The US or China accuses the other of violating an AI agreement

Where things stand (as of 2026-10-11)

I found no qualifying accusation published since the question opened on 2026-09-16. There are now real AI agreements that a violation claim could target:

1. Lima, Nov 2024 (Biden–Xi). The two leaders "affirmed the need to maintain human control over the decision to use nuclear weapons" and agreed to develop military AI "in a prudent and responsible manner" (White House readout, 2024-11-16).
2. Washington state visit, Sept 2026 (Trump–Xi). Both governments published matching texts: a "China-U.S. AI Dialogue" (the US calls it the "SI Dialogue") with the next exchange in or by November 2026, plus a "bilateral communication channel for AI incidents." Sources: the White House fact sheet, 2026-09-25 and the Foreign Ministry's "eight deliverables", 2026-09-26. The dialogue sits inside the Bessent–He Lifeng economic channel. Incident definitions and triggers are not specified, and US officials say export controls and chips are excluded (Fudan analysis, 2026-09-29; ORF, Oct 2026).

Reading I adopted. A generic Chinese line such as "seriously violates the important consensus of the two heads of state" mostly does not count, even when the dispute is about AI. To count, the statement must tie the alleged breach to an identifiable AI commitment, such as the AI-dialogue or incident-channel consensus or the nuclear-AI pledge.

Reference class and base rate

• Direct base rate. Zero. Neither government has formally accused the other of breaching an AI commitment since the first ones in 2023–24. Those commitments were thin, though.
• Analogues show such accusations do happen once a commitment is substantive and relations sour:
  • The 2015 Obama–Xi cyber-theft pledge led to a Dec 2018 Pompeo–Nielsen joint statement accusing China of violating it (Washington Post, 2018-12-21).
  • Trade deals in 2019, 2020 and 2025 drew violation charges within months.
  • In 2026, China's Commerce Ministry has repeatedly said US measures "seriously violate the important consensus reached by the two heads of state" (2026-08-05) or "disregarded the Beijing consensus" (2026-06-13).
• AI disputes so far. In fights over distillation, the Commerce Ministry has already cited the leaders' AI-dialogue consensus, but only to urge or warn, not to allege a breach:
  • 2026-07-27: it "hopes the US will earnestly implement the consensus… strengthen AI dialogue."
  • 2026-09-09: the leaders agreed to AI dialogue, but China will "resolutely counter" if the US acts against Chinese AI firms.
• The trigger is being threatened. Bessent said sanctions and Entity List listings are "on the table" (TechCrunch, 2026-07-22). The NSA, FBI and CISA named six Chinese labs in an advisory (2026-09-08). No concrete penalty has been imposed yet.

Paths to YES and NO

Main YES path (roughly half the probability). The US imposes concrete penalties on Chinese AI labs, and the Commerce or Foreign Ministry replies that this violates the leaders' AI consensus. My rough chain: about 50% that such action happens by end-2028; about 0.25–0.3 that the reply names the AI consensus with breach language and is accepted.

Other YES paths:

• New, more substantive AI commitments, followed by noncompliance charges (cyber analogue). Possible subjects: distillation, non-state access to dangerous models, incident-notification duties, chips.
• Disputes over the incident channel after a real cross-border AI incident.
• A different US administration from 2029.

Main NO factors:

• Today's commitments are procedural and hard to "breach."
• China's usual AI rhetoric is "hegemonism" and "politicization," not breach of an agreement.
• Strict identification rules apply.
• The 2024 dialogue lapsed after one meeting, and this one could too.
• Both leaders want calm through the APEC summit in Shenzhen (Nov 2026), the G20 in Miami (Dec 2026), and the truce expiry on 2027-01-10.
• Expert forecasts put substantive US–China AI deals at about 5% (Metaculus treaty question; LEAP panel on a military-AI deal by end-2027).

Hazard model

I computed the series in Python as the cumulative product of quarterly hazards:

Period	Hazard per quarter	Reason
Q4 2026	1.2%	Summit season, little appetite for confrontation
Q1 2027	2.2%	Truce deadline
Rest of 2027	1.8%
2028	1.6%
2029–31	1.1–1.3%	AI consensus may go stale; new agreements possible

This gives about 8.5% by end-2027, 14% by end-2028 and 26% by end-2031.

Strongest case against. If resolvers count generic "violates the consensus" statements made in AI disputes, or if the US sanctions Chinese AI labs soon, my numbers are too low; that would push the 2031 figure toward 35–40%. If the AI dialogue quietly lapses, they are too high.

Question. By the last day of each quarter, has the US or Chinese government formally accused the other of violating an agreement between them on AI?

Definitions.

• Governments: the United States federal executive government and the central government of the People’s Republic of China, including their heads of state, ministries or departments, armed forces, and authorized diplomatic representatives or spokespersons acting in an official capacity.
• Agreement between them on AI: an identifiable agreement accepted by both governments that contains at least one mutual or unilateral commitment explicitly concerning artificial intelligence, its development, deployment, safety, governance, military use, or AI-specific computing resources. Treaties, executive agreements, memoranda of understanding, and nonbinding political commitments recorded in joint statements or official readouts count. A broader or multilateral agreement counts if both governments accepted the relevant AI commitment. Signature, ratification, enforceability, and a verification mechanism are not required. Agreements made before the question’s opening date count.
• Evidence of agreement: a published joint text, matching official acknowledgments by both governments, or an official readout explicitly reporting both governments’ acceptance of the identified commitment. An accusation alone cannot establish that an agreement existed. The full agreement need not be public if these sources establish the relevant commitment.
• Formally accuse: publish an official statement, diplomatic note, sanctions notice, government report, or official transcript expressly asserting that the other government has breached, violated, failed to comply with, or failed to honor the identified AI commitment. Equivalent wording in Chinese counts. The statement must attribute noncompliance to the other government, including its failure to meet an agreed obligation to regulate private actors; accusing only a private company does not count.

Resolves YES if a qualifying official accusation is published within the Resolution window and by the quarter-end. The agreement and the AI commitment alleged to have been breached must be identifiable from the accusation and supporting official records. Neither proof of the breach, acknowledgment by the accused government, sanctions, nor an adjudication is required. A denial, later withdrawal of the accusation, or later compliance does not reverse YES.

Exclusions (resolve NO).

• Warnings that a country might violate an agreement, requests for clarification, investigations into possible violations, and general expressions of concern without an affirmative allegation of noncompliance.
• Allegations of violating domestic export controls, intellectual-property law, general trade obligations, or international law without identifying a breached AI-specific commitment accepted by both governments. A dispute involving AI chips does not by itself make a general tariff or trade agreement an agreement on AI.
• Unaccepted proposals, unilateral policies, general endorsements of responsible AI without an identifiable commitment, and private commercial agreements.
• Statements by legislators, former officials, political parties, media outlets, anonymous officials, or personal social-media accounts unless subsequently published as an official government accusation through a qualifying source.
• Reports of a private diplomatic accusation without publication of the accusation or an official government account asserting it.

Resolving source. Official publications of the White House, US Department of State, Department of Defense, Department of Commerce, Department of the Treasury, Office of the US Trade Representative, or their diplomatic missions; and the PRC State Council, Ministry of Foreign Affairs, Ministry of National Defense, Ministry of Commerce, or their diplomatic missions. Official government archives and the US Federal Register also count. Successor executive offices and agencies count. This is a narrower source rule than Public confirmation (A2): third-party reporting alone cannot resolve the question, and the accused government’s dispute does not disqualify an accusation. Apply Resolution date (A3) and the Resolution window; use the publication date of the qualifying accusation, not the agreement date or alleged breach date.