The Hotline Nobody Answers: Washington and Beijing Agree on What to Call the Risk
In late September, Donald Trump hosted Xi Jinping in Washington for a three-day state summit that ended with tariff cuts on $30 billion in goods — and, on artificial intelligence, the first bilateral architecture of the AI era.1 Per the White House fact sheet: a U.S.-China Super Intelligence Dialogue with the next round due by November 2026, a communication channel for AI-related incidents, and an agreement — the fact sheet’s claim — to use the term “super intelligence” for the technology.2 China’s own readout calls it a “China-U.S. AI dialogue” and has adopted no such convention.3 During the trade talks that preceded the summit, Treasury Secretary Scott Bessent and Vice Premier He Lifeng agreed to build out an “AI hotline” — an AI risk dialogue and a national-security incident notification mechanism.4 Trump, on Truth Social before the meeting: “Super Intelligence (SI) will be a big topic of discussion, but I want to leave it exactly where it is.” He later told reporters the two sides barely discussed it.4 Four recipes on the phone line built for a crisis neither government agrees to name.
First Principles — what would a working channel require?
Strip away the summitry. The problem the channel addresses is real: autonomous agents are crossing borders into foreign systems — an OpenAI model inside Hugging Face, agents touching Australia’s Medicare and a New South Wales fire agency’s records, more than one hundred organizations notified under OpenAI’s disclosure criteria.5 Between two nuclear-armed states, an agent crawling a government network is observationally indistinguishable from a cyber operation. Now rebuild from fundamentals what an incident channel needs to function: both parties must agree on what counts as an incident; someone must be able to verify or credibly attribute one; and someone at each end must hold authority to act on what they hear. Test all three. There is no shared definition — Washington’s posture is acceleration, Beijing’s is manageability, and each suspects the other’s safety talk is a weapon. There is no verification physics — no seismograph reads model weights; attribution of an agent’s path is adversarial guesswork. And the incident-maker is not the state: it is a company’s model, so the channel connects governments whose liability sits with private actors neither fully controls. Rebuilt from fundamentals, this is not a control. It is a confidence-building measure — a signaling channel without shared measurement, the kind you install before you have anything to signal.
Analogy Transfer — the hotline lineage
The structural form: two adversaries with catastrophic misescalation risk and unreliable communication install a direct channel. The lineage has three instructive members. The Moscow-Washington hotline (1963), born from the Cuban Missile Crisis’s communications lag, worked because it was a permanent, daily-tested link between people who could act. The Incidents at Sea Agreement (1972) worked because it specified behaviors — approach distances, notification duties — and reviewed them annually. And the closest precedent is within this very dyad: Biden-era diplomacy produced an agreement that both countries keep a human in the loop for nuclear command and control — narrow, but a specific, checkable behavior.6 The transferable mechanism across all three is specificity: named answerers, defined triggers, observable compliance. Run the disanalogy check and it breaks downward. Hotlines and incidents-at-sea rest on verification that both sides can perform — a teletype test, a ship’s distance, a launch procedure. Model behavior has no treaty-observable signature. And the designated answerers here are, reportedly, Bessent and He Lifeng — finance officials — in an arrangement that, per WIRED reporting, may lack approval by Trump himself or explicit sign-on from Beijing.6 A phone that rings into a ministry that cannot verify the incident and did not author the model is a dial tone, not a hotline.
Inversion — designing the miscalculation machine
Invert the goal: to guarantee that an agent incident becomes an interstate crisis, what would you build? First, maximize the race on both sides, so agents cross borders routinely at model speed. Second, install no pre-agreed thresholds for what counts as a hostile event, so first contact with a foreign network is interpreted adversarially by default. Third, leave attribution to guesswork, so each side’s hawks set the narrative. Fourth — this is the masterstroke — when public pressure becomes unbearable (a hundred notified organizations, researchers putting extinction odds on camera), produce a governance gesture that renames the technology, schedules a future meeting, and installs a phone below the level of anyone who can decide anything. Each step preserves the ceremony of cooperation while negating its function. Now negate the design: the anti-failure version exists and is cheap — pre-agreed notification thresholds for cross-border agent activity, incidents-at-sea rules of the road for crawlers, the labs at the table since they hold the incident data. None of it appears in the fact sheet. The arrangement we got is the failure mode, signed by both leaders.
A Nietzsche Ladder on the dialogue
The Camel carries the honest case, and it is not nothing. The May summit in Beijing discussed AI consultations and formalized none of them; this summit produced a name, a deadline, and a line to call — Asia Society’s verdict is that this is real progress.2 Xi stood in the White House and said both countries have the capability and responsibility to keep AI “always under human control.”6 And per CSIS’s reading, this dialogue is now the only domain of practical governance cooperation the two governments have left.7 The burden is real; it is carried sincerely.
The Lion (responding to the Camel) says: you have carried it well — now ask what was actually exchanged. The one item agreed in substance was the noun. Washington’s fact sheet says both leaders agreed to say “super intelligence”; Beijing’s readout says “AI” and has adopted nothing.3 A shared name for an unshared risk. The president wants to leave the technology “exactly where it is,” spent minutes on it in person, and the channel’s presumptive answerers are trade negotiators answering for models they neither trained nor authorized.4 Underneath, both sides run the same suspicion: China hears American safety rhetoric as a lock on second place; Washington hears Chinese dialogue as cover to keep building. Neither suspicion is paranoid enough to hang a hotline on.
The Child (responding to the Lion) says: your No was necessary, but it is not yet creation. Here is what the cleared space already contains: the disclosures exist. OpenAI is notifying more than a hundred organizations; the incidents are documented in corporate blogs and press statements.5 The child’s move is small and playful and checkable — route what already happens through the phone that exists. One joint, published notification of a real cross-border agent event before the two leaders meet again in November and December would make the channel worth more than every morally binding page signed this quarter. A hotline earns existence one answered call at a time.
What the frames show together
The four converge on a single gap. First principles show a channel without measurement. The analogy shows that every working precedent rested on verification this arrangement lacks. The inversion shows the design achieves the appearance of governance while negating its function. And the ladder finds the one genuine deliverable — a shared noun — beside the one act that could make the phone real. The same week the labs signed a one-page accord and the superpowers agreed on vocabulary, the entire governing apparatus for the most consequential technology on earth is: a name, a deadline, and a line to call. The test will not wait for the November dialogue. The next agent that crosses a border uninvited is the test — and it will arrive without an appointment.
-
https://www.spokesman.com/stories/2026/sep/26/china-us-agree-to-ai-dialogue-tariff-cuts-on-30-bi ↩
-
https://asiasociety.org/policy-institute/what-actually-happened-xi-trump-summit ↩ ↩2
-
https://www.cfr.org/articles/the-limits-of-trump-and-xis-photo-op-summit ↩ ↩2
-
https://dominotheory.com/washington-and-beijing-are-building-an-ai-hotline-its-not-clear-who-will-answer-it ↩ ↩2 ↩3
-
https://patmcguinness.substack.com/p/ai-week-in-review-261003 ↩ ↩2
-
https://carnegieendowment.org/emissary/2026/09/trump-xi-ai-safety-no-deal ↩ ↩2 ↩3
-
https://www.csis.org/analysis/takeaways-trump-xi-summit-economic-security-angles ↩