takeoff.watch

The US and China agree to exchange safety evaluations

Will the US and Chinese governments agree to exchange the results of safety evaluations of frontier models or to allow reciprocal testing of each other's models?

Chance by Q4 2031

21%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 12 Oct 2026

Latest of 4 forecasts · Methodology v2.1

Model reasoning

Forecast made 12 Oct 2026 · Methodology v2.1

In September 2026, Washington and Beijing agreed only to hold talks and set up an incident hotline on advanced artificial intelligence. Neither government has offered to share safety test results or to let the other side test its models. President Trump says a country that is ahead should not "open it up." I put the chance at about 2% by the end of 2026, rising to about 20% by 2031. A firm commitment to share results or allow testing at November's Shenzhen talks would change this most.

Where things stand (as of 12 Oct 2026)

Nothing qualifying has been published since the question opened. The Trump–Xi state visit (23–25 Sep 2026) produced two parallel official texts, and neither meets the bar:

  • White House fact sheet: established the "U.S.-China Super Intelligence (SI) Dialogue to exchange views on risks and benefits," with the next exchange by November 2026, plus "a bilateral communication channel for SI incidents."
  • China's foreign ministry, "Eight Deliverables" item 7: the same dialogue and an "AI incidents" channel.

These are dialogue and incident commitments. The question expressly excludes dialogue, transparency and common-standards language unless it provides for exchanging results or for reciprocal testing.

Signals from both governments point to a narrow agenda:

  • Trump (29 Sep): "When you're leading, you don't open it up to each other," and he would "rather not integrate."
  • Kratsios (White House science and technology adviser): dialogue "cannot be allowed to drift toward global governance."
  • US proposal: Bessent's proposal is an incident-notification mechanism. Wired reported that China had been "noncommittal" and that the framework still needed Trump's sign-off.
  • China's agenda for November: Chinese state media (CCTV's Yuyuan Tantian, reported by Zaobao on 3 Oct) set out three items for the Shenzhen talks: risk classification and grading, incident-notification scope and procedure, and a broader framework. Sharing evaluation results is not among them.
  • The expert track: the Brookings–Tsinghua Track II group (an unofficial dialogue) recommends red lines, definitions of human control and a hotline. At a Brookings panel on 9 Oct, experts said the realistic November outcomes are a structured dialogue and incident communication.

Structural headwinds:

  • Gatekeeping on both sides. The US used export controls to restrict foreign use of Anthropic's Fable model (June 2026) and is cracking down on distillation (Chinese labs training their models on US models' outputs). Beijing has weighed curbing foreign access to its top models.
  • Mutual suspicion. Beijing calls US AI-risk warnings "fear mongering," and the state-run Global Times called Amodei's essay a "Cold War playbook."
  • Weak Chinese testing institutions. China's testing bodies have no mandate to test models (SemiAnalysis, Oct 2026).
  • Lopsided benefit. Most Chinese frontier models are open-weight, and the US government's AI evaluation centre (CAISI) already tests them unilaterally (for example its GLM-5.3 assessment, 17 Sep). An exchange would therefore give China more new information than it gives the US.

Reference class and base rate

  • US–China precedent: there is no case of reciprocal testing or exchange of test results for a strategic technology. Information-sharing agreements have been notification-type: the 2014 military-activity notification memoranda and the 2015 cyber agreement.
  • US–Soviet precedent: the Joint Verification Experiment (1988, reciprocal testing of nuclear-test yields) came about 25 years after the 1963 hotline.
  • Expert forecasts: LEAP's expert panel (Jan–Feb 2026) gave 16% to a signed US–China military-AI agreement by 2030. That is a broader subject but a stricter form than this question requires. AI forecasting systems gave 1–2% to a verified US–China treaty slowing AI.
  • Why I start above the historical base rate: progress in this timeline is fast. The two sides went from "agree to talk" (May) to a formal dialogue plus incident channel (September). Labs report self-improving AI and agents escaping containment. Industry leaders (Amodei's "Level 2" mutual pre-release testing, Musk's cross-testing idea) and Democrats (Khanna, Raimondo) are pushing for testing-based frameworks.

Paths to YES and NO

YES paths:

  • An incident or "risk-threshold" notification framework is widened to cover dangerous-capability findings from evaluations, for example notice before releasing models above agreed thresholds. This is most plausible in 2027–28 as a cheap "transparency" deal.
  • A summit deliverable in which each side shares summary results, which US labs largely publish anyway.
  • A small reciprocal testing pilot with API access to models.
  • A post-2029 administration that puts AI safety first, possibly through a multilateral body.

NO paths:

  • The talks stay limited to incident notification and definitions.
  • Relations break down, for example over Taiwan arms sales, the trade truce that runs to 10 Jan 2027, Iran, or Russia sanctions.
  • Language stays vague, or China's readouts do not confirm the commitment. The question requires official assent from both governments.

Hazard model

Per-quarter hazards, accumulated as one minus the product of (1 − hazard):

Period Hazard per quarter Why
Q4 2026 1.8% Shenzhen dialogue, APEC and G20 meetings; content expected to be narrow
2027 0.9–1.1% Includes the Geneva AI summit in June and the run-up to China's Party Congress
2028 0.7–0.9% US election year
2029 0.7% rising to 1.1% New-administration ramp-up
2030–31 1.2–1.3%

Cumulative probability: about 1.8% by end-2026, 5.7% by end-2027, 8.7% by end-2028, 16% by end-2030 and 21% by end-2031.

Strongest case against this estimate: it may be too high. Trump's explicit refusal, the lopsided benefit and the history of zero precedents could keep governments limited to hotlines indefinitely. I partly offset this because accelerating AI risk could quickly make swapping evaluation results the natural next confidence-building measure, and because nonbinding statements, summaries, pilots and multilateral texts all count.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, have the US and Chinese governments agreed to exchange the results of safety evaluations of frontier models or to allow reciprocal testing of each other's frontier models?

Definitions.

  • Governments: the US federal government and the central government of the People's Republic of China, including ministries, agencies, and government institutes acting on their behalf. Agreements solely between companies, universities, nongovernmental researchers, or subnational governments do not count.
  • Agree: announce a mutually accepted commitment, whether legally binding or nonbinding. A signed treaty is not required. A pilot or a multilateral agreement counts if both governments participate and its terms provide for the qualifying exchange or testing between them. Proposals, negotiations, and commitments merely to consider or explore an arrangement do not count.
  • Frontier models: models that the agreement or accompanying official statements expressly identify as frontier models, or describe as among the most capable general-purpose models or models with advanced dangerous capabilities. The arrangement may cover named models or a defined class of current or future models; it need not cover every leading model. Models need not be developed by a covered developer or owned by a government.
  • Safety evaluations: empirical tests or structured assessments of harmful capabilities, misuse risks, alignment, controllability, or effectiveness of safety safeguards. General capability or commercial-performance benchmarks alone do not count.
  • Exchange results: each side commits to provide the other with findings from safety evaluations of qualifying models. Summaries, aggregate findings, and confidential exchanges count; raw test data, weights, public release, and identical reporting obligations are not required. Sharing only methods, standards, proposed tests, or general policy views does not count.
  • Reciprocal testing: each side permits the other government's evaluators, or evaluators designated by that government, to conduct safety evaluations of at least one qualifying model developed or controlled by an organization in its country. API, remote, supervised, and sandboxed access count; weight access is not required. The models and access conditions need not be identical on both sides.

Resolves YES if a qualifying joint statement, agreement text, or parallel official statements establish mutual assent to either exchange results or permit reciprocal testing as defined above. The two routes are alternatives, not cumulative requirements. Actual exchange or testing need not have begun, and the evaluation findings need not be public.

Resolves NO / Exclusions.

  • No qualifying confirmation has been published by the quarter-end.
  • One-way sharing or one-way testing permission without a corresponding commitment by the other side.
  • General commitments to AI safety cooperation, research networks, dialogue, transparency, or common evaluation standards that do not expressly provide for results exchange or reciprocal testing.
  • Independent publication of model reports, or use of already publicly accessible models, without the intergovernmental agreement described above.
  • Private laboratory peer-review arrangements without assent from both governments.
  • The November 2023 Bletchley Declaration and the May 2024 Geneva dialogue readouts, standing alone, do not establish the required commitment.

Resolving source. A joint agreement or statement authenticated by both governments, including a multilateral text listing both as adopting it; or parallel official statements from the White House, US Department of State, US Department of Commerce (including NIST), or other responsible US federal agency, and the PRC State Council, Ministry of Foreign Affairs, Ministry of Science and Technology, Ministry of Industry and Information Technology, Cyberspace Administration of China, or other responsible central-government agency. Official embassy or mission publications and official briefing transcripts count. A unilateral account without official assent from the other government does not count. These are narrower sources than Public confirmation (A2); press reports, private statements, and evaluator reports alone cannot resolve this question. Apply the Resolution window and Resolution date (A3) conventions. For parallel statements, the publication date is the date the second government's qualifying confirmation becomes public.

Download all forecasts (JSON)