Beijing and Washington agreed during President Xi Jinping's September 23-25 state visit to establish a formal China-U.S. AI Dialogue and a bilateral channel for AI-related incidents, according to China's Ministry of Foreign Affairs as reported by MediaNama. The first exchange is scheduled for November 2026. The Chinese readout does not say what counts as an "AI incident," which agencies participate, or whether the channel covers cybersecurity, autonomous systems or military uses.
That gap is the whole story. The channel is a good idea. What it is for will decide whether it helps or harms.
The strongest case for the channel
The case for a hotline is serious. Two states with the most capable AI systems, and an adversarial relationship, currently have no agreed way to tell each other that something has gone wrong. Crisis communication links exist because misreading an event as an attack is more dangerous than the event itself. If a model-driven cyber intrusion, a malfunctioning automated system, or a fast-spreading synthetic-media event hits both countries, a phone number beats a press statement. Cheap, early contact between officials is the least controversial form of risk management there is. Al Jazeera's correspondent at the White House described the summit as offering "a lot more pomp than progress," but the same report cited an analyst calling continued bilateral dialogue perhaps the biggest deliverable for both sides. That is a fair reading: the value here is in keeping talks alive.
Precedent exists, and it is narrow
This is not the first US-China AI understanding. At their November 2024 meeting in Lima, the two presidents "stressed the need to maintain human control over the decision to use nuclear weapons," according to China's foreign ministry readout of that meeting. The same readout notes they had co-sponsored each other's AI resolutions at the UN General Assembly. Those commitments were specific, short and about one catastrophic scenario. The new channel is the opposite: open-ended in scope and undefined in trigger.
Specificity matters because a hotline only works if both sides agree in advance what warrants a call. Without that, each government will apply its own definition. For Washington, an AI incident might mean a model used in a cyberattack. For Beijing, whose officials this month criticised calls for tighter chip controls as "Cold War" tactics, it might include exports restrictions or the spread of content it deems harmful. MediaNama notes that Beijing's response to Anthropic CEO Dario Amodei's proposal for tighter controls on advanced chips reaching China called for "an open, inclusive and benevolent approach to AI." The same phrase can describe very different policy goals.
The risk to speech and innovation
Our concern is not the existence of a channel but its potential drift. If "AI incident" is defined broadly enough to include harmful information, the channel becomes a route for governments to press each other, and each other's companies, on content. China's regulatory model already treats generated content as a matter of state responsibility. A dialogue that quietly imports that logic into a US-China setting would put pressure on open-model developers and researchers who publish safety findings. Incident reporting works in aviation and cybersecurity because the taxonomy is technical: what failed, how, and with what consequence. It fails when "incident" means "output someone disliked."
A narrow definition helps innovation too. Companies building or deploying models across borders need predictable rules. A vague government-to-government mechanism creates uncertainty about whether a model failure could become a diplomatic matter and a pretext for restrictions.
Do not duplicate the multilateral track
The bilateral channel also arrives alongside a slower multilateral process. The UN General Assembly adopted resolution A/RES/79/325 by consensus on 26 August 2025, creating the Independent International Scientific Panel on AI and a Global Dialogue on AI Governance. TechPolicy.Press reports that the panel has 40 experts serving in a personal capacity, must disclose conflicts of interest, and that critics flagged funding shortfalls and limited civil-society participation. The UN track brings in more than 100 countries and produces evidence anyone can inspect. The bilateral track is faster and more decisive but opaque.
The sensible division of labour is that the US-China channel handles operational contact between two states, while the UN panel supplies shared definitions and evidence. A bilateral channel that invents its own definitions in private would undercut the only forum where the wider world, including India, Brazil and African states, has a seat.
What to watch in November
Three tests will show whether this is substance or ceremony.
- A published definition. Even a one-page shared taxonomy limited to technical failures, such as loss of control of an autonomous system, a model-enabled attack on critical infrastructure, or a safety failure with cross-border effects, would make the channel credible.
- Named agencies. A channel run by technical agencies and crisis officials reads differently from one run by propaganda or content regulators.
- A speech carve-out. An explicit statement that the mechanism does not cover lawful expression, research publication or open-weight releases would remove the biggest downside risk.
None of this requires slowing AI development. It requires only that the two governments say plainly what they are agreeing to. Proportionate governance starts with a clear definition of the harm being addressed. Until Washington and Beijing publish one, the channel deserves cautious support and close scrutiny, not applause.