The US and China verify each other's AI follows an agreed specification

Where things stand (16 Oct 2026)

Nothing that could resolve this exists yet. The September 2026 Trump–Xi state visit produced only a "U.S.-China Super Intelligence (SI) Dialogue to exchange views on risks and benefits", a next meeting "by November 2026", and a promised communication channel for AI incidents (White House fact sheet, Sept 2026). The two governments have no agreed written specification, no inspection mechanism, and no international body both have authorised. Working-level teams are reportedly trying to fill in the details of the incident framework so it can be announced at APEC in Shenzhen in November (SCMP, 16 Oct 2026).

The political will points the other way.

• US: Trump said the US is "not going to be putting on brakes… When you're leading, you don't open it up… I'm not looking to open it up" (AP/PBS, 26 Sept 2026). At the UN General Assembly he "totally rejects any attempt to construct a globalist scheme to control" AI (Carnegie, 30 Sept 2026).
• China: Beijing says it will not let the US "unilaterally define" risk thresholds, and the two sides distrust each other's commitments (Atlantic Council).
• Multilateral: Twenty countries plus the EU called for an international body to "set standards, enable verification". Neither the US nor China signed (Al Jazeera, 22 Sept 2026). The US is not in China's World AI Cooperation Organization (WAICO), and China is not in the US-led groupings.

Expert views agree. A Council on Foreign Relations analyst calls a deal "exceedingly unlikely under current strategic conditions" and without precedent (CFR, Sept 2026). A CNAS analysis says the verification infrastructure "will take years to develop", and the field has only about 50 engineers (CNAS, 22 Sept 2026). Even the AI Futures Project's advocacy scenario ("AI 2040: Plan A", which its authors say is not a prediction) puts a US–China deal in 2029 and the first, inference-only verification retrofits in 2029 (ai-2040.com).

Reference class and related forecasts

• Arms-control history: going from first dialogue to intrusive, affirmative verification of a rival's strategic technology has taken years to decades. INF was negotiated from 1981, signed in 1987 and inspected from 1988. The Chemical Weapons Convention took about 17 years before inspections began. The fastest case, New START, rested on an existing inspection regime. The US and China have never had any arms-control inspection arrangement.
• Markets:
  • Polymarket's "US x China agree to pace the AI frontier in 2026" stands at about 6% Yes, and that market asks only for an agreement.
  • Metaculus has about 25% that the US and China are party to any AI treaty that "controls or monitors" AI development by 2030. That is a much broader class.
  • Metaculus has about 5% for an AGI-limiting treaty signed before 2030.

This question is narrower than any of those. It needs all of the following:

1. an agreed specification of AI behaviour or operation;
2. inspection by the other government or a jointly authorised international body. Checks of compute, chips, facility safeguards or management processes are excluded;
3. completed affirmative conformity findings covering each country's national top tier of AI systems;
4. explicit affirmation by both governments on official sites.

Paths to YES

• A. Crisis-driven deal under Trump (2027–28). Incidents are escalating in this timeline: agent swarms, sandbox escapes, and early AI self-improvement acknowledged by the labs. A serious disaster could push Trump, who likes deals and has a warm rapport with Xi, into agreeing "red lines" with mutual testing. Testing-based verification is technically feasible now. I put this path at roughly 1% by 2031.
• B. A new US administration from 2029. Democrats are favoured at about 60% for 2028. Ro Khanna and others are pushing a "binding, enforceable, verifiable" treaty and an "International AI Inspection Agency". China would still have to accept intrusive checks, and inspections would have to be finished by the end of 2031. I estimate about 0.35 (US pursues) × 0.10 (China agrees by 2030) × 0.5 (inspections finished and affirmed in time) ≈ 1.75%.
• C. Very fast AI progress. AI-assisted verification could make an AI-mediated deal possible. I add about 0.7% that does not overlap with A and B.

Adding these gives roughly 3.5–4% by the end of 2031.

Hazard schedule

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

• Q4 2026: 0.02%. There is no specification or inspection, so it is essentially impossible.
• 2027: rising from 0.04% to 0.10% per quarter.
• 2028: 0.12–0.14% per quarter, with Trump still in office.
• Q1 2029: 0.12%, a dip for the transition.
• From Q2 2029: rising to 0.36% per quarter by late 2031, reflecting a possible new administration, maturing verification tools and growing salience.

Resulting cumulative probabilities: 0.3% by end-2027, 0.8% by end-2028, 1.5% by end-2029, 2.5% by end-2030 and 3.9% by end-2031.

Strongest case against my estimate

For higher: AI progress in this timeline is extraordinarily fast. Every major US lab has endorsed "pacing", the public strongly favours rules, and relations are in a détente. A single catastrophe could compress years of diplomacy into months. For lower: no rivals have ever inspected each other's crown-jewel technology this quickly. The question's exclusions remove the most likely forms of verification, which are compute- and chip-based. Each government would also have to publicly certify that the other side inspected its most capable systems. These roughly balance, so I keep a low single-digit estimate.

Question. By the last day of each quarter, have the governments of the United States and China each stated that inspectors from the other country or from an international body have verified that its most capable AI systems follow an agreed written specification?

Definitions.

• Governments: the United States federal executive government and the People's Republic of China's central government, acting through their heads, ministries, agencies, or officially authorized representatives. Subnational governments, individual legislators, political parties, and unofficial advisers do not count.
• Its most capable AI systems: the highest-capability tier of AI systems developed or operated within the government's national jurisdiction, including privately developed systems, as expressly identified by that government or by an inspection framework that government endorses. The statement must characterize the inspected scope as covering the country's most capable systems, not merely the most capable systems within one agency, company, application sector, or selected sample. Public disclosure of classified system names is not required; an unqualified official statement that the national highest-capability tier is covered suffices. No independent leaderboard ranking is required.
• Inspectors from the other country: inspectors officially appointed or authorized by the other national government; citizenship alone does not qualify.
• International body: an intergovernmental organization or multinational inspection organization formally authorized by both governments to perform the relevant verification. An ordinary commercial auditor does not qualify solely because it operates internationally. Either country may use either inspection route, and they need not use the same inspectors.
• Agreed written specification: an identifiable written document, version, or set of annexes accepted by both governments, setting requirements for the behavior, objectives, constraints, or operation of the AI systems being verified. A treaty is not required. Country-specific provisions within a shared agreement count; unrelated domestic standards do not. The full text may be confidential if both governments publicly acknowledge the document and its status as the agreed specification.
• Verified: a completed inspection, audit, test, or technical examination resulting in an affirmative finding of conformity to the specification for the covered systems. Remote examination counts. Verification need not constitute a mathematical proof or guarantee of future behavior, but agreement to future inspections or a statement of intended compliance does not count.

Resolves YES if qualifying official publications establish all of the following:

1. The United States government states that qualifying inspectors have verified conformity of its most capable AI systems to the agreed written specification.
2. The Chinese government makes the corresponding statement for its most capable AI systems.
3. Both confirmations concern the same agreed specification or framework, including any mutually accepted country-specific annexes.

A joint statement expressly making both confirmations counts. Separate statements may be published on different dates. Resolution occurs on the publication date of the last document necessary to satisfy these conditions, under the Resolution date convention (A3) and Resolution window. Both governments' qualifying confirmations must fall within the Resolution window; the agreement and inspections themselves may predate it.

Resolves NO / Exclusions. Resolve NO at a horizon if either government's qualifying confirmation is absent. Announcements of negotiations, inspection access, a proposed inspection body, inspection completion without an affirmative conformity finding, domestic self-certification, and inspections limited to less capable systems do not count. Verification only of training-compute limits, chip inventories, facility safeguards, or an organization's management processes does not establish that the AI systems follow a specification. Statements that systems comply except for unresolved requirements do not count as affirmative conformity findings. Advisory scientific assessments, including those of the UN Independent International Scientific Panel on Artificial Intelligence, do not count unless the body is authorized to inspect and the other requirements are met. An inspecting body's report alone cannot substitute for either government's statement; explicit official adoption of its relevant finding can.

Resolving source. Official statements, readouts, transcripts, reports, or jointly issued documents published by the White House (whitehouse.gov), U.S. Department of State (state.gov), another responsible U.S. federal executive agency, China's State Council (gov.cn), Ministry of Foreign Affairs (mfa.gov.cn or fmprc.gov.cn), or another responsible Chinese central-government agency, including their official successor domains. Treaty texts and inspecting-body reports may establish the specification, mandate, and inspection findings, but both governments must expressly affirm or adopt the respective verification result. This is a narrower source requirement than Public confirmation (A2); press reporting, silence, and lack of dispute cannot substitute for governmental affirmation. Apply Resolution date (A3) and the Resolution window.