takeoff.watch

The US and China set up a joint frontier AI project

Will the US and Chinese governments establish, by signed agreement, a frontier AI project in which both take part in development or governance?

Chance by Q4 2031

2%

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

I put the chance of a signed US–China agreement setting up a joint frontier AI project at about 0.5% before 2029 and about 2% by the end of 2031. The September 2026 summit produced only a dialogue and an incident hotline, and Trump explicitly refused to "integrate" with China. Even the boldest US proposals seek limits and inspections, not building models together. A severe AI disaster leading both governments to merge development efforts would most change this.

Where things stand (as of 12 Oct 2026)

Nothing published since the question opened on 16 Sep 2026 comes close to qualifying. AI was a central topic at Xi's state visit to Washington (23–25 Sep 2026), but only two AI outcomes came out of it:

  • a "U.S.-China Super Intelligence (SI) Dialogue" to "exchange views on risks and benefits", with the next round by November 2026;
  • a bilateral communication channel for SI incidents.

Both are confirmed by the White House fact sheet and China's eight-point consensus. They are a dialogue and a hotline, and the question explicitly excludes both. The day after the summit, Trump publicly ruled out joint development: "I would rather not integrate, because we're leading by a lot… When you're leading, you don't open it up to each other" (Al Jazeera and Times of India, 26 Sep 2026).

The administration's follow-up points the same way:

  • the White House Accord on Super Intelligence (29 Sep), a voluntary pledge by companies;
  • the "Super Intelligence Force" under Director of National Intelligence Jay Clayton (October), whose stated aim is to "ensure that America continues to lead the World."

Chinese signals about the November round, reported by Pekingnology from a state-media outlet that cited Chinese negotiators, concern grading risks, rules for reporting incidents, and China's voice in writing rules. Nobody is talking about co-development. China's ambassador urges openness and "drawing on each other's strengths" (Newsweek, October 2026). That suggests Beijing might accept a joint project if one were offered, but Washington is the bottleneck.

Reference class and base rate

The reference class is rival great powers jointly developing, or jointly governing, the strategically decisive technology of their era under a signed agreement. That has essentially never happened:

  • The Baruch Plan for international control of atomic energy failed.
  • US–Soviet cooperation in détente stuck to non-decisive areas. Apollo–Soyuz was a joint space mission, and ITER, a fusion project that grew out of the 1985 US–Soviet summit, is the only frontier-science body with both the US and China as governing members.

Even the most cooperation-minded proposals in this debate are about limits and verification, not building models together:

  • Dario Amodei's "levels" of cooperation run from information sharing to a pause.
  • Rep. Ro Khanna's proposed five-pillar treaty adds kill switches and an inspection agency.
  • The Center for American Progress asks for a hotline and working groups.
  • A 20-country declaration calls for an oversight institution.

A joint frontier project (the "CERN for AI", MAGIC or AI 2027 "Consensus-1" idea) appears mainly in advocacy and fictional scenarios. Even the optimistic "Plan A" scenario puts a joint regime only in the late 2020s or early 2030s. On Metaculus, "US-China AGI-limiting treaty signed before 2030" stood around 4–5% (Oct 2025 and Apr 2026). This question needs something narrower than such a treaty: a specific project, an explicitly stated frontier-capability goal, and substantive development or governance roles for both governments. Funding, hosting, observer status and evaluation-only arrangements are excluded.

Paths and estimate

  • Trump's term (to January 2029): about 0.5% in total. The administration is in race mode, Congress is hawkish, and US export controls and security worries about sharing technology work against it. The residual path is a serious AI incident plus Chinese parity plus Trump's appetite for deals. Even then, cooperation would more likely be incident response or limits than joint development.
  • 2029–2031: about 1.7% in total. A new president might campaign on AI safety, which is now a bipartisan public worry (Sanders and Bannon have shared a stage on it). My rough decomposition: a substantive US–China AI treaty by the end of 2031 at about 10%, times about 15–20% that it creates a jointly governed frontier lab with an explicit frontier goal, plus a small chance of a multilateral or "goodwill" project.
  • Hazard shape. The quarterly hazard rises from about 0.02% (Q4 2026) to about 0.18% (2031), reflecting faster AI capabilities and the change of administration. I built the 21 quarterly hazards in Python and took cumulative = 1 − ∏(1 − h). That gives about 0.46% by the end of 2028 and about 2.1% by the end of 2031.

Strongest case for a higher number. By the labs' own accounts (Amodei, OpenAI's chief scientist), AI that improves itself is starting, and rogue-agent incidents such as OpenAI agents hacking Hugging Face are pushing AI risk into mainstream politics. In fast-takeoff worlds, unusual arrangements become more plausible, such as a co-designed AI used to enforce a treaty. I gave this some weight, which is why the late-period hazard rises. I did not go higher because the specific signed, officially confirmed, explicitly frontier joint-project outcome is a small slice even of those worlds.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, have the US and Chinese governments established, by signed agreement, a frontier AI project in which both take part in development or governance?

Definitions.

  • US and Chinese governments: the United States federal government and the central government of the People's Republic of China, including ministries, agencies, and representatives signing on their behalf. Participation solely by subnational governments, private companies, universities, or state-owned enterprises does not count unless the agreement expressly gives them a role on behalf of their national government.
  • Signed agreement: a written instrument signed by authorized representatives of both governments, whether bilateral or multilateral. A treaty, executive agreement, memorandum of understanding, or signed joint declaration counts; legal enforceability and legislative ratification are not required. An unsigned announcement or an agreement to negotiate does not count.
  • Establish: the signed instrument creates a specified joint project or formally brings both governments into an existing project. It must assign substantive roles, not merely express an intention to explore cooperation. Construction, training, or model release need not have begun.
  • Frontier AI project: an identifiable program to develop or operate one or more general-purpose AI models or systems intended to match or exceed the leading general-purpose AI systems available worldwide at the time of the agreement. The agreement or accompanying official description must explicitly identify this frontier or world-leading capability objective; merely calling a project innovative, advanced, or AI-related is insufficient. A project need not be limited to a single model or owned exclusively by the two governments.
  • Take part in development or governance: each government must have an assigned substantive role in the same project. Development includes conducting or commissioning model research, architecture design, training, or post-training. Governance includes a formal decision-making, approval, or veto role over the project's research program, training, safety requirements, access, or deployment. The two governments may have different roles; one may participate in development and the other in governance. Funding alone, infrastructure provision alone, advisory or observer status, and generally applicable regulatory jurisdiction do not count.

Resolves YES if qualifying official sources confirm that both governments have signed an agreement establishing a project meeting all the definitions above and document each government's qualifying role. The full agreement need not be public if official statements establish its signature and all substantive requirements. A project involving additional countries or private participants counts. Subsequent withdrawal, cancellation, or failure does not reverse a YES.

Exclusions (resolve NO).

  • AI diplomatic dialogues, general governance forums, standards bodies, evaluation exchanges, reciprocal testing, or compliance-monitoring arrangements that do not give both governments qualifying roles in a specific frontier AI development or operating project.
  • Cooperation limited to chips, minerals, energy, datacentres, supply chains, or distribution of existing models without qualifying participation in the frontier AI project itself.
  • A joint safety research project that only studies or evaluates other developers' models, rather than developing or operating its own qualifying frontier system.
  • Private US–China collaborations without both national governments' participation under the signed agreement.
  • Proposed agreements, signatures by only one government, and non-specific commitments to future joint projects.

Resolving source. The signed agreement published by either government or an official treaty depositary; or official statements from the US White House, Department of State, or other participating federal agency, and China's Ministry of Foreign Affairs, State Council, or other participating central-government ministry or agency. A published agreement signed by both is sufficient. If the agreement is unpublished, both governments must officially confirm signature, and their official accounts must together establish all substantive requirements without contradicting one another. These are narrower resolving sources than Public confirmation (A2); media reports alone do not count. Apply the Resolution window and Resolution date (A3); where multiple publications are needed, use the date of the last required publication.

Download all forecasts (JSON)