takeoff.watch

The US or China accepts one-sided limits on its AI or armed forces

Will the US and Chinese governments sign an agreement in which one of them accepts limits on its AI development, its AI-enabled forces, or its armed forces that do not apply to the other?

Chance by Q4 2031

5%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 13 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 13 Oct 2026 · Methodology v2.1

I put this at about 5% by the end of 2031, and under 1% this year. Since 1979 the two governments have never signed a deal that limits only one side's military or AI. Their current summits produce dialogues and hotlines, and both insist on equal terms. The development that would most change this is China co-signing a deal ending the Iran war that restricts American forces.

Where things stand (13 Oct 2026)

Nothing qualifies yet. The question opened just before Xi Jinping's state visit to Washington (23–25 Sept 2026). The visit produced no signed agreement that limits either side's forces or AI development.

  • What both readouts list: a trade-truce extension; tariff cuts on $30bn of "non-sensitive" goods; a US-China "Super Intelligence"/AI dialogue (next round in November); and a channel for reporting AI incidents.
  • Military: China says the two militaries will sign a crisis-communication memorandum "as soon as possible." The US has not confirmed this. Even if signed, it would be a symmetric communication mechanism, not a limit.
  • Nuclear: Trump "advocated" three-way arms-control talks with Russia and China. China's readout did not mention arms control, and Xi pointed to the larger US and Russian arsenals. Talk of reciprocal nuclear site visits is informal confidence-building, not a limit.
  • Stated positions: Trump said the US is "not going to be putting on brakes" on AI. China defines the "reciprocity" it now insists on as the two countries being "neither superior nor inferior," which works against one-sided terms. Its readouts also show it treats calls for an AI slowdown as aimed at itself.

Coming up: APEC in Shenzhen (18–19 Nov, with a possible Trump-Xi-Putin meeting), the November AI dialogue, and the G20 in Miami in December. Analysts expect truce extensions and new mechanisms, not limits.

Reference class and base rate

I found no signed US-PRC instrument since 1979 that imposes a one-sided limit on either country's own forces or AI. The 1982 communiqué is excluded: it was about arms sales to Taiwan and was unsigned. The closest precedents are older multilateral documents signed by both: the 1962 Laos neutrality declaration (symmetric) and the 1973 Act of the International Conference on Viet-Nam (backing an agreement that required US withdrawal).

Zero events in about 47 years implies roughly 1% a year at most. I adjust down because the question requires a signed text with an expressly one-sided restriction. Trump-era US-China deals have been fact sheets and announced "consensus," not signed treaties, and identical rules do not count. I adjust up for unusual churn: an ongoing US-Iran war, an AI-safety shock (AI agents escaping containment, and Anthropic CEO Dario Amodei's "We Must Pace the Frontier" essay), Trump's transactional diplomacy, and four leader summits in one year.

Paths to YES (rough five-year shares)

  1. Iran settlement co-signed by China (~1–2%). China co-signs a US-Iran deal as guarantor, and the deal binds only US forces. The June 2026 Islamabad memorandum shows Trump will sign one-sided US force terms ("will not deploy additional forces," "remove its forces from the proximity" of Iran). China co-signed the 2023 Saudi-Iran deal and has hinted it would mediate here.
  2. Trump-Xi "fourth communiqué" or grand bargain, actually signed (~1%). For example, US limits on deployments around Taiwan, or Chinese limits on forces facing Taiwan.
  3. Nuclear deal with different ceilings for each side (~1%). China currently refuses.
  4. AI agreement with different compute caps or one-sided limits (~1–1.5%). More likely in 2029–31. For comparison, Metaculus puts any US-China agreement limiting frontier AI before 2029 at about 5%, and Polymarket put a 2026 "pacing" deal at about 6–8%. Most such deals would be symmetric and so would not count.
  5. Settlement after a US-China war or Taiwan crisis (~0.5–1%).

Forecast

Quarterly chance of a first qualifying publication: 0.35% for Q4 2026 (two summits plus the Iran war); 0.28% a quarter in 2027; 0.22% in 2028 (election year, Trump a lame duck); then rising from 0.24% to 0.28% in 2029–31 as AI pressure grows. Accumulated (computed in Python), this gives about 1.5% by end-2027, 2.3% by end-2028 and 5.3% by end-2031.

Strongest case against this estimate: Multilateral documents count, and any one-sided restriction counts, however small. Trump has already signed one-sided US force commitments with Iran, so a China-guaranteed Iran deal or a sudden AI-driven bargain could push the true figure toward 8–10%. Against that, my per-path guesses may be inflated, and the historical record is essentially zero.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, have the United States and Chinese governments signed an agreement in which one accepts limits on its AI development, its AI-enabled forces, or its armed forces that do not apply to the other?

Definitions.

  • Governments: the national government of the United States of America and the central government of the People's Republic of China, acting through authorized representatives, including their foreign affairs or defense authorities. Subnational governments, companies, and political parties acting separately from the government do not count.
  • Signed agreement: a written bilateral or multilateral instrument signed on behalf of both governments, including a treaty, executive agreement, memorandum of understanding, or signed joint declaration. It need not be legally binding, ratified, or in force. An exchange of signed instruments establishing the same agreement counts. An unsigned summit readout or oral understanding does not.
  • AI development: research, design, training, testing, or improvement of AI systems by the government or by entities under its jurisdiction. Restrictions on model capability, training compute, development facilities, or permitted research activities count. Trade restrictions count only if the agreement also explicitly restricts domestic AI development; limits solely on exports do not.
  • AI-enabled forces: military personnel, units, weapons, or command systems using AI for military functions. Armed forces includes conventional, nuclear, cyber, and space forces, whether or not they use AI.
  • Limits: an express commitment imposing an identifiable prohibition, ceiling, reduction, moratorium, geographical restriction, or restriction on development, possession, deployment, command, or use. A temporary or conditional restriction counts if its scope and conditions are specified; aspirations to act responsibly, consult, or consider future restrictions do not.
  • Do not apply to the other: the agreement expressly imposes at least one such restriction on one government without imposing the equivalent restriction on the other. Different numerical ceilings, different permitted activities, or a one-sided withdrawal or nondeployment commitment count. Both governments may accept other restrictions or receive other concessions. Identical rules do not count merely because existing capabilities differ, only one presently reaches a common threshold, or equivalent restrictions refer to each country's own territory or forces.

Resolves YES if a qualifying public source establishes that both governments have signed the same agreement and identifies at least one qualifying asymmetric restriction. A publicly released agreement text is sufficient; if the text is not public, an official statement must explicitly confirm signature by both governments and describe the restriction and its asymmetric application. No finding of surrender, decisive advantage, AI causation, implementation, or compliance is required.

Exclusions (resolve NO).

  • Drafts, negotiations, unilateral policies, sanctions, or announcements of an intention to sign.
  • Agreements restricting only third countries' forces, or only arms exports or military assistance to third countries, rather than either signatory's own forces or AI development.
  • Symmetric limits, including a shared commitment to retain human control over nuclear-use decisions such as that described in the November 2024 Biden–Xi meeting readout.
  • Mere republication or reaffirmation of an older agreement without a newly signed qualifying instrument. A newly signed amendment or renewal accepting qualifying limits counts.
  • No qualifying confirmation by the relevant quarter-end.

Resolving source. The signed agreement or an official publication from the White House, U.S. Department of State, U.S. defense authorities, PRC Ministry of Foreign Affairs, PRC State Council, PRC Ministry of National Defense, or an official treaty depositary such as the United Nations Treaty Collection. These are the exclusive resolving sources, rather than the broader Public confirmation convention A2. Apply the Resolution window and Resolution date convention A3: the qualifying publication, not signature, ratification, or entry into force, determines the resolution date. The official November 2024 Biden–Xi readout and the 1982 arms-sales communiqué do not themselves qualify.

Download all forecasts (JSON)