Trump-Xi AI Safety Alert Plan Faces a US-China Trust Crisis
Trump and Xi weigh an AI risk alert channel, but export controls, missing technical oversight and rivalry threaten meaningful US-China safety cooperation.
Summary
On September 23, 2026, ahead of Donald Trump and Xi Jinping’s Thursday and Friday meetings, Treasury Secretary Scott Bessent said Washington and Beijing had discussed using an existing channel to flag dangerous or unpredictable AI systems and share common goals. China has not publicly accepted the proposal, which lacks technical officials able to assess alerts, although a follow-up meeting is scheduled for November. Trust remains weak as Trump expands export controls, Congress considers further chip and hardware restrictions, and the Justice Department accuses six leading Chinese AI makers of stealing from US frontier laboratories, allegations China disputes.
Georgetown Center for Security and Emerging Technology fellow Sam Bresnick called the channel better than nothing but warned US safety policy can look like containment, particularly when Trump and Sam Altman promote American-led standards. Omdia chief analyst Su Lian Jye said effectiveness requires shared risk classifications, notification triggers, response protocols and reviews. Representative Ro Khanna urged a ban on recursively self-improving AI and safeguards against biological and nuclear misuse. Stanford Digital Economy Lab fellow Alvin Wang Graylin said Chinese open models could soon exceed 80 percent market share, frontier capability leads typically last only months, and hundreds of Chinese model audits could complement US frontier-model expertise. He favors bilateral and multilateral safety channels, narrower Joe Biden-style export controls, and reasonable regulation to counter threats to workers, economies and national security, including smaller models used for biological, chemical or cyber weapons.
Positives
- A November follow-up meeting indicates Washington and Beijing see enough value in the AI discussions to continue them.
- The proposed channel would let either government flag dangerous or unpredictable AI behavior directly to the other.
- China’s experience from hundreds of model audits could complement US expertise in testing frontier systems.
- Su Lian Jye identified concrete requirements, including shared risk classifications, notification triggers, response protocols and regular reviews.
- Ro Khanna proposed specific safeguards against recursively self-improving AI and biological or nuclear misuse.
Risks & concerns
- China has not publicly accepted the alert mechanism, leaving its participation and durability uncertain.
- No technical officials are assigned to evaluate alerts, weakening the channel’s ability to distinguish urgent threats.
- Expanded export controls, proposed chip restrictions and disputed theft allegations deepen Chinese suspicion of US safety initiatives.
- Chinese open models could soon exceed 80 percent market share while the US remains focused on short-lived frontier capability leads.
- Smaller specialized models could enable rogue actors to launch biological, chemical or cyber attacks despite superintelligence-focused policies.