Will the U.S. and China agree to an AI incident hotline by the end of 2026?
We put the chance that the U.S. and China formally agree to an AI incident hotline by December 31, 2026 at 36 percent, as of September 26. Washington proposed it before the Washington summit, but Beijing's readouts endorsed only continued AI dialogue, and U.S.-China crisis channels usually take far longer than three months to agree.
Why it matters: A channel for warning each other about dangerous AI incidents, such as an autonomous cyberattack or a model misused for weapons, could cut the risk that a software failure is mistaken for an act of war between the two largest powers.
TPG forecast
36%
Unlikely
Less likely than the market thinks
Version 1: 36% (range 24% to 50%) · Kalshi 51%
Evidence cutoff Sep 26, 2026 · published Sep 26, 2026 · Less likely than the market thinks
This forecast is on its first version. Each update adds a point; old versions are never edited.
The numbers behind the call
May 2024
last dedicated U.S.-China AI safety talks before September 2026, a gap of more than two years[3]
2008
year the U.S.-China military hotline began, the model for an AI channel[1]
48 hours
response window on the military hotline, where U.S. officials say calls sometimes go unanswered[2]
95 percent
of AI war-game simulations in which leading models engaged in nuclear signaling, per a 2026 study[4]
1997
year Clinton and Jiang agreed to create the first U.S.-China presidential link[6]
What it means for you
National security
A working alert channel gives Washington a way to ask Beijing whether an AI-driven cyber incident was state-ordered or a rogue system, before anyone retaliates. Without one, attribution in a crisis relies on guesswork.
Businesses and critical infrastructure
Hospitals, banks and utilities are the likeliest targets of AI-enabled attacks. Faster government-to-government warnings could shorten the time between an incident abroad and a defensive alert at home.
Tech investors and AI companies
U.S. negotiators say export controls on advanced chips are not part of the talks, so a hotline would not loosen sales to China. It would, however, signal a calmer bilateral climate, which markets have tended to reward.
Congress and AI policy
A bilateral channel could become the template for rules on frontier AI incidents that lawmakers in both parties have been pressing for, and it would test whether China honors crisis channels better than the military hotline.
The read
Three lenses, always in this order. How the method works.
01
Psychology
What do the decision makers need?
Beijing wants to be seen as a responsible AI power and to shape global rules through its own initiatives, but it resists bilateral arrangements that imply transparency about its systems or that look like U.S.-designed guardrails. Xi's line at the White House, continuing AI dialogue and jointly guarding against misuse, gives China credit for engagement without committing to an alert obligation. China's record with the military hotline, where U.S. officials say calls sometimes go unanswered, shows it values channels it controls more than channels that bind it.
Trump's incentive runs the other way from his rhetoric. He has said he wants to leave AI exactly where it is and that the Justice Department is the guardrail, yet a hotline restricts nothing and can be sold as a Cold War style red phone and a personal win with Xi. His Treasury secretary put the proposal on the table, so the White House owns it. The question is whether he spends leverage on it at APEC or the G20, where trade will crowd the agenda.
02
History
What happened the last times this came up?
Reference class: U.S.-China crisis communication channels. The presidential link agreed in 1997 and the defense telephone link running since 2008 each came after years of proposals. The closest analogue for a fast YES is November 2014, when China hosted APEC in Beijing and the two sides signed memoranda on notifying each other of major military activities and on rules for air and sea encounters. China as host wanted deliverables, and it will host APEC again in Shenzhen in November 2026.
Where the analogue breaks: the 2014 memoranda followed more than a year of working-level drafting inside the two militaries, while the AI proposal was first tabled on September 20, 2026 inside the trade channel, run by finance and trade officials rather than technical or security agencies. The last substantive bilateral AI milestone was the 2024 Lima statement keeping human control over nuclear launch decisions, which was a principle, not a mechanism, and dedicated AI safety talks then lapsed for more than two years.
03
Statistics
Base rate, adjustments, the number.
Base rate: of the five U.S.-China leader summits since November 2023 (Woodside, Lima, Busan, Beijing and Washington), two produced a new AI commitment of any kind and none produced an incident-notification mechanism. We estimate about a 40 percent chance that a given summit yields an AI commitment and about a 50 percent chance that such a commitment includes a qualifying alert mechanism, roughly 20 percent per summit. Two leader meetings remain before the deadline, APEC in Shenzhen in November and the G20 in Miami in December: 1 minus 0.8 times 0.8 gives 0.36.
Adjustments roughly cancel. Up: China confirmed it held its first AI talks with the United States and both sides agreed to meet again on AI, and China hosts APEC with an incentive to announce deliverables. Down: the September summit, the best venue so far, produced no AI agreement, and both Chinese readouts omitted the alert mechanism. We stay at 0.36, well below Kalshi's 0.51.
Between the lines
What people spent versus what they said, and what was left out.
What was left out
Xinhua's readout of the New York talks said only that the sides discussed AI, and Xi's Oval Office readout called for continued AI dialogue and guarding against misuse, without mentioning the U.S. notification proposal.
Beijing has accepted the dialogue frame but not the alert obligation. If it were close to yes, state media would already be claiming the idea as a shared initiative.
Personnel move
The AI talks sit inside the Bessent and He Lifeng economic channel, and Greer said chip export controls are not part of them.
Housing AI in the trade track makes a hotline a possible sweetener in a year-end trade package, which is the main path to YES. It also means the negotiators lack the technical mandate to design incident categories quickly, which is the main reason for delay.
Cheap talk
Trump posted that he wants to leave AI exactly where it is and that the Justice Department is the guardrail, and told the UN the United States will encourage AI, not rein it in.
The president is not spending capital on AI safety. A hotline survives only if it is framed as a crisis phone, not a safety regime, and only if Xi wants to hand him a win.
Sequencing
The U.S. tabled the proposal on September 20, four days before the summit, and the two sides agreed to meet again on AI afterward.
A proposal made that late was not meant to be signed at this summit. The realistic target is APEC in Shenzhen, where China controls the agenda and can present any channel as its own idea.
Our number after adjusting the base rate for what is different this time: 36%. It sits inside the historical interval.
Check our statistics yourself
- Reference class
- 0 of 5 (0%)
- Exact 95% CI (Clopper-Pearson)
- 0.0 to 52.2%
- Bayesian 95% credible (uniform prior)
- 0.4 to 45.9%
- Binomial test of our 36%
- p = 0.167
Our 36% sits inside the historical interval: the forecast is consistent with how cases like this have gone.
Reproduce in JASP (free, jasp-stats.org)
- Download base-rate.csv (one row per case in the reference class, outcome 1 or 0) and open it in JASP.
- Frequencies, then Binomial Test. Variable: outcome. Test value: 0.36. Tick Confidence interval. JASP reports the same p value and Clopper-Pearson interval shown here.
- Frequencies, then Bayesian Binomial Test, with a Beta(1, 1) prior. The posterior 95% credible interval matches ours.
Reference class: Computed by TPG from the Woodside (2023), Lima (2024), Busan (2025), Beijing (May 2026) and Washington (September 2026) summits. Tech Times and CNBC reporting document the Lima AI statement, the lapse in AI talks and the September 2026 outcome; the per-summit rate used in the statistics lens is our estimate.
Base rate
0%
of five U.S.-China leader summits since 2023 produced an AI incident-notification mechanism[3][5]
Computed by TPG from the Woodside (2023), Lima (2024), Busan (2025), Beijing (May 2026) and Washington (September 2026) summits. Tech Times and CNBC reporting document the Lima AI statement, the lapse in AI talks and the September 2026 outcome; the per-summit rate used in the statistics lens is our estimate. (n = 5)
Closest historical parallels
APEC Beijing military notification memoranda (2014)
Alike: China hosted APEC and agreed with the U.S. on a mechanism to notify each other of major military activities, a close cousin of an incident alert system.
Outcome: Signed during Obama's November 2014 visit to Beijing.
Where it breaks: Those texts were drafted by the two militaries over more than a year; the AI proposal is weeks old and run by trade officials.
Lima Biden and Xi statement on AI and nuclear weapons (2024)
Alike: A leader-level AI commitment announced at an APEC summit.
Outcome: Both affirmed that humans, not AI, should control decisions on nuclear weapons use.
Where it breaks: It was a statement of principle with no operational channel; a hotline requires both governments to build and staff a mechanism.
How it could play out
Every path we see, sized by probability. Paths that resolve YES add up to our 36%.
- Dialogue continues, no mechanism in 2026 No54%
The sides announce working groups or principles but defer an incident-notification commitment into 2027.
- Hotline agreed at APEC in Shenzhen Yes20%
China as host announces an AI risk communication channel with the U.S. as a summit deliverable.
- Hotline agreed at the G20 in Miami or in a year-end trade package Yes16%
The alert mechanism is bundled with a broader trade deal or truce extension in December.
- AI talks stall No10%
New U.S. chip controls, sanctions tied to Iran, or a Taiwan dispute freeze the AI track.
What would change our number
Chinese state media or the Foreign Ministry names the notification mechanism as a shared goal
+12 pointsA second U.S.-China AI working-level meeting is held with technical agencies, not only trade officials
+8 pointsTrump confirms travel to APEC in Shenzhen
+3 pointsWashington announces new AI chip export controls aimed at China before APEC
-10 pointsTPG 36% vs Kalshi 51%
Kalshi's contract, with near-identical rules, trades at 0.51 for agreement before January 1, 2027. We are 15 points lower for three specific reasons: the September summit, the most favorable venue so far, produced no AI deal; both Chinese readouts left the alert mechanism out; and every prior U.S.-China crisis channel took well over three months from proposal to agreement. Traders may be over-weighting the APEC host effect.
Market price recorded September 26, 2026 from Kalshi. Analysis, not investment advice.
How this question resolves
Resolves YES if, between September 26, 2026 and December 31, 2026, the governments of the United States and China formally announce an agreement to establish, or implement, a bilateral government-to-government mechanism for notifying each other about AI incidents that could affect national security (an AI hotline or alert system). The announcement must appear in an official readout, joint statement, agreement text, press release or formal briefing; one government's statement counts only if it says both governments agreed. An agreement to hold an AI dialogue, shared principles on AI safety, a unilateral proposal, or continuing negotiations do not count. Resolves NO otherwise.
Closes: December 31, 2026. Judged by: White House, Treasury or State Department statements, China's Foreign or Commerce Ministry and Xinhua readouts, and Reuters or AP reporting
Questions people ask
What is the US China AI hotline?
+
It is a U.S. proposal, made by Treasury Secretary Scott Bessent in talks with Chinese Vice Premier He Lifeng on September 20, 2026, for a notification mechanism that would let each government warn the other about AI incidents serious enough to threaten national security. As of September 26, 2026, China had not publicly accepted it.
Did Trump and Xi agree on AI?
+
Not on anything concrete. At their September 2026 Washington summit, Xi said the two sides can continue AI dialogue and guard against misuse, and Trump said he wants to leave AI exactly where it is. No AI agreement or hotline was announced, though the two sides agreed to keep talking about AI.
Will the US and China cooperate on AI safety?
+
Some cooperation is likely, but a formal incident hotline in 2026 is less likely than not. TPG puts the chance of an agreed U.S.-China AI incident notification mechanism by December 31, 2026 at 36 percent, with the best opportunities at APEC in Shenzhen in November and the G20 in Miami in December.
Does the US have a hotline with China?
+
Yes. The two countries agreed in 1997 to create a direct presidential link, and they have had a military hotline since 2008 for senior defense officials. It gives each side 48 hours to respond, but U.S. officials have complained that sometimes no one answers. There is no AI-specific channel yet.
Sources
Every source is dated on or before the September 26, 2026 evidence cutoff. Nothing published later was used.
- [1]What's the US-China AI hotline that Trump plans to pitch to Xi Jinping?. Al Jazeera, Sep 21, 2026.
“The two countries have had a military hotline since 2008, designed to allow senior defence officials to communicate during potentially dangerous incidents.”
- [2]What's the US-China AI hotline that Trump plans to pitch to Xi Jinping?. Al Jazeera, Sep 21, 2026.
“It gives both sides 48 hours to respond, though US officials have complained that sometimes no one picks up.”
- [3]US-China AI Hotline Needs Text, Tiers, and Technicians: Trump-Xi Summit Opens Today. Tech Times, Sep 24, 2026.
“the two countries met in Geneva in May 2024 for dedicated AI safety talks”
- [4]US-China AI Hotline Needs Text, Tiers, and Technicians: Trump-Xi Summit Opens Today. Tech Times, Sep 24, 2026.
“occurred in 95 percent of games, with GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash all treating nuclear threats as routine strategic tools”
- [5]China's Xi urges U.S. to cooperate on AI. CNBC, Sep 24, 2026.
“The two sides can continue AI dialogue, exchange views on risks and benefits, and together guard against the misuse or malicious use of AI”
- [6]What's the US-China AI hotline that Trump plans to pitch to Xi Jinping?. Al Jazeera, Sep 21, 2026.
“In 1997, former US President Bill Clinton and his then-Chinese counterpart Jiang Zemin agreed to create a”
- [7]U.S. proposes exchanging AI safety alerts with China, Bessent says. NBC News, Sep 20, 2026.
“did not mention the U.S. proposal and said only that the two sides had discussed AI.”
- [8]China confirms first AI talks with U.S. have taken place, hints at trade truce extension. CNBC, Sep 24, 2026.
“China's Commerce Ministry confirmed Thursday that its senior trade negotiators had held their first talks with the U.S. on artificial intelligence.”
Version history and integrity
Published forecasts are never edited. When the evidence changes we publish a new version beside the old one, each locked with a sha256 hash of its full contents. When this question resolves it is scored in public on our track record.
| Version | Cutoff | Forecast | Lock hash |
|---|---|---|---|
| v1 | Sep 26, 2026 | 36% Unlikely | 042745047fdd6ed9f4f4f935e4fedbe50e2760f6061cecad25ab20247d955502 |
Download every forecast: JSON or CSV (CC BY 4.0). Public sources only. Analysis, not investment or legal advice.