The United States has proposed that Washington and Beijing create an AI safety notification mechanism for incidents that rise to a national-security level. The idea is not a treaty, a ban on AI, or an agreement China has accepted. It is a proposal for the two governments to communicate when high-risk AI events could threaten both countries—a narrow but potentially important step between the world’s two largest AI powers.
At a glance
- US Treasury Secretary Scott Bessent said the proposal was discussed with Chinese Vice Premier He Lifeng in New York on September 20.
- The proposal concerns national-security-level AI incidents and a possible formal US–China AI dialogue.
- China’s response was not publicly confirmed, and advanced-chip export controls were not part of the proposed mechanism.
What was proposed?
Reuters reported that Bessent proposed a notification system for common goals and threats involving AI-related incidents that reach a national-security threshold. He said the two sides also discussed creating a US–China AI dialogue for President Donald Trump and President Xi Jinping to consider at their coming summit.
That wording matters. A notification mechanism would be a channel to share information after or during a serious incident; it would not automatically regulate private companies, require either country to reveal sensitive model details, or settle wider disputes over trade, military technology and data.
What could count as an AI incident?
The officials did not publish a detailed definition. In practice, a system like this could be relevant when AI is implicated in a cyberattack affecting critical infrastructure, a dangerous loss of control in a high-consequence setting, or an incident involving nuclear, biological, financial or military security risks. The details would be the hardest part: each government would need to decide what must be disclosed, how quickly, through which officials, and what information can remain classified.
Existing crisis channels between governments show why narrow communication can still be valuable. When a fast-moving incident is misunderstood, leaders can assume intent where there was an accident, criminal activity or a third-party attack. A reliable contact point does not remove strategic competition, but it can make escalation less likely.
Why are the talks happening now?
AI has become a central part of US–China competition, alongside semiconductors, cloud capacity and military technology. The countries are advancing different policy approaches while retaining deep economic ties. Reuters reported that the talks also covered possible tariff reductions for a limited set of non-strategic goods and preparations for the Trump–Xi meeting.
The AI discussion was deliberately separate from US controls on sophisticated AI chips and semiconductor-manufacturing equipment. That separation is a reminder that safety cooperation and technology competition can proceed at the same time—and that progress in one area does not mean the other disagreements are resolved.
Why it matters
An AI incident-alert system would test whether two rivals can build limited safety guardrails without first agreeing on the future of AI. The value would depend on operational details, not the headline: who gets notified, what triggers an alert, how false alarms are handled and whether messages reach decision-makers quickly.
It also matters beyond Washington and Beijing. AI systems, software supply chains and cloud services cross borders. A serious failure can affect companies and users far from the country in which it began. The early aim of a notification channel is therefore less about solving every AI governance question than reducing the risk that one extreme event becomes a wider international crisis.
For the business side of the technology race, see our analysis of CXMT’s fifth-generation DRAM platform. For a recent example of why secure deployment matters, read our report on the Google Gemini cybersecurity incident.
What happens next?
The immediate question is whether China agrees to consider the proposal and whether the upcoming leaders’ meeting produces a formal mandate. If it does, negotiators would still need to define scope, confidentiality, verification and links to existing diplomatic and cybersecurity channels.
Until then, the accurate description is a US proposal—not a deal. That distinction is important for readers and investors following a relationship where a single announcement can have implications for AI policy, chip supply chains and global markets.
