Donald Trump wants the US to beat China in the AI race. But as he meets with Chinese President Xi Jinping, both countries are considering how to address the risks posed by rapidly advancing technology.
In recent months, AI systems have taken more serious actions online. OpenAI announced that one of its research models had been compromised as part of its AI platform Hugging Face, while Anthropic announced it had discovered that AI was being used to automate some of its cyberattacks. The incident has raised concerns that increased autonomy in models could make attacks faster and harder to contain.
Ahead of Thursday’s meeting between President Trump and President Xi Jinping, U.S. Treasury Secretary Scott Bessent discussed with Chinese Treasury Secretary He Lifeng over the weekend the prospect of a “U.S.-China AI dialogue,” including channels for incidents “up to the national security level.”
Arok Mehta, director of the Wadhwani AI Center at the Center for Strategic and International Studies, told CNBC that “leaders in both countries are fearful of the growing cyber capabilities of AI models, especially agents acting autonomously to evade containment and infiltrate websites.”
But while fear has intensified, tensions have not subsided.
The U.S. government continues to restrict China’s access to Nvidia’s cutting-edge AI chips, accusing Chinese AI companies of using “distillation” (training models based on the output of more advanced U.S. systems), a claim Beijing rejects.
What has been said so far?
The United States took an early lead in the development of large-scale language models (LLMs) thanks to companies such as OpenAI and Anthropic, but in recent months the Chinese model has significantly improved its capabilities and is increasingly being adopted by companies around the world, including in the United States.
President Trump has previously dismissed warnings that AI poses an existential threat to humanity, but the risks have been amplified after OpenAI CEO Sam Altman and Anthropic head Dario Amodei called for a slowdown of the industry and regulatory oversight.
“If we don’t beat AI, we’re going to be in a very bad position,” President Trump told reporters earlier this month.
But in a post on Truth Social on Monday, he said the government “will rein in the situation if necessary.”
AI safety and chip access
Mehta told CNBC that the areas where the US and China are most likely to agree could be a common definition and framework on AI safety for powerful models, and an emergency communication mechanism to discuss incidents.
Chris McGuire, senior fellow for China and emerging technologies at the Council on Foreign Relations, told CNBC that a channel linking national security incidents would be a “modest” but positive step.
But he added that it was “very unlikely” that there would be an agreement to slow the breakneck pace of AI development.
McGuire said China has historically used AI dialogue “to complain about U.S. export controls,” and the U.S. has used it to prevent cutting-edge AI chips from being sold to China.
“Chinese state media has largely dismissed concerns expressed by major U.S. AI research institutes about potentially catastrophic risks, consistent with China’s long-standing skepticism of all U.S. arms control proposals,” it added.
Despite US restrictions, several Chinese companies were reportedly able to access the chips’ computing power remotely through data centers in Southeast Asia.
U.S. Trade Representative Jamison Greer said chip restrictions were not on the agenda during preliminary talks over the weekend, but Melanie Hart, senior director of the Atlantic Council’s Global China Hub think tank, said China could push for less regulation.
“China is asking the United States to adopt China’s AI model and lift some U.S. restrictions on China’s access to semiconductor chips,” he told reporters on Tuesday.
Distillation may also be a subject of discussion. U.S. companies and officials have accused Chinese labs of profiting from the practice, calling it theft. China denies this claim.
In July, Bessent threatened sanctions against companies engaging in such behavior, saying the Trump administration had found “the watermarks of large American language models on many Chinese models,” which he called “unacceptable.”
Mehta said progress in AI collaboration depends on two things. “First, it does not undermine the ability of the United States and China to compete on technological merit.”
The second is a verification mechanism, he added. “Policymakers are not going to be willing to take the other person’s word for it, so they need to find ways to ensure that the information flowing between the two countries is accurate, and that means both technical and governance validation tools.”
