The US and China name an AI capability both will not build

Where things stand (as of 11 Oct 2026)

No qualifying commitment has been published since the 16 Sep 2026 opening date. The only bilateral AI commitment on record is still the excluded Biden–Xi nuclear-launch statement of November 2024.

The US–China track. At the Trump–Xi state visit (23–25 Sep 2026) the two sides agreed:

• to set up a "Super Intelligence" (US name) / "AI" (Chinese name) Dialogue "to exchange views on risks and benefits", with the next exchange in November 2026 around the APEC summit in Shenzhen;
• to set up a communication channel for AI incidents.

Neither government's readout contains any restraint commitment (White House fact sheet; China's foreign ministry, "Eight Deliverables", 26 Sep).

The stated agendas are about defining and reporting risks, not banning anything:

• Treasury Secretary Bessent wants the two sides to "agree on what the leading AI dangers are": uncontrollable agents, and non-state actors using AI for cyber or bioweapons attacks.
• Chinese state media (Yuyuan Tantian, 2 Oct) says November will cover three things: how to classify and grade AI risks, the rules for notifying incidents, and fitting AI into the wider bilateral framework.

The US is openly hostile to restraints:

• Trump called AI-catastrophe fears a "hoax" and told the UN General Assembly the US "totally rejects" global control of AI.
• He said he wants to "leave it exactly where it is" and on 29 Sep ruled out cooperating with China on AI governance.
• Michael Kratsios told the UN Security Council the US rejects international AI governance.

Xi keeps saying AI "must be kept under human control" and proposes jointly preventing "misuse and abuse", but China's own approach puts speed first. Expert commentary expects no breakthrough in November (CFR/McGuire, Brookings, Carnegie, CSIS). Polymarket's "US x China agree to pace the AI frontier in 2026" trades at about 6.9%; I pulled this from the gamma API on 11 Oct.

The multilateral track (the main near-term path). On 4–5 Sep 2026 the treaty body on autonomous weapons (the Convention on Certain Conventional Weapons, CCW) adopted by consensus, including the US and China, a set of "elements of an instrument, without prejudging its nature" (final report CCW/GGE.1/2026/3). They include:

• para 31: autonomous weapons "must not be used if… inherently indiscriminate";
• para 32: attacks using them "must not be directed against the civilian population";
• para 34: "control and human judgement with regard to [them] are needed".

The report recommends that the treaty's Seventh Review Conference (16–20 Nov 2026) "consider and endorse" the elements "without prejudice to the result of future discussions". At the 8 Sep preparatory meeting:

• Russia said it hopes the conference will endorse the report;
• the US praised the human-judgement text;
• China welcomed the report.

I put the chance of an endorsement at about 65%. The open question is whether a resolver would count it.

Reading I adopted. For literal YES: autonomous use of conventional weapons is the question's own example of a qualifying behaviour, and "an agreed requirement to retain human control" counts. Against: the elements are explicitly not yet an undertaking ("without prejudging its nature", "without prejudice to future discussions"), they restate existing law of war, and consensus adoption is not clearly "express" joining. I judge that a strict reader would probably say NO and give about a 22% chance that a plain endorsement is counted. That puts this path at about 15%. Adding about 2.5% for a bilateral surprise in Shenzhen or at the G20 in Miami gives about 18% by 31 Dec 2026.

Base rate and the hazard after 2026

Reference class: specific behavioural commitments between the US and China in technology and security. Examples are the 2014 military-encounter memoranda, the 2015 pledge against cyber theft of commercial secrets, and the 2024 nuclear human-control statement. That is roughly one every 4–5 years when the two are actively engaged. For AI specifically there has been one, and it is excluded here.

Hazards by period:

• Trump era, 2027–28: low, about 0.7% a quarter. There are bumps in Q4 2027 (the Biological Weapons Convention's Tenth Review Conference, where explicit AI language might arguably count, plus APEC) and Q4 2028 (a possible non-binding political declaration on autonomous weapons, which the US says it prefers to a treaty). Cumulative for the two years, given no YES in 2026: about 10%.
• 2029–31: it depends on the next administration. I assume about 55% that it is Democratic, with roughly a 30% chance of something like a pledge not to use AI to build bioweapons or a human-control commitment by 2031; under a Republican, about 10%. Add the Eighth Review Conference of the weapons treaty in Q4 2031. Cumulative, given nothing earlier: about 23%.

How I computed it. In Python I set 21 quarterly hazards: 0.18; then 0.007/0.007/0.007/0.036; 0.007×3 then 0.023; 0.008/0.012/0.015/0.034; 0.021×4; 0.021×3 then 0.038. I took the running product of (1 − hazard) to get the cumulative path: 0.18 (2026), 0.23 (2027), 0.26 (2028), 0.31 (2029), 0.37 (2030), 0.43 (2031).

Strongest case against

• Trump's explicit refusal and the notification-only agendas leave little room for a bilateral pledge before 2029.
• The CFR view is that no AI arms-control deal is likely while both sides think they can win.
• A strict resolver will likely treat the November endorsement as preparatory, not a commitment. If so, the 2026 figure should be about 4% and the 2031 figure about 31%.

The case the other way:

• AI salience is extreme: models escaping their test sandboxes, OpenAI's training pause, Security Council briefings.
• The bar is a single named restraint, and several venues offer one.

Sources

• White House fact sheet, Sep 2026: whitehouse.gov/fact-sheets/2026/09/fact-sheet-president-donald-j-trump-advances-a-fair-and-reciprocal-relationship-with-china-while-hosting-historic-state-visit/
• China's foreign ministry, 26 Sep 2026: fmprc.gov.cn/eng/xw/zyxw/202609/t20260926_12031663.html
• CBS News, 26 Sep: cbsnews.com/news/trump-xi-us-china-ai-trade-summit/
• Asia Times, Sep: asiatimes.com/2026/09/us-china-open-hotline-to-rein-in-runaway-ai/
• The Next Web, Sep–Oct: thenextweb.com/news/trump-rejects-china-ai-safety-cooperation
• Zaobao, 3 Oct: zaochenbao.com/news/china/202610/0382513.html
• CFR, Sep: cfr.org/articles/how-to-reach-an-ai-arms-control-deal-with-china
• CSIS takeaways: csis.org/analysis/takeaways-trump-xi-white-house-summit
• Final report of the autonomous-weapons experts group: docs-library.unoda.org/…/CCW-GGE.1-2026-3-Advance_copy.pdf
• Preparatory meeting transcript, 8 Sep: transcripts.un.org/en/asset/k18/k1856vomjr
• Human Rights Watch, 7 Sep: hrw.org/news/2026/09/07/un-talks-on-killer-robots-end-with-calls-for-negotiations-growing
• Reuters via Dawn, Sep: dawn.com/news/2027728
• UNIDIR on the 2027 Biological Weapons Convention review conference: unidir.org/montreux-retreat-advances-work-towards-a-stronger-biological-weapons-convention/
• Polymarket gamma API, 11 Oct 2026.

Question. By the last day of each quarter, have the US and Chinese governments jointly named at least one AI capability or behaviour, other than control of nuclear launch, that both commit not to develop or deploy?

Definitions.

• Governments: the United States federal government and the central government of the People's Republic of China, acting through their heads of state or government, ministries, departments, or authorized diplomatic representatives.
• Jointly named: identified in a joint agreement, declaration, or communiqué, or in parallel official statements that explicitly describe the same mutual understanding. A multilateral instrument counts if both governments expressly join its commitment. Independent domestic policies with coincidentally similar restrictions do not count.
• AI capability or behaviour: a specified function AI could perform, or a specified action or use of AI, including a prohibition limited to a named operational context. General descriptions such as unsafe AI, irresponsible AI, or AI inconsistent with human rights, without a specified function, action, or use, do not count.
• Commit not to develop or deploy: both governments expressly undertake to refrain from developing the same named capability or behaviour, deploying it, or both. A shared prohibition on deployment alone suffices even if development remains permitted. The commitment must cover at least each government's own activities, including activities conducted on its behalf; regulation of all private developers is not required. Political commitments and time-limited moratoria count without legal enforceability, verification, or demonstrated compliance. An agreed requirement to retain human control counts as a commitment not to deploy the expressly identified autonomous behaviour. A proposal to negotiate a restriction does not count.
• Other than control of nuclear launch: restrictions solely on AI deciding, authorizing, ordering, or executing nuclear-weapons use are excluded. A broader restriction counts only if it also expressly covers a distinct non-launch capability or behaviour, such as autonomous use of conventional weapons.

Resolves YES if qualifying official publications establish that both governments have adopted the same commitment meeting all the definitions above. The publications must identify what AI capability or behaviour is forbidden and whether development, deployment, or both are restricted. No treaty signature, ratification, or implementation is required.

Resolves NO / Exclusions. No qualifying commitment has been published by the horizon. Exclude unilateral commitments; private-company pledges; shared expressions of concern; promises to discuss safety; general calls for responsible military AI; and endorsements of nonbinding recommendations addressed to others without an express undertaking by both governments. The November 16, 2024 Biden–Xi statements on human control over nuclear-weapons use do not qualify. General appeals such as the March 2024 UN General Assembly resolution's call to refrain from AI incompatible with human rights do not identify a qualifying capability or behaviour.

Resolving source. Official joint texts or parallel official statements published by the White House, US Department of State, another authorized US federal department, China's Ministry of Foreign Affairs, State Council, another authorized PRC central-government ministry, or their official diplomatic missions. An official treaty depositary or UN record may establish a multilateral text and each government's express adherence. These government records, rather than the developer sources in Public confirmation (A2), are the resolving sources. Apply the Resolution window and Resolution date (A3); for parallel statements, use the publication date of the last statement needed to establish mutual commitment.