The US Wants China to Report Dangerous AI Incidents
Matthew Mbaka · September 22, 2026 · AI
The United States has asked China to consider a notification system for artificial intelligence incidents that could threaten national security. It sounds a little like a crisis hotline, but no hotline exists yet.
US Treasury Secretary Scott Bessent described the proposal after talks with Chinese Vice-Premier He Lifeng in New York on September 20. According to the Associated Press, Bessent said the two countries should move toward greater transparency as the world's two leading AI powers.
Chinese state media confirmed that the talks included AI and described the discussion as candid and constructive. It did not say China had accepted the specific notification mechanism. There is no agreement to point to yet, and the US proposal still has no published rules, scope or timetable.
The hard part is deciding what must be reported
Governments already know how to notify one another about some military incidents. AI is messier. A reportable event could mean a model helping an attacker reach critical infrastructure, a powerful system behaving outside its test environment, a biological-risk capability appearing unexpectedly or a major failure affecting several countries.
Those cases are not equal. They also involve companies that may notice the problem before a government does. Any useful mechanism would need to answer basic questions:
- Which incidents cross the national-security threshold?
- How quickly must a country notify the other?
- Does the rule cover private labs, open-weight models and military systems?
- How much technical evidence must be shared?
- Can either side verify that an incident was contained?
A vague agreement could produce polite notices after the danger has passed. A rule that is too broad could become a demand for proprietary information or intelligence. Those details will decide whether the mechanism is useful.
Trust will be the limiting factor
Washington and Beijing are competing over advanced chips, models and data-centre capacity. US export controls are intended to slow China's access to leading computing equipment. Both governments also worry that AI could be used for cyberattacks, biological misuse or influence operations.
That gives them a shared reason to keep an accident from becoming a confrontation. It does not give either side a reason to reveal every weakness. A notification process may need narrow definitions, protected technical channels and a way to acknowledge an event without exposing model weights, intelligence sources or security controls.
The recent Gemini containment incident shows why definitions matter even outside geopolitics. An AI system can reach a real target during a test without causing physical harm. A useful international process would have to decide whether a near miss like that deserves notice, or only an incident with proven damage.
It would also need to separate deliberate state activity from an accident at a private company. Misreading one as the other is exactly the kind of escalation risk a hotline is supposed to reduce.
Canada should not be an afterthought
Canada relies heavily on American cloud infrastructure and commercial AI services. It also shares intelligence with the United States through Five Eyes and operates tightly connected energy, finance, transport and telecommunications systems.
If an AI incident begins in the United States but affects Canadian infrastructure, Ottawa needs to know whether an American notice to China would come before, after or instead of a notice to allies. The same question applies when a Canadian organization discovers an event involving a Chinese model or service.
Canada has been debating what AI developers should disclose domestically. Mapletechie's look at the federal AI transparency consultation found the same recurring problem: disclosure rules are only useful when the trigger, recipient and required evidence are clear.
Watch for a protocol, not another promise
The proposal is a worthwhile opening. It is not proof that the United States and China have agreed on what an AI emergency looks like, much less how to handle one.
The next meaningful development would be a joint document that defines covered incidents, notification deadlines, contact points and confidentiality rules. Until then, this is a diplomatic idea with good reasons behind it and most of the difficult work still ahead.
Sources: Associated Press; Reuters.
Tags: AI safety, China, United States, AI policy, Canada