US-China AI Safety Talks Face Hurdle Over Definition of 'Serious' Incidents Ahead of APEC

US and Chinese officials have agreed in principle to a new "superintelligence" dialogue and an incident-alert channel, but remain far apart on what kind of AI failure warrants a cross-border warning. With President Donald Trump and President Xi Jinping set to meet November 18-19 at the APEC summit in Shenzhen, the gap illustrates how each side's domestic security priorities shape its approach to AI risk.

A dialogue mechanism, but no shared vocabulary

Treasury Secretary Scott Bessent told the National Committee on US-China Relations on November 10 that Washington and Beijing would advance a bilateral dialogue on "superintelligence" — the Trump administration's term for advanced AI — and operate a direct channel to alert each other to "serious incidents" involving such models. The two sides would also exchange best practices for testing advanced models before release and commit to keeping "human judgment firmly in command of the technology" [1][4].

The framework was first announced after Xi's September visit to Washington, where leaders agreed to set up an AI dialogue and an incident communication channel, though without specifying how either would operate [1]. Bessent said the Shenzhen meeting would advance the dialogue [1][4].

Divergent definitions of safety

The sticking point is what counts as a reportable incident. Analysts say the two governments view AI safety through different national-security lenses.

"The US is focused on loss of control of an AI system — a model or network of agents that starts engaging in unpredictable, potentially nefarious behaviour in ways that could cause harm and increase risks," said Karman Lucero, senior fellow at Yale Law School's Paul Tsai China Centre. "China's view of national security centres first on regime stability and content control" [1].

Xiao Qian, deputy director of Tsinghua University's Centre for International Security and Strategy, acknowledged convergence on human control as a starting point but said the two sides must first address common concerns such as misuse and abuse of AI before building crisis-communication channels [1]. She suggested initial consensus might be easiest on "incidents with cross-border implications that reach a significant level of risk" [1].

Lucero warned that bridging the gap over what defines an AI safety incident "could prove difficult" and that both sides would need "concerted efforts to clearly define what constituted a reportable incident" [1].

Vague thresholds, no binding rules

Bessent's public remarks did not define severity thresholds, notification timelines, or which agencies would staff the channel [4]. The explainx.ai analysis notes that the United States lacks a single mandatory registry for all civilian frontier-model failures, while China has Cyberspace Administration tiered reporting for certain AI events, and it remains unclear how Chinese labs would report through a US Treasury-led hotline [4].

The channel is explicitly non-binding — a confidence-building measure modeled on Cold War hotlines rather than an arms-control treaty [4]. Bessent paired cooperation language with a reminder that the two countries should "harbor no illusions" about their differences [4].

APEC backdrop: Nvidia scandal and trade friction

The Shenzhen meeting arrives amid a cascade of other issues. Nvidia CEO Jensen Huang resigned this month after reports of an affair with a suspected Chinese spy, though the company said no secrets were compromised [2][3]. Trump has expressed continued confidence in Huang and reportedly considered bringing him on the trip [5].

Analysts at CSIS said the White House is downplaying expectations for the Trump-Xi meeting, describing it as a "courtesy call" with more substantive economic talks expected at the G-20 in Miami three weeks later [5]. The AI dialogue is one of several tracks — alongside rare-earth exports, tariff truces, and chip-export controls — where both sides have signaled willingness to talk but have not locked in deliverables [3][5].

What happens next

Xiao said a realistic next step would be agreeing on "reportable incident types, notification thresholds and case examples" [1]. Bessent's remarks treat the channel as already agreed and awaiting implementation work [4], but if Shenzhen produces no mention of contact lists, severity rubrics, or lab reporting duties, confidence in near-term operations should be downgraded [4].

For now, the two capitals have built diplomatic plumbing for a conversation they have not yet agreed how to have.