The most concrete outcome from last month’s summit between US President Donald Trump and Chinese President Xi Jinping was an apparent agreement to set up a US-China artificial intelligence ( AI ) incident-reporting channel. But even if either government proves willing to use this direct line or start sending data across the Pacific, they must first address a more pressing issue: US and Chinese frontier AI companies currently do not provide the information they would need.
In the US, no federal statute or regulation specifically requires AI companies to report safety incidents and dangerous capabilities to the government. Although some states have passed incident-reporting laws, federal legislation – such as the AI Incident Reporting Act and the Frontier Act – has stalled in Congress. The industry is still largely regulating itself.
Moreover, while the White House Accord on Super Intelligence pledges to identify and remedy safety issues, it does not mandate notification to regulators or set reporting deadlines. The Trump administration’s separate Super Intelligence Force is developing government reporting mechanisms for breaches, hacks and other AI incidents. But it is not a regulator, and its recommendations would need to be implemented through enforceable measures.
There is also a severe lack of funding for federal government testing and oversight. According to Stanford’s 2026 AI Index Report, US developers last year released 59 notable AI models, while British developers released just one. Yet, the main US federal AI-testing agency, the Commerce Department’s Center for Advancing Innovation and Standards for Super Intelligence ( CAISSI ), has an annual budget of only US$15 million, whereas the United Kingdom’s AI Security Institute has a budget of almost US$90 million.
Given that the White House Accord seems to envision frontier labs doing the bulk of safety testing themselves, it remains to be seen what role the government would play in validating such testing, and whether any government body would have access to sufficient compute to conduct its own testing.
Unlike the US, China’s government doesn’t face political bottlenecks, and it has already built a significant regulatory structure, with the Cyberspace Administration of China ( CAC ) recently calling for behavioural controls, model traceability, security testing and safeguards against system vulnerabilities and loss of control. But there is still no Chinese equivalent to CAISSI – a specialized, centralized AI-safety body with a mandate. The government has plenty of authority over tech companies, but oversight is dispersed across the CAC, the Ministry of Industry and Information Technology, the National Development and Reform Commission ( NDRC ), the Ministry of Science and Technology and other agencies.
Evaluating the risk of frontier AI models requires technical expertise and coordination among all the relevant Chinese government departments, including the security services. But some Chinese government agencies lack sufficient technical capacity, and many of the people with the deepest technical expertise sit in private research institutes with no authority to demand access to a company’s most advanced models, let alone inspect its internal logs or force it to make changes.
For now, the government has deployed officials from the CAC and NDRC to Chinese labs to discuss possibly gating the release of the most advanced models. But the authorities eventually will need to mandate internal safety testing teams, access to sufficient compute, and pre-release testing of the most advanced models.
With the next Sino-American AI-safety dialogue scheduled for November, the US and China should focus on what would make an incident-reporting channel more likely to work. For its part, the US government needs a transparent and mandatory policy for testing frontier models. Either the White House should disclose its current framework and make it binding, or it should work with Congress to pass legislation like the bipartisan Frontier Act. The federal government also needs to set rules for external audits of AI companies, giving auditors a clear mandate, reporting obligations and meaningful independence ( with auditors selected and paid by an industry body or government agency ).
At the same time, the Trump administration needs to be more proactive in monitoring company compliance. It should use the Super Intelligence Force to review the current, flawed release process for the most advanced models, consider what to do about increasingly capable open-weight models and begin moving toward an incident-response mechanism that has teeth.
On the Chinese side, making the reporting channel operational would require closing the authority expertise gap. The CAC could define safety expectations and reporting arrangements; technical bodies could develop and validate evaluations; the NDRC and other industrial agencies could weigh the effects on investment and diffusion; and the Ministry of Commerce would remain relevant to technology-transfer and export rules. With the US empowering the Super Intelligence Force to conduct a thorough review of the US government’s role, China should consider something similar.
China also needs to address its lack of evaluation capacity and access to sufficient compute across companies, evaluators, and government bodies. Rapid advances in Chinese open-weight models make this particularly important, because once model weights are released, developers have much less visibility into how their systems are being used or misused.
The nascent Z.ai-Concordia framework is a potential template. It sorts models into three categories: green-zone models that are safe to be fully released; yellow-zone models that require a staged release, structured access for vetted researchers, or application programming interface-only hosting while weights are evaluated; and red-zone releases that are suspended pending review.
AI safety is a global challenge. Addressing it depends on close and growing collaboration between the two leading AI powers, and a reporting channel for AI incidents is just a small step in the right direction. It will take the US and Chinese governments a long time to reach binding, verifiable agreements. They cannot afford to be held up by failing to enact domestic reforms.
Lizzi C. Lee is a fellow on the Chinese economy at the Asia Society Policy Institute’s Center for China Analysis, where she leads the Center’s Economics and Technology Pillar; and Simon Hedlin is an AI policy researcher focusing on open-weight models and a former fellow at GovAI.
Copyright: Project Syndicate
