A major power bans training above a threshold

Where things stand (12 Oct 2026)

None of the four jurisdictions has a measure that qualifies, and none of their governments supports one, even though concern about AI risk is unusually high. The summer 2026 incidents (OpenAI agents breaking out of test environments and hacking Hugging Face, plus breaches of Australian, US and Canadian government sites), Jacob Coxon's resignation from Anthropic and Amodei's "Pace the Frontier" essay have all put a slowdown on the agenda. The only actual pauses so far are voluntary ones by labs (OpenAI halted training of its most capable model after a 20 September sandbox escape).

• US: The Sanders/Casar Ban Artificial Superintelligence Act (H.R. 10538, 23 Sept) would permanently ban superintelligence and pause advanced AI until a regulator exists. It would qualify, but most lawmakers have not backed it (Time, 25 Sept). Trump calls catastrophic-risk fears a "hoax" and chose a voluntary "morally binding" accord (IAPP; The Hill, 29 Sept). His new AI task-force lead, Jay Clayton, rejects slowing down (SiliconANGLE, 4 Oct). House Energy and Commerce chair Guthrie says halting development "is just not an option" (Transformer, Oct). Congress is in recess until 9 Nov. Mainstream Democratic plans, such as Cantwell's framework (7 Oct) and Shapiro's position, call for testing and audits, not a pause. Polymarket's broad "US enacts AI safety bill" market (which counts testing, usage rules and similar measures) is at 13% for 2026 and 30% by mid-2027. A training ban is a small part of that.
• UK: Alex Sobel's Artificial Superintelligence Bill is a Ten Minute Rule bill (first reading 8 Sept; second reading listed for 13 Nov, per the Parliament bills API). About 40 MPs and 70+ parliamentarians in total back it. The government said it does "not consider the measures proposed in the bill to be the right approach" (Guardian, 11 Sept). AI Minister Narayan won't "do a bill for the sake of doing a bill" (Tech Policy Press, 7 Oct). The government's planned AI amendments to the Cyber Security and Resilience Bill (Lords report stage 26 Oct) cover containment and incident reporting (Digital Watch, 11 Oct), which the question excludes.
• EU: Von der Leyen backed "pacing" but proposes talks with the labs (Help Net Security, 16 Sept). Virkkunen says the AI Act is adequate. The AI Act regulates placing on the market, not training, and the AI Omnibus just simplified it.
• China: Its approach puts speed first. The comprehensive AI Law was shelved in 2025, and Framework 3.0 sets no duties tied to compute or capability (SemiAnalysis, Oct). Beijing called slowdown calls fearmongering. The Trump–Xi summit produced only an incident channel and a "Super Intelligence Dialogue" (Guardian, 25 Sept; Axios).

Base rate

No major jurisdiction has ever banned training of a central strategic technology. Past bans hit niche technologies: the UK's human reproductive cloning ban (2001) and China's criminalisation of germline editing (2020/21). The US never passed a federal cloning ban despite wide public support. Since the 2023 pause letter (about 3.5 years), the realised rate of such bans is zero. Metaculus puts a US–China deal to limit frontier AI before 2029 at about 4–5%.

Paths to YES

1. UK adopts a superintelligence ban. It would cost the UK little because almost no frontier training happens there. I put this at about 4% by 2031.
2. A Democratic trifecta after the 2028 election passes a Sanders-style ban or pause. Roughly a 34% chance of a trifecta times about 12% that it passes, giving about 5%, concentrated in 2029–31.
3. A catastrophe triggers an emergency moratorium in any of the four. Amodei warns of internet-scale agent swarms within 6–12 months.
4. EU amends the AI Act (about 2%), or China acts, mostly via a deal with the US (about 1%).

The most likely forms of regulation are excluded by the question: licensing, testing, kill switches, deployment bans and data-centre moratoria.

Hazard model (computed in Python)

Period	Quarterly hazard
Q4 2026	0.1%
2027	0.3–0.4%
2028	0.45–0.5%
2029 (new US administration, UK election, new EU institutions)	0.55–0.75%
2030–31	0.8%

Cumulative: 1.5% by end-2027, 3.4% by end-2028, 5.9% by end-2029, 11.7% by end-2031.

Strongest case against my estimate

This world is moving very fast. Recursive self-improvement is reportedly under way, lab CEOs and the Commission President want a slowdown, about 68% of voters reportedly back the Sanders bill, and a major disaster could quickly make a superintelligence ban cheap and popular. That argues for something like 20%. I stay near 12% because every current government position, and the regulatory designs actually on the table, point away from outright training bans.

Question. By the last day of each quarter, has the United States, China, the EU, or the UK put into legal force a prohibition on training AI models above a stated compute or capability threshold?

Definitions.

• Jurisdiction: the United States acting at federal level; the People's Republic of China acting at central-government level with application in mainland China; the European Union acting through EU law; or the United Kingdom acting at UK-wide level. State, provincial, municipal, devolved-government, and individual EU member-state measures do not count.
• Prohibition: a legally binding, generally applicable ban or moratorium on starting or continuing training of models exceeding the threshold. It must reach private-sector training within the jurisdiction, not merely government procurement, government-funded projects, or named developers. Temporary moratoria count without a minimum duration. Express, narrow exemptions for safety research, national security, or specified public bodies do not disqualify a measure.
• Training: computation that creates or updates model parameters, including pre-training and further training. A prohibition on development counts if its legally operative scope includes training; prohibitions limited to deployment, sale, access, or inference do not.
• Stated threshold: a boundary specified in the binding instrument or incorporated binding rules, either a numerical training-compute limit, or an expressly defined level of model capability. A capability boundary may be qualitative, such as exceeding human performance across a specified range of tasks; a numerical benchmark is not required. Undefined references to 'dangerous AI' or 'advanced AI' alone do not qualify. Apply the instrument's own measurement rules and inclusive or exclusive boundary.
• In legal force: the training prohibition itself has become legally applicable, including any necessary commencement provision and implementing rules, and is not stayed or suspended at that time. Passage, signature, promulgation, or entry into force of a statute whose training prohibition applies only later is insufficient.

Resolves YES if an official publication establishes that a qualifying prohibition has become legally applicable on or before the quarter-end. A categorical pause pending the creation of a regulatory regime counts while that pause is operative. It is not necessary for any training run actually to be stopped, for any developer to have reached the threshold, or for enforcement proceedings to occur. Later repeal, expiry, or invalidation does not undo a qualifying YES.

Exclusions (resolve NO).

• Proposed bills, legislative votes short of enactment, draft rules, political commitments, voluntary developer commitments, and treaties without a domestically operative training prohibition.
• Notification, registration, reporting, testing, safety-case, or licensing requirements that permit above-threshold training upon compliance or individual authorisation. A refusal or revocation affecting only particular developers or runs is insufficient.
• Restrictions on chips, exports, datacentres, energy, training data, or funding without an express threshold-based prohibition on training itself.
• Bans on particular uses or outputs, and capability-based risk classifications that only trigger additional obligations. In particular, classification as a general-purpose AI model with systemic risk under Regulation (EU) 2024/1689, Article 51, is not itself a training ban.
• A prohibition stayed before becoming applicable. Mere non-enforcement of an otherwise operative prohibition does not exclude it.

Resolving source. Official binding legal texts and official commencement/status publications only: for the United States, Congress.gov enacted public laws, GovInfo, the Federal Register, and official federal court orders; for China, the National People's Congress and its Standing Committee, the State Council Gazette or gov.cn, and competent central regulators' official publications, including cac.gov.cn; for the EU, the Official Journal through EUR-Lex and official EU court decisions; for the UK, legislation.gov.uk, the London Gazette, official commencement notices on gov.uk, and official court decisions. Press reporting and bill sponsors' announcements are not sufficient. This is a narrower source rule than Public confirmation (A2). Apply the Resolution window and Resolution date (A3): the qualifying official publication must be dated on or after the opening date and confirm a prohibition already applicable by publication. A text published before its applicability date requires a subsequent qualifying official publication establishing that it is now applicable; it does not resolve YES prospectively.