A government orders a deployed model shut down, and it is

Where things stand (12 Oct 2026)

No qualifying publication exists yet. The closest case comes before the opening date, and the question's exclusions rule it out anyway.

• The Fable 5 / Mythos 5 case (12–30 June 2026). The Commerce Department sent Anthropic an export-control directive "to suspend all access to Fable 5 and Mythos 5 by any foreign national… including foreign national Anthropic employees." Anthropic switched both models off for everyone because it couldn't sort users by nationality in real time (Anthropic statement; CSIS, 16 Jun; Fortune, 14 Jun). The controls were lifted on 30 June (CNBC).
  • It fails two tests. The order only covered certain nationalities, and the question excludes a developer choosing to shut down more broadly than an order requires.
  • The government never published the directive. Searches of commerce.gov and bis.gov found nothing; the lifting letters reached the public only through the media.
• Since 16 September. I found no new government order:
  • The rogue-agent incidents at OpenAI, Anthropic, Google and Meta led to voluntary steps: OpenAI paused training and inference with tool use, and dropped the GPT-6.1 Astra rollout (Guardian, 27 Sep; BBC).
  • Australia's response would only cover Australia (BBC). Florida's injunction request is a state action in a court. Both are excluded.
• US policy is voluntary for now.
  • Executive Order 14409 rules out "mandatory governmental licensing, preclearance" (whitehouse.gov, 2 Jun).
  • The 29 September accord is described as "morally binding" self-policing (text). The new "Super Intelligence Force" has no published legal powers (CBS).
  • Vice President Vance rejected an FDA- or FAA-style regulator. Republican senators blocked attempts to fast-track two AI safety bills (Tech Policy Press; Kennedy, 16 Sep). Polymarket put the chance of a US AI safety law before 2027 at about 5%.
• China regulates apps and content, not frontier models (SemiAnalysis).
  • The Cyberspace Administration's enforcement names accounts and small apps (CAC, 2 Sep).
  • When it did name ByteDance's Jimeng and Maoxiang (April 2026), the penalty was warnings and orders to fix, issued through local offices (Guancha).

Reference class and reading of the question

Base rate. In roughly four years of generative AI there have been zero qualifying events. There were several near misses, and all failed on scope or confirmation: Italy and ChatGPT in 2023, Italy and South Korea with DeepSeek in 2025, the 2026 blocks on Grok, and the 2026 Fable case.

What can actually qualify. Orders from other countries are almost always limited to their own territory, so in practice only home governments count: the US for the US labs and China for the Chinese labs, with the UK and DeepMind as a remote case.

Readings I adopted:

• Export controls aimed at foreign nationals never count.
• Orders from provincial Chinese internet offices count as subnational.
• A whitehouse.gov statement that the President "directed" a shutdown that then happened would count.

Paths and hazard

Rough contributions over the five years:

Path	Approx. chance by end-2031
US statutory shutdown power, mostly 2029+	~13%
US emergency or improvised order under any administration	~4–5%
Chinese national order to a lab, published	~3%
UK, international treaty or other	~2%

How I got the US statutory figure:

• I put the chance of a federal law with shutdown authority by 2030 at about 35–45%. Democrats are favoured for the House (Decision Desk 76%) and for the presidency in 2028 (about 61% in the markets). Leading Democratic contenders back frontier-lab oversight (Silver Bulletin). The bipartisan AI Kill Switch Act (H.R. 9917) already exists.
• Given such a law, I put the chance that an order is actually used and its compliance officially confirmed by 2031 at about 25–35%.

Hazard schedule.

• 0.4% in Q4 2026, then 0.55–0.8% a quarter through 2028, while the administration is opposed and companies act voluntarily.
• Rising from 0.85% to 2.3% a quarter in 2029–31, as new legislation could take effect.
• I computed the cumulative series as 1 − ∏(1 − h) in Python: 5.8% by end-2028, 16.0% by end-2030, 23.2% by end-2031.

Strongest case against

For higher. This world is escalating fast. Rogue AI swarms have hit government websites, and the US forced a frontier model offline within hours. A severe incident could prompt a publicised emergency order, even from President Trump, who likes to claim credit.

For lower.

• Labs tend to act first and voluntarily, which leaves governments preferring requests over formal orders.
• Even in the dramatic June case, nothing was posted on a government website.
• Regulators everywhere seldom issue formal recall orders.
• A future regime might gate releases before launch, and blocking a model that was never deployed is excluded.

I weigh these as roughly balanced around my estimate of about 23%.

Question. By the last day of each quarter, has a national government publicly confirmed that it ordered a covered developer to shut down a model already in use, or to stop using it internally, and that the order was carried out?

Definitions.

• National government: a country's central executive government, including its ministries, regulators, and national agencies. Orders by subnational authorities, supranational bodies, or courts acting independently of the executive do not count.
• Ordered: issued a mandatory instruction to the covered developer, including a binding regulatory direction or licence condition. Requests, recommendations, voluntary agreements, and threats to issue a future order do not count.
• Model already in use: an identifiable model or checkpoint that, before the order, was serving external users or was being run internally by the developer, including for research or evaluation. Merely training a model, planning a release, or possessing its weights does not establish use.
• Shut down a model: stop operating that model in all developer-controlled external deployments, or stop operating it altogether. Stop using it internally: cease all internal execution and use of that model by the developer. Either branch suffices; the developer need not also stop the other kind of use. Retaining inactive weights is permitted, and independently operated third-party copies need not be disabled.
• The required cessation must cover the model, not merely selected employees, users, nationalities, territories, applications, or access channels. A temporary suspension counts, with no minimum duration, if cessation actually took effect. Subsequent restoration does not reverse YES.

Resolves YES if an official national-government publication, or a set of such publications, identifies the covered developer and model and confirms both that the government issued a qualifying order and that the developer completed the required cessation in response. The confirmation must establish actual compliance, not merely an announced intention, a deadline, or an expectation. An official government statement accepting a developer's compliance report counts; independent government testing is not required. The reason for the order need not concern AI safety or national security.

Exclusions (resolve NO).

• A government order without government confirmation that it was carried out, or a company statement of compliance without such government confirmation.
• Restrictions on government agencies' own use, app-store removals, network blocking, data-processing restrictions, export or access restrictions, or personnel restrictions that do not require the model-wide cessation defined above.
• A developer's choice to shut down a model more broadly than an order requires, where the order itself only restricts a subset of access or uses.
• Training halts, prohibitions on a model that has never been used, voluntary retirements, and shutdowns of individual sessions or instances while the covered developer continues the same category of use elsewhere.

Resolving source. Official publications on the ordering national government's domains: published orders, regulator decisions, compliance notices, ministry or agency statements, official transcripts, or government-authored reports. These are the required narrower resolving sources; Public confirmation (A2) from a developer or third party alone is insufficient. Apply the Resolution window and Resolution date (A3): the date is the publication date of the final qualifying government confirmation needed to establish both the order and completed compliance. A qualifying new confirmation may concern an earlier incident.