The United States has put a narrow but consequential idea on the table in its AI discussions with China: notify each other when an artificial-intelligence incident could affect national security. U.S. Treasury Secretary Scott Bessent said on September 20 that Washington had proposed the mechanism during weekend talks in New York with Chinese Vice Premier He Lifeng, according to AP.
That description matters because the proposal is not an operating channel. The fact pattern supplied for the talks does not establish a hotline, a formal protocol, a reporting timetable, or an agreed definition of an incident. It also does not show that China accepted, tested, staffed, or implemented the idea. The mechanism remains a U.S. proposal discussed between the two officials.
The September 20 conversation took place ahead of planned talks between President Donald Trump and President Xi Jinping. That timing placed AI alongside broader U.S.-China engagement, while leaving the incident-notification idea at an early diplomatic stage.
Bessent’s National-Security Proposal Sets a Narrow AI Test
The proposal, as described by Bessent and reported by AP, focuses on artificial-intelligence incidents that could affect national security. Reuters’ September 20 account characterized the proposed U.S.-China AI dialogue as addressing common goals, common threats, and incidents that reach a national-security level. That framing points to a potential exchange about exceptionally serious events rather than a general reporting system for every model failure or safety concern.
Even that narrow description leaves major questions open. The verified account does not identify a technical trigger, a reporting deadline, a responsible agency, or an escalation process. It does not say whether a notification would cover an incident originating in a model, a deployment, a cyber operation, or another setting. Those details cannot be treated as settled because neither the AP account nor the corroborating Reuters report establishes them.
The unresolved status also distinguishes the proposal from the earlier idea of a consultation forum. Reuters reported that Trump and Xi had discussed a forum for consultations on AI development in May, but that forum had not been formalized or launched. The September proposal therefore sits within a broader effort to create a channel for dialogue, while the available reporting does not show that either the forum or the incident mechanism became an institution.
China’s Response and the Dialogue’s Missing Institutional Shape
The Chinese response to the notification proposal was unclear in Reuters’ account. An Xinhua readout briefly acknowledged that the two sides discussed AI, but the supplied reporting does not establish Chinese endorsement of the U.S. mechanism. That distinction is central: a bilateral conversation about a proposal is not evidence of a joint commitment.
For policy officials, the lack of institutional shape is as important as the subject matter. There is no verified indication of who would receive a notification, how either side would validate an alert, or what would happen after one was sent. The available evidence also does not support describing the idea as active, operational, or formally agreed. Any future arrangement would need those elements to move from diplomatic concept to working practice, but the sources do not say that such work has begun.
The proposal follows the opening of a U.S.-China AI dialogue after New York talks and subsequent plans for a Shenzhen test of AI safety talks. Those earlier items provide historical context, but they do not change the status of the specific incident-notification mechanism described by Bessent and the September 20 reports.
Future AI-Safety Topics Remain Reuters-Reported Expectations
Reuters, citing analyst George Chen, reported that future discussions could address AI weaponization, AI-safety principles, critical-infrastructure protection, and cyberattack prevention. Those subjects should be understood as reported expectations, not a finalized bilateral agenda. The supplied material does not show that the governments agreed to discuss each topic, assigned negotiators, or set a schedule for addressing them.
The range of possible subjects helps explain why a national-security incident mechanism could be attractive to Washington. AI-related events may be difficult to separate from cyber risks, critical infrastructure, or military concerns, yet the proposal described in the available evidence does not specify how those boundaries would be drawn. A notification arrangement could only be assessed after the two governments define what qualifies and how much information each would share. None of that has been established in the available reporting.
Reuters, again citing Chen, reported limited prospects for cooperation, even as it described common goals and threats as part of the proposed dialogue. The proposal therefore remains dependent on a relationship in which the Chinese response was unclear and an earlier consultation forum was not formalized.
For the AI industry, the immediate significance is therefore diplomatic rather than operational. Companies, researchers, and infrastructure operators have no verified bilateral reporting channel to use under this proposal. For policymakers, the next meaningful development would be evidence that the two sides have defined the mechanism’s scope, participants, and procedures. Until then, the September 20 discussion is best understood as a proposed step in a still-unsettled U.S.-China AI dialogue, not as a functioning safeguard.