Bessent outlines US-China AI warning channel as reporting rules remain open

There is also an unresolved question about the systems covered. The White House said the leaders agreed to use the term “super intelligence” for the applicable emerging technologies. China's September 26 readout instead describes an AI Dialogue and a channel for AI incidents.

That analysis distinguishes superintelligence—a hypothetical intellect far surpassing human capabilities across almost all domains—from AI's broader, existing applications. Bessent's use of the administration's preferred terminology is not evidence that either country has such systems. Nor do the two governments' different labels establish a shared technical boundary for incident reporting.

Selected 2026 milestones leading to Bessent's description of the planned warning channel. An operational launch date remains unspecified. Source: SIPRI.

Earlier safety talks coexist with technological competition

The groundwork predates this year's summits. A Chinese government account records the first intergovernmental AI dialogue in Geneva in May 2024, covering risks and governance. It also records Beijing's objections to U.S. restrictions on China in AI. The November 5 analysis notes that Xi and then-U.S. President Joe Biden affirmed the need for human control over nuclear-weapons decisions at their November 2024 meeting.

That history limits claims of novelty: neither bilateral AI risk discussions nor commitments to human control began with Bessent's latest remarks. The September agreement's specific addition is an incident channel, with a scheduled follow-on dialogue to work through its practical implications.

Competition remains explicit. In remarks after the September meetings, Trump said Washington would not slow its AI efforts and expressed reluctance to integrate with China, citing what he claimed was an American lead. That claimed lead is his assessment, not a benchmark established by the agreement. The summit followed the trade war he had escalated the previous year.

The industrial constraints also persist. The November 5 analysis identifies U.S. restrictions on Chinese access to semiconductors and computing power, alongside Chinese restrictions on critical minerals used in chip-making equipment, datacentre cooling and robotics. The White House's September readout acknowledged continuing U.S. concerns over shortages of rare earths and other critical minerals. An incident channel does not itself change access to those inputs.

The governments are pursuing narrower commercial arrangements in parallel. The White House said their Board of Trade, a government-to-government negotiating mechanism chartered in May, had agreed on recommendations for more favorable tariff treatment for $30 billion of non-sensitive goods in each direction. The figure describes goods covered by the recommendations; it is not a completed tariff reduction. Bessent describes the broader U.S. aim as reducing economic risks while maintaining trade ties.

The next test is how warnings reach the channel

The November 5 analysis identifies the decisions still needed: which incidents qualify, who reports them, which agencies handle them and what response follows. It notes that unusual model behavior is often detected first by the companies developing and operating the systems, while it remains unclear whether frontier AI laboratories would be required to report through the bilateral channel.

Existing domestic arrangements provide only part of the answer. The analysis describes U.S. reporting as dependent on the incident, affected infrastructure, agency rules and private frameworks, with no single mandatory national system covering all general anomalous AI incidents. It points to a tiered Chinese reporting system administered by the country's cyberspace regulator.

The scope of that comparison matters. China's national incident-reporting measures apply to cybersecurity incidents involving network operators in China. They assign national and provincial coordination responsibilities to cyberspace authorities. Their existence does not by itself establish that every unusual AI model behavior must be reported, or how information would pass into the bilateral channel.

Industry participation is therefore an implementation issue, not simply a matter of diplomatic attendance. The November 5 analysis records stronger trade-official and corporate representation at the 2026 summits, including Bessent's role, but argues that technical expertise must move into the working dialogue. Having executives at summit dinners does not establish their companies' reporting duties or what incident data they will share.

Bessent says he will accompany Trump to Shenzhen the following week for another meeting with Xi. Alongside the promised November AI exchange, that provides a near-term diplomatic opportunity. The evidence of a functioning safeguard would be more specific: agreed incident definitions, designated contacts, a route for developers to report relevant failures and procedures for acting on warnings.

England’s university regulator pushes AI-free tests and AI skills

The Office for Students wants more AI-free assessment and better AI skills teaching, backed by existing quality rules. Its research raises doubts about unsupervised coursework, but universities say assessment reform is already under way.