperation against “discriminatory barriers based on ideological differences.” In thinly veiled terms, Fu was charging that U. S. restrictions on China harmed the developing world. Much of the allure of China’s pitch rests on how the country’s developers share their technology. To win the support of these countries as they built out their AI infrastructure, China needed a bloc of its own. In July 2025, Premier Li Qiang proposed the World Artificial Intelligence Cooperation Organization (WAICO), which promised developing countries technology, training, and a greater say in writing global AI rules. In a speech marking its launch a year later, Chinese leader Xi Jinping called for “openness,” shorthand for access to U. S. chips and the adoption of Chinese open-weight models. After urging countries to take the risks of AI seriously, Xi added that they should, “in the meantime,” jointly oppose “overstretching the national security concept in the field of AI and placing one country’s security over that of others.” Cooperation on risks, in Xi’s telling, was subordinate to overcoming U. S. restrictions. WAICO has quickly gained standing among its target audience. An early endorsement by UN Secretary-General António Guterres lent it immediate credibility in the eyes of developing countries, and Chinese officials have since promoted WAICO in frequent speeches and in bilateral meetings with countries as varied as Brazil, Pakistan, and Serbia. Beijing has signaled that membership will come with perks. Its draft export controls on leading models, released in October, promise to help WAICO’s 37 members build their AI capabilities, suggesting members could have easier access to models restricted for the rest of the world. Xi personally led this diplomatic offensive in the lead-up to the Washington summit with U. S. President Donald Trump in September. During a speech in Kyrgyzstan, he pledged to expand a Chinese AI center for the Shanghai Cooperation Organization that helps members build local language models and train engineers. In Cairo, he and Egyptian President Abdel Fattah el-Sisi signed agreements on AI, data centers, and semiconductors, which have become a staple of Xi’s trips abroad. And in New Delhi, Xi proposed an open-source AI zone for BRICS members and told other leaders that a “small yard, high fences” approach would only “box oneself in and divide the world.” AI risks were nowhere on the agenda. The message instead was that U. S. restrictions would isolate Washington, not Beijing. Much of the allure of China’s pitch rests on how the country’s developers share their technology. Chinese developers release their best models with open weights, which any government or company can download, run on its own hardware, and adapt to its own language. For governments wary of depending on American companies whose exports Washington can restrict, open-weight models promise a greater degree of control. Chinese developers rarely conduct or publish the results of any safety testing for these models, leaving others to discover after release whether, for example, they can be used to launch cyberattacks or build weapons. But their accessibility allows Beijing to argue that any effort to contain the risks of Chinese models is an attempt to deny AI to the developing world. CHANGING THE SUBJECTChina’s approach to bilateral AI talks with the United States has deflected discussions of risk just as assiduously as its global AI diplomacy has. In May 2024, the United States and China held talks in Geneva, which I joined as part of the U. S. delegation. The U. S. objective of the talks was to test whether an in-depth technical exchange with China on AI risk was possible. The answer was no. Chinese diplomats steered the discussion away from risks for hours and declined to even explain China’s own AI regulations beyond a superficial overview, despite having technical experts from the relevant agencies in the room. The Chinese position came through most clearly in the opening and closing remarks of the delegation’s most senior officials. They began with an appeal for cooperation on capacity-building projects in the developing world. They then cast U. S. technology restrictions as an attempt to contain China’s development and the U. S. emphasis on safety as a barrier preventing developing countries from accessing and using AI. The implication was clear—export controls were the primary impediment to cooperation on AI safety. The one concrete outcome, a November 2024 statement in which former President Joe Biden and Xi affirmed “the need to maintain human control over the decision to use nuclear weapons,” asked little of Beijing. This September, after the summit between Trump and Xi in Washington, China’s Ministry of Foreign Affairs published an “eight-point consensus,” Beijing’s account of what the two sides had agreed to. It described the dialogue as an exchange on the “risks and benefits” of AI. But whereas the White House framed the dialogue as a security issue, listing it alongside nuclear arms control, China’s Ministry of Commerce stressed instead that talks had taken place, and would continue, under the two countries’ economic and trade mechanism. Folded into trade talks, a dialogue on AI risks and benefits leaves Beijing free to parry U. S. requests on risks, such as limits on openly releasing models with dangerous capabilities, by demanding that Washington relax export controls. Xi did propose to “jointly prevent the misuse and abuse of AI,” conveniently eliding China’s own frequent use of AI against the United States in various ways, such as hacking campaigns and influence operations. He added that AI should remain under human control and be developed for good, a principle Beijing has repeated since 2023 but about which it has offered little further detail or follow-through. More pointed was his insistence that future dialogue should “draw on each other’s strengths, not guard against each other” and his request that Chinese companies be “treated fairly,” both implicit rebukes of U. S. restrictions. JUST WIN, BABYThis record suggests Beijing will come to future meetings, beginning with the next discussion in November, with proposals designed to put Washington in a bind. Although U. S. Trade Representative Jamieson Greer has made clear that export controls are not on the table, Beijing will raise them anyway. If Washington engages, Chinese negotiators will claim a standing channel where U. S. restrictions on AI technology are up for negotiation. If it refuses, Beijing will continue to evade U. S. entreaties on safety while accusing Washington of stifling opportunity for developing countries. It hopes that every round of the dialogue will further cement an impression among allies such as Japan and the Netherlands that Washington will trade away restrictions over their heads, and that they should accommodate China’s AI ambitions, too. But Washington is in a stronger position than its diplomacy has let on. The world’s leading models are still American, and, together with allies, the United States controls key chokepoints in the semiconductor supply chain. Export controls on those chokepoints are not a bargaining chip and should be strengthened. Chinese AI companies regularly lament their lack of high-end chips and have resorted to smuggling chips and illicitly distilling U. S. models. U. S. developers also conduct increasingly extensive and transparent safety testing, which Chinese developers avoid. Taken together, these strengths give Washington the means to press Beijing and an attractive offer for the rest of the world. It should use them, advertising its rigorously tested models to compete for global adoption while pressuring China on safety in the dialogue. The first chance to apply that pressure is the next bilateral meeting. Washington should set two conditions for it and future meetings. Chinese developers must commit to publishing safety evaluations at least as extensive as those that leading U. S. developers publish today on equivalent benchmarks for the most serious risks, such as cyberattacks and biological weapons, and to disclosing previous and future incidents. The United States can make the same commitments, as its companies already publish such evaluations, increasingly disclose incidents, and open their models to independent and government evaluators before release. Otherwise, the incident channel to which the two sides agreed in September will run one way, with China free to point to U. S. incidents while reporting none of its own. America must offer countries a more compelling vision for AI adoption and governance. These conditions would change the terms of the dialogue. Since 2023, Washington has raised risks and Beijing has steered the conversation toward development. As long as Chinese companies face no scrutiny for the risks of their models, Beijing can tout them for their openness and innovation without having to answer questions from the rest of the world about their safety. Concordia AI, which tracks China’s AI ecosystem, found that only five of China’s ten leading developers published safety test results for any model over the past year, and none did so consistently. If China agrees, governments and companies choosing which models to use will be able to judge Chinese and American developers by the same criteria. If China refuses, the onus will be on Beijing to explain why fair treatment for Chinese companies should not extend to safety testing. Regardless, the U. S. government should ramp up its own tests of new Chinese models and publish the results. It should also avoid unforced errors, such as Trump’s waving away of AI safety concerns as a “hoax,” which hand Beijing an easy retort to U. S. appeals. More broadly, effective AI diplomacy ultimately requires convincing other countries to join the American AI ecosystem. Countries that train their engineers, build their local language models, and run their data centers on Chinese technology will come to depend on it. To counter that dependence (and prevent it from turning into political alignment), Washington needs to present a positive vision of AI governance that pairs verifiable safety commitments—the kind Chinese developers avoid—with the access and investment developing countries want. U. S. companies already offer the best technology, but Washington has yet to put real diplomatic weight behind them. It should draw from its evolving approach at home to rally partners around shared safety measures. Under the White House Accord on Super Intelligence, signed in September, six leading U. S. companies committed themselves to internal controls, external audits, and independent oversight. The G-7 Code of Conduct, which built on voluntary commitments secured under the Biden administration, could incorporate the accord’s measures and serve as a common standard for partners. Countries that agree to its measures could receive priority for technology and financing. Without such incentives, the AI competition may be determined by price and access alone, where China has the edge. Much of Washington has treated talks with Beijing as a place to manage the shared risks of AI apart from the competition. China has always treated them as an arena of the larger rivalry. To maintain its lead and make real progress on safety, the United States should start doing the same.