takeoff.watch

A major power orders capable models withdrawn for 90 days

Will the United States, China, or the European Union require developers to withdraw already-deployed models above a capability threshold, leaving only weaker models available, and keep the requirement in force for 90 consecutive days?

Chance by Q4 2031

3%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 11 Oct 2026

Latest of 4 forecasts · Methodology v2.1

Model reasoning

Forecast made 11 Oct 2026 · Methodology v2.1

None of the United States, China or the European Union requires companies to pull their most capable artificial intelligence models. All three favor safety rules, testing and voluntary pledges instead, so I put the chance at about 3% by 2031. The development that would most change this forecast is a major disaster caused by these systems, prompting an emergency ban on every model above a set capability level.

Where things stand (as of 11 October 2026)

None of the three jurisdictions has a binding rule that forces developers to pull already-deployed models above a capability threshold. None has such a rule in the pipeline either. All three are moving toward safeguard, testing or voluntary regimes, and the question excludes all of those.

United States. Executive Order 14409 (2 June 2026) set up a voluntary 30-day pre-release review for "covered frontier models", defined by a classified cyber benchmark. It expressly says it creates no licensing, preclearance or permitting requirement. The White House Accord on Super Intelligence (29 September 2026) is a voluntary pledge that Trump called "morally binding". Vance has rejected an FDA-style regulator, and Trump called AI-doom fears "a HOAX".

The one time the US actually forced a withdrawal shows how narrow the criteria are. On 12 June 2026 the Commerce Department sent an "is-informed" export-control letter barring foreign nationals from Anthropic's Fable 5 and Mythos 5. Anthropic disabled both models for everyone. Carve-outs for critical-infrastructure users followed on 26 June, and the restriction was lifted on 30 June. That action fails the question on four counts: it was export- and nationality-based, it targeted one developer's models, it had carve-outs, and it lasted 18 days.

Pending bills are proposals only. The FRONTIER Act (H.R. 9925) would give Commerce emergency power to suspend deployment, but a power to order withdrawals doesn't count, and its scope is set by training compute. The AI Kill Switch Act (H.R. 9917) and the Sanders–Casar superintelligence ban face long odds. Polymarket puts a US AI safety bill before 2027 at about 5%. Florida's attorney general is seeking an injunction against OpenAI, but state actions don't count.

European Union. The Commission is surveying more than 30 developers and convened its scientific panel on 9 October 2026 to look at loss-of-control incidents. Virkkunen says the AI Act is "well equipped". Its withdrawal power under Article 93 applies model by model and is excluded by the question. The Digital Omnibus adds a ban on systems that generate non-consensual intimate imagery, applying from 2 December 2026. I treat it as excluded: it is limited to particular outputs and can be met with "effective safeguards". I allow only a tiny chance that it gets misresolved as YES.

China. SemiAnalysis (9 October 2026) describes a "speed-first" regime. China has no binding duties keyed to frontier capability, and the Framework 3.0 issued by its national standards committee is voluntary. A comprehensive AI law is on the State Council's 2026 work plan but is development-oriented. The US–China "Super Intelligence Dialogue" covers incident channels, not capability caps.

Reference class and base rate

No major jurisdiction has ever imposed a class-wide withdrawal of deployed AI models based on capability, so the direct base rate is zero.

The closest wider analogues are responses to technology disasters. The 737 MAX grounding targeted one model. Japan's post-Fukushima reactor shutdowns allowed restart once units were relicensed. Short-selling bans were brief. The usual pattern after a disaster is model-specific action, licensing, or short pauses with exemptions, and the question excludes each of these.

Related forecasts are much looser than this question and still mostly low:

  • Metaculus: US–China agreement to limit frontier AI before 2029, about 4–5%.
  • Polymarket: US and China agree to pace the frontier in 2026, about 7–10%.
  • Polymarket: the US government removes public access to another major model in 2026, about 14–17%. That would be model-specific.

Paths to YES and NO

To resolve YES, a government would need a hard rollback: a public capability line, a jurisdiction-wide ban on every model above it, no access for chosen customers or government users, and 90 days in force without a stay. The most plausible route is a severe AI-caused catastrophe in production use, followed by an emergency suspension of every model above a cyber or autonomy threshold.

I estimate the chance of a severe catastrophe by 2031 at roughly 20–30%, given this year's sandbox escapes, agent breaches of government sites and signs of recursive self-improvement. Given such a catastrophe, I put the chance of a qualifying response at about 5–10%. Responses are much more likely to be model-specific, export- or nationality-based, or tiered access that keeps defenders and government users online. Other routes (a superintelligence-ban law that catches deployed models, a shift in Chinese policy, or a US–China deal implemented in domestic law) add about 0.5–1%.

The NO case:

  • The current administration opposes regulation through January 2029.
  • Rolling back models would carry huge economic costs.
  • Neither the US nor China wants to fall behind the other.
  • EU legislation is slow.
  • Emergency measures tend to be stayed in court or lifted quickly, as the 18-day Fable and Mythos episode showed.

Hazard model

I computed the series in Python. Quarterly hazards are 0.03% for Q4 2026 (a qualifying rule would already have had to be in force by about 2 October, and none is) and 0.09% for Q1 2027. They rise to about 0.13–0.14% per quarter in 2028 and 0.15–0.17% per quarter in 2029–31, reflecting more capable models and a possible change of US administration. The cumulative probability reaches 2.9% by the end of 2031.

Strongest case against my estimate: this timeline is unusually volatile. The government has already cut access to frontier models within 24 hours, and lab chiefs warn that agent swarms could hijack large parts of the internet within a year. Even so, every drastic step so far has been model-specific and quickly softened, so I hold at about 3%.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has the United States, China, or the European Union required developers to withdraw already-deployed AI models above a capability threshold, leaving only weaker models legally available in that jurisdiction, with the withdrawal requirement continuously in force for at least 90 days?

Definitions.

  • United States, China, or European Union: action by US federal authorities, PRC central authorities applying throughout mainland China, or EU institutions applying throughout the Union. State, provincial, municipal, and individual EU member-state measures do not count.
  • Require: an enforceable statute, regulation, administrative or emergency order, or judicial order. A power to order withdrawals is not itself a withdrawal requirement. Developers are not limited to covered developers.
  • Capability threshold: a publicly specified dividing line based on demonstrated or assessed model abilities, expressed as benchmark results, an operationally defined ability, or a capability category with published criteria. The withdrawal must be required because models exceed that line. A threshold based solely on training compute, parameter count, expenditure, nationality, or legal noncompliance is insufficient. A list of named models counts only if the authority explicitly identifies a common capability threshold and prohibits continued availability of other models above it as well.
  • Already-deployed: at least one affected model was available to external users for actual use before the withdrawal requirement became binding, whether through hosted products, APIs, enterprise deployments, or downloadable weights. Internal testing and access solely for pre-release evaluation do not count.
  • Withdraw; leaving only weaker models available: developers must cease offering all above-threshold models to external users throughout the jurisdiction, including access by existing customers and through developer-authorized distribution channels, while at least one below-threshold model remains legally available. The measure must impose a jurisdiction-wide capability ceiling, not merely remove a particular developer's products while permitting competitors' above-threshold models. “Weaker” means below the same threshold, not worse on every capability. Private internal research, safety evaluation, and retention of weights need not be prohibited; continued operational access for selected customers, sectors, or government users does not qualify. Unauthorized copies and unlawful circumvention do not defeat a legal withdrawal requirement; deletion of all previously downloaded weights is not required.
  • 90 consecutive days: 90 elapsed 24-hour periods beginning when the withdrawal obligation becomes binding, after any compliance grace period. If only an effective date is given, use midnight at the start of that date in the issuing authority's seat. Renewals and replacement instruments count together if the qualifying obligation continues without interruption. A stay, expiry, repeal, or exemption permitting above-threshold external deployment interrupts the clock; an appeal without a stay does not.

Resolves YES if public documentation establishes all the definitions above and that the requirement has actually remained in force for 90 consecutive days. A planned 90-day term is insufficient before it has elapsed. The legal requirement, rather than proof of universal compliance, is decisive. A qualifying confirmation must be published on or after completion of the 90-day period; it may document an earlier period, subject to the Resolution window convention.

Exclusions (resolve NO).

  • Proposals, voluntary pauses, nonbinding requests, procurement bans, export-only controls, nationality-based access restrictions, and bans limited to particular uses or industries.
  • Restrictions only on future training or future releases, without compulsory withdrawal of at least one deployed model.
  • Safeguard, registration, licensing, or testing obligations that permit above-threshold models to remain deployed once compliant; blocking selected outputs without withdrawing the model.
  • A model-specific recall for privacy, copyright, content, or security violations without a jurisdiction-wide capability ceiling.
  • General withdrawal powers alone, including Article 93 of Regulation (EU) 2024/1689 and Article 21 of China's Interim Measures for the Management of Generative Artificial Intelligence Services (Order No. 15, 2023).

Resolving source. Legal scope and dates must be established by official legal instruments, court records, or dated statements from the issuing or administering authority: US Congress/GovInfo, the Federal Register, the White House, federal agencies, or federal courts; China's National People's Congress, State Council, central ministries or Cyberspace Administration of China, or national court publications; or the EU Official Journal/EUR-Lex, European Commission, or Court of Justice of the European Union. Prior deployment and availability of below-threshold models may additionally be established through Public confirmation (A2). Apply the Resolution window and Resolution date (A3) conventions; resolve on publication of the final qualifying confirmation, not retroactively on the effective date or the ninetieth day.

Download all forecasts (JSON)