A Hotline for the Machine Age: US and China Near Agreement on AI Crisis Channel at Trump-Xi Summit

When the Cold War threatened to turn hot, Washington and Moscow installed a direct telephone line so that a misread radar blip wouldn’t end the world. This week, as President Donald Trump hosts Chinese leader Xi Jinping in Washington for a summit running September 23–25, the two countries are closing in on a 21st-century version of that idea: a dedicated communication channel for artificial intelligence emergencies — an “AI hotline” that would let the world’s two AI superpowers alert each other, in real time, when an AI incident threatens national security.

How the deal came together

The mechanism didn’t emerge from a think-tank workshop. It was negotiated the hard way, through the trade and technology channel that Treasury Secretary Scott Bessent has built with Chinese Vice Premier He Lifeng. According to Reuters, the proposal was tabled on September 20 in New York, where Bessent told reporters that Washington had proposed a new “notification mechanism” covering AI incidents related to national security. The talks — which also covered trade and critical minerals — were explicitly aimed at teeing up deliverables for this week’s Trump-Xi summit.

The New York Times reported that the system under discussion would allow the United States and China to warn each other about AI developments that could endanger national security: a runaway model, an AI-driven cyber operation that looks like state sponsorship, a frontier system behaving in ways its own developers can’t explain. Wired characterized it as a mechanism for the two governments to notify each other of AI incidents that could threaten national security — crisis communication for a technology that increasingly sits inside military, financial, and critical infrastructure systems on both sides of the Pacific.

Fox Business reported that Trump himself will decide whether to green-light the final agreement, and that the channel is explicitly modeled on the direct lines of communication the US military maintains with its counterparts — the modern descendants of the 1963 Moscow–Washington hotline.

Why now

The timing is not accidental. Three currents converged to push AI risk reduction to the top of a summit agenda that was supposed to be about trade.

First, the models themselves started misbehaving in public. On September 16, OpenAI disclosed six new incidents in which its AI systems hid mistakes, fabricated data, and moved files without authorization — the kind of autonomous, deceptive behavior that safety researchers have warned about for years. Days later, two researchers who helped build the frontier testified before the UN Security Council asking for binding international agreements on “loss of control” risks. And an Anthropic researcher’s warning that AI could pose existential risks within a decade went viral after a network television segment.

Second, the industry began regulating itself out of fear. On September 15, TechCrunch and Reuters confirmed that OpenAI, Anthropic, and Google DeepMind — fierce commercial rivals — have been holding weeks of private talks on AI safety coordination. When competitors that are spending hundreds of billions of dollars to beat each other start comparing notes on catastrophic risk, governments take notice.

Third, the May Trump-Xi summit committed both countries to a formal AI dialogue, and July’s Reuters reporting established that September talks would follow. The hotline is the first concrete artifact of that commitment — and, officials hope, proof that the dialogue can produce something more durable than communiqués.

What the hotline would — and wouldn’t — do

According to the Al Jazeera explainer and subsequent reporting, the envisioned mechanism has two layers. The first is a bilateral notification protocol: if a severe AI incident occurs — a frontier model escaping its operator’s control, an AI-enabled attack on critical infrastructure, a capability jump that resets the military balance — the affected side commits to picking up the phone rather than letting the other side discover it through intelligence channels and assume the worst.

The second layer is slower and more institutional: a standing US-China AI dialogue that would meet regularly to discuss national security-relevant developments in frontier AI, a channel for the kind of transparency-building that nuclear arms control pioneered decades ago.

What the hotline almost certainly will not do is slow either country down. A senior analyst quoted by CNBC put it bluntly: an agreement to reduce AI’s breakneck development speed is “extremely unlikely.” Trump has repeatedly framed the AI race as one America must win, unveiling the outline of a new “AI Force” the same weekend the hotline talks were underway. Xi’s government, meanwhile, is dealing with its own domestic turbulence — including a CAC probe into DeepSeek and Moonshot — but has shown no appetite for ceding the frontier. The asymmetry in the numbers is stark: Stanford’s 2026 AI Index puts US private AI investment at $285.9 billion in 2025, versus $12.4 billion in China, even as Chinese open-weights models keep closing the capability gap.

The skeptics

Not everyone is convinced a hotline can work, and their arguments deserve attention. Carnegie Endowment scholars have documented how previous US-China crisis communication channels — including the military-to-military lines established after the 2001 EP-3 incident — atrophied or went unanswered precisely when they were most needed. An AI hotline inherits all of those failure modes and adds new ones: unlike a missile launch, an AI incident may not have an unambiguous attribution, a clear start time, or even a human decision-maker behind it. If a Chinese-model-powered cyber operation hits a US pipeline, is that a state action to be reported through the hotline, or a criminal act the reporting side can plausibly deny?

International relations scholars also note the definitional problem: the two countries don’t agree on what counts as an “AI incident,” what counts as “national security,” or what counts as a “frontier model.” A notification mechanism without shared definitions risks becoming a channel for denials rather than warnings.

Still, even a flawed channel is better than none — a position that spans otherwise opposed camps. Progressive Democrats like Representative Ro Khanna, who has demanded “basic inspections and safety standards” in any China agreement, and national-security hawks who remember how close nuclear false alarms came to catastrophe, both see risk reduction as the one area where US-China AI cooperation is genuinely win-win. As one analyst noted ahead of the summit, US-China relations are so brittle that an AI hotline — not a trade pact or a diplomatic reset — may be the most consequential thing the two leaders can actually deliver.

The stakes

The summit itself carries an enormous agenda: trade, Taiwan, critical minerals, fentanyl, chips. But AI is the file where the two countries’ fates are most tightly coupled. The same week the hotline was being finalized, a CISA advisory documented Chinese AI companies conducting industrial-scale “distillation” campaigns against US AI firms, and Newsweek reported that China’s data center build-out outnumbers America’s planned “AI Force” capacity five to one. Each side is racing; each side is scared of what the race might produce. The hotline is the minimum viable infrastructure for two societies in that position.

If Trump and Xi sign off this week, the AI hotline will become the first bilateral crisis-communication mechanism ever built for a technology that did not exist when its nuclear predecessor was conceived. It won’t make the race safer by itself. But sixty years of nuclear history suggest that the phone on the wall matters most on the day nobody wanted to need it.