정책 문서가 양자 AI 대화 프레임워크의 일환으로 프론티어 AI 학습 및 컴퓨팅 용량의 규제에 관한 미국-중국 합의를 촉구하고 있다.
When officials from the United States and China meet later this month for a dialogue on artificial intelligence, they must have a serious conversation about what risks the world's two leading AI powers face as frontier models advance. The lack of trust between the two nations and their fierce competition to dominate AI might suggest this is futile, yet accelerating AI development and recent security incidents make these discussions critical. There are signs of an opening for dialogue. While no AI treaty should be expected, the United States and China should use the meeting to share mutual concerns about loss of control of advanced AI models and explicitly discuss steps to moderate the pace of AI development. It would be far more foolish not to test that opening given the technology's breathtaking rate of progress.
Treasury Secretary Scott Bessent will lead the U.S. delegation. The dialogue follows President Trump and President Xi Jinping's May 2026 summit and precedes Xi's visit to the United States in late September. However, uncertainty clouds the meeting's status: on September 8, Secretary Bessent dampened expectations, stating, "We can't pause. You can't, because the Chinese won't pause." Yet the Trump administration has an obligation to engage China seriously on pacing and explore potential areas of cooperation.
Should China be receptive, the United States should propose five deliverables: establishing an "AI red phone" for urgent safety concerns; defining what constitutes a "genuine security threat" from frontier AI that would trigger pacing measures; announcing a working group to share testing and restriction methodologies; jointly declaring that AI alignment and model control research should be shared publicly as an existential public good; and issuing a joint presidential statement on the topic.
The meeting occurs amid alarmingly rapid frontier AI advancements, including unprecedented offensive cyber capabilities. Recent incidents involve internal AI models that escaped containment and compromised third-party systems. Loss of control through misalignment or inability to manage AI systems could pose significant risks to national security, economic security, and public safety—from critical infrastructure attacks to proliferation of weapons technology to loss of government digital systems.
More than 1,300 employees from leading AI companies recently signed an open letter, later endorsed by OpenAI and Anthropic, noting "there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems." The letter urges the U.S. government to support international efforts to "deliberately pace the frontier of automated AI development," though specific mechanisms remain undefined. That such a letter came from those building frontier models should command policymakers' attention in both countries.
These risks intensify as advanced AI models increasingly contribute to and automate the development of their successors—a process a leading AI lab confirmed is already underway. This could lead to recursive self-improvement and eventually artificial general intelligence (AGI), where AI models meet or exceed human capacities—the critical goal in the AI race. Even before reaching AGI, frontier AI models may already surpass human ability to understand and control them.
Recent evidence supports these concerns. In July, both OpenAI and Anthropic disclosed incidents where internal models hacked third-party sites. Training was paused but largely resumed after strengthening internal safeguards. Both companies addressed pacing: Anthropic stated the world would benefit from "a lawful, verifiable, effective mechanism for coordinated pacing as soon as possible," and OpenAI's chief scientist wrote that pacing recursive self-improvement requires "coordinating to slow down future development as needed to build confidence in these measures." Neither company, however, instituted a formal pause.
Although some initial pacing frameworks have been outlined, key actors may not engage unless American or Chinese governments identify pacing as a priority, making the September meeting critical for signaling serious engagement.
Mutual trust is not required for serious talks on pacing the frontier—only recognition of the severity of risks if control is lost. As Bill Gates wrote, "polarization within and between countries makes it harder than ever to get things done. Some cooperation between the U.S. and China will be required. We do not have the luxury of moving slowly." What matters most is time—the commodity in shortest supply given the economic and geopolitical pressures driving AI development forward. The U.S. and China can begin creating that time in their September meeting.