takeoff.watch

A regulator halts a lab's training run or deployment

Will a regulator in the United States, China, the EU, or the UK refuse or revoke a licence or authorisation, citing AI capability or safety, and so halt a covered developer's training run or the deployment of one of its frontier models?

Chance by Q4 2031

26%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 14 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 14 Oct 2026 · Methodology v2.1

Governments are stepping in more often on advanced artificial intelligence; this summer the United States forced Anthropic to pull two models for 19 days using export rules. But this question needs a formal refusal or cancellation of a legally required permit, with official safety reasons, and few such permit systems exist yet. I estimate about 2% by end-2026, rising to about 26% by end-2031. A law letting officials approve or block model releases would raise this most.

Current status (as of 14 Oct 2026)

Nothing has qualified since the question opened on 16 Sep 2026. Governments are intervening much more, but almost none of it takes the legal form this question requires.

  • The June precedent falls before the window, and it is ambiguous anyway.
    • On 12 Jun 2026 the Commerce Department sent Anthropic an "is-informed" export-control letter. It required a licence before any foreign national, inside or outside the United States, could access Fable 5 or Mythos 5.
    • Anthropic could not filter users by nationality, so it disabled both models for everyone. The controls were withdrawn on 30 Jun, after 19 days.
    • The letter imposed a new licence requirement. It did not deny an application or revoke a licence Anthropic already held. The question's definitions say "general legal permission ... without an identifiable licensing or approval decision does not count", which points against counting it. Commerce never published the letter. Its public grounds were vague ("diversion risks").
    • My reading: a repeat of this would qualify only about 35–40% of the time. Explicit licence denials or revocations with official grounds would clearly qualify.
  • The United States now has a "voluntary", secret pre-release review framework (Executive Order 14409, June 2026), plus requests such as OpenAI's limited GPT-5.6 preview. The September 2026 accord is non-binding. The new "Super Intelligence Force" is advisory, though on 9 Oct it declared incident reporting "not optional". Voluntary pauses, such as OpenAI's September training pause and its cancellation of GPT-6.1 Astra, are excluded.
  • Federal legislation has stalled. A Thune–Cruz–Klobuchar draft would let the Commerce Secretary approve or block frontier releases, but it is unreleased and contested. Trump calls the risk a "hoax", and Vance rejects an FDA-style regulator.
  • China has a real pre-approval regime (filings and security assessments with the Cyberspace Administration of China, CAC). But its frontier approach is speed-first and refusals are not published. Its April 2026 action against ByteDance AI apps produced only warnings and orders to fix problems.
  • The EU relies on AI Act enforcement powers, which the question excludes.
  • The UK has no licensing power. The Burnham government is weighing narrow, binding national-security legislation.

Reference class and base rate

From 2023 to mid-2026 there were no unambiguous qualifying events in the four jurisdictions; the June case is the one near-miss. The environment is now far more intense: rogue-agent incidents, claims of artificial general intelligence, and active governments. So I set the hazard well above that historical near-zero, at about 1.3–1.4% a quarter.

Paths, as a component hazard model

I computed the quarterly hazards in Python and accumulated them. Component totals through 2031, treated as independent:

Path Total by end-2031
United States ad hoc action (export controls), judged as qualifying ~9–10%
United States statutory approval regime, from about 2028 ~8%
China CAC suspending or revoking a frontier model's filing, published ~3%
United States licence-based action against a Chinese model ~2.5%
UK ~2%
US states ~2%
EU ~1.3%
Infrastructure or compute permits ~1%

The union under independence is about 25.6%. I shaded slightly for positive correlation and offsetting tail scenarios.

Near term. Polymarket prices, pulled from its gamma API on 14 Oct, put the chance that the US government removes public access to "another major AI model" in 2026 at 14.5% (Anthropic 10.5%, Chinese models 12.5%). I treat these as overpriced long shots. Multiplying by the chance such an action meets the licensing test gives a hazard of about 2% for the fourth quarter of 2026.

Later years. Hazards come mainly from formal regimes: a federal law (Polymarket gives 34.5% that a broad federal AI framework becomes law before 2028, but most candidates lack blocking power), possible UK legislation, and a post-2028 administration.

Strongest case against

Governments almost always act through requests, voluntary frameworks, orders and investigations, not formal licence refusals. Developers negotiate or withdraw rather than accept a published denial; compare how rarely the foreign-investment review committee formally blocks deals. Secrecy also makes it hard to get official grounds into the public record. These points could justify about 12%.

Strongest case for a higher number

Recursive self-improvement is reported, incidents are escalating, and the government has already shown it will force models offline. Formal licensing of deployment or training could arrive quickly after a crisis. These points could justify 40% or more. I settle at about 26% by end-2031.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has a regulator in the United States, China, the EU, or the UK refused or revoked a licence or authorisation, citing AI capability or safety, and thereby halted a covered developer's training run or the deployment of one of its frontier models?

Definitions.

  • Regulator: a public administrative authority empowered to grant, refuse, or withdraw the relevant permission, at national, EU, member-state, state, provincial, or local level in one of the named jurisdictions. Courts, legislatures, private certification bodies, and advisory institutes acting without such powers do not count.
  • Licence or authorisation: an identifiable legally required permission to conduct the training or deployment, including mandatory approval of a security assessment or regulatory registration where acceptance is a prerequisite to proceeding. A permission for indispensable infrastructure or compute also counts if its refusal or revocation demonstrably halts the identified run or deployment. General legal permission to conduct business, without an identifiable licensing or approval decision, does not count.
  • Refuse or revoke: an operative administrative decision denying an application or renewal, or withdrawing an existing permission. Temporary withdrawal or suspension of the permission counts; a pending application, investigation, request for information, or threatened refusal does not.
  • Citing AI capability or safety: the decision's stated grounds expressly include the model's capabilities, harmful outputs or behaviour, misuse potential, loss-of-control risk, or inadequacy of safeguards or safety evidence. This need not be the sole ground. National-security or public-order risks attributed to the AI's capabilities or outputs count. Generic references to safety in a statute do not suffice.
  • Frontier model: a general-purpose model identified by the developer or regulator as among that developer's most capable models at the time of the decision, or as a planned successor intended to match or exceed them. Smaller, older, or application-specific models do not qualify merely because the developer is covered. The frontier requirement applies to the deployment branch; the training branch covers any identified model-training run of a covered developer.
  • Halt: actual cessation or prevention of commencement of an identified training run or of making the identified model available in the regulator's territorial jurisdiction. A blocked launch counts if a concrete proposed deployment is identified in the decision or developer statement. A territorial halt counts even if training or deployment continues elsewhere. There is no minimum duration. Restricting individual features or uses while the same model remains generally available in that territory does not count.

Resolves YES if qualifying public documents establish all of the following:

  1. A regulator has issued the refusal or revocation defined above, identifying a covered developer, its controlled operating entity, or a provider of indispensable infrastructure or compute for its identified run or deployment.
  2. AI capability or safety is expressly cited as a ground for that decision.
  3. The refusal or revocation has taken effect and has actually halted the run or deployment. A legally operative refusal that prevents a specifically proposed run or launch from starting satisfies this requirement; an order to cease an existing activity needs confirmation that it ceased.

An appeal does not prevent YES unless the decision is stayed before any halt occurs. Later reversal, permission to resume, or relocation does not undo a qualifying YES.

Exclusions (resolve NO).

  • Fines, compulsory testing, systemic-risk designations, general prohibitions, or withdrawal orders that do not refuse or revoke an identifiable licence or authorisation.
  • Voluntary pauses, commercial withdrawals, or compliance delays without a qualifying administrative decision.
  • Decisions based only on copyright, training-data consent, privacy notices, competition, corporate ownership, trade restrictions, energy supply, land use, or ordinary facility safety, without a stated AI-capability or AI-safety ground.
  • Denial of one customer's application or a sector-specific use that leaves the developer's model deployment generally available.
  • An announcement of future action without an effective decision and the required halt by the horizon.

Resolving source. The issuing regulator's published decision, order, official register entry, or official explanatory statement establishes the decision and its grounds. Eligible issuers include the European Commission acting through the AI Office, competent EU member-state authorities, the Cyberspace Administration of China and competent Chinese sectoral or provincial authorities, and competent US or UK licensing authorities, including successors. The halt may be established by those official records or by a statement on the covered developer's domain or a developer-authored report meeting Public confirmation (A2); company statements alone cannot establish otherwise unpublished regulatory grounds. China's Interim Measures for the Management of Generative Artificial Intelligence Services (effective 2023-08-15), Articles 17 and 23, illustrate assessments, filings, and permits that count only when they meet the definitions above (text). EU AI Act enforcement powers alone are not a licensing refusal (Commission guidance). Apply the Resolution window and Resolution date (A3); if multiple publications are required, use the publication date of the last document needed to establish all conditions.

Download all forecasts (JSON)