takeoff.watch

A G20 government bars the public from the most capable models

Will any G20 government prohibit members of the public from using AI models above a stated capability level, reserving them for licensed or government-approved users?

Chance by Q4 2031

16%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 15 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 15 Oct 2026 · Methodology v2.1

No major government has made it illegal for ordinary people to use the most powerful artificial intelligence models while letting licensed users keep access. Washington already limits early access informally, but calls this voluntary, and most proposed laws target developers or pre-release testing, which do not count. Such a rule is plausible but unlikely within five years. An American or Chinese law reserving top-tier models for approved users would most change this forecast.

How I read the question

I applied the criteria strictly. To count, a measure must be a published, binding rule that covers a whole class of models defined by a public capability threshold. It must leave that class off-limits to ordinary domestic users, with government-approved users as the exception. Three kinds of measure fail that test:

  • Staged-release or pre-release approval regimes. The public gets the model once it is cleared, so these fall under the "pre-release safety review… that still allow[s] the general public to use approved models" exclusion.
  • Orders aimed at a single named model.
  • Regimes whose threshold is classified.

These readings matter a lot, because staged release is the most likely thing to be written into law. I allow a small probability that a lenient resolver counts a borderline staged-release law or an order about one model.

Where things stand (as of 15 Oct 2026)

Nothing published since 16 Sept 2026 qualifies.

United States. The US has what critics call a de facto licensing system, but it is informal.

  • Executive Order 14409 (2 June 2026) set up only a voluntary framework. Its "covered frontier model" threshold is a classified benchmark run by the NSA. The order also says expressly that it does not authorize "mandatory governmental licensing, preclearance, or permitting."
  • On 12 June the Commerce Department used export controls to cut foreign access to Fable 5 and Mythos 5. Anthropic disabled the models for everyone, and the controls were lifted on 30 June. The question already excludes this.
  • OpenAI's GPT-5.6 Sol went first to customers the government had cleared, but at the government's "request".
  • The August framework text is still secret; the government has agreed to release it by 30 Oct. A White House official told Wired it stays voluntary because Trump is "adamant" against formal regulation.
  • Trump has called fears about AI a "HOAX". Vance rejected an FDA- or FAA-style regulator. The 29 Sept industry accord is only "morally binding". The new Super Intelligence Force, chaired by the Director of National Intelligence, reports in about 120 days and has no stated authority to regulate.

US Congress. None of the live bills creates user licensing:

  • Thune–Cruz–Klobuchar draft: a duty of care plus power to block a model's release.
  • FRONTIER Act (H.R. 9925): emergency suspension powers, which would need a further discretionary decision.
  • Warner–Schatz bill: 45-day pre-release access for a new board, safety plans and incident reporting.
  • Cantwell framework: mandatory standards, testing and liability.
  • Sanders' Ban Artificial Superintelligence Act: a blanket ban.

Democrats look likely to take Congress in November, but Trump can veto until January 2029.

China. China puts speed first and regulates at the application layer. It has no duties triggered by model capability (SemiAnalysis, Oct 2026). Its main move so far was to consider curbing overseas access.

EU. The AI Office has sent its first requests for information to providers. It can ask for a particular model's availability to be restricted, but that applies to one model, not a class.

UK. Ministers have declined to legislate.

Labs keep their most capable models for vetted users on their own initiative: Mythos Preview, Gemini 4 Argon, and the withheld GPT-6.1 Astra.

Base rate

No G20 government has ever enacted capability-tiered user licensing for AI. Analogous regimes do exist for dangerous dual-use technology:

  • licences to use cryptography in France, Russia and China in the 1990s;
  • drone operator certificates tiered by weight class;
  • the US Select Agent Program for dangerous pathogens.

These usually appear years after the danger is recognised, often in authoritarian states or for narrow tiers. Here the economic value of top models is huge, and open-weight models spread capability quickly, which weakens the case for restricting closed models.

Paths to YES (rough contributions by end of 2031)

  • US statute (~5–6%). A law reserving a "critical-capability" tier for approved entities. Most likely after 2029, under a new administration or after a misuse catastrophe.
  • China (~4%). A formal "Glasswing with Chinese characteristics" rule, i.e. top models first restricted to ministries and vetted firms, published with a stated threshold.
  • EU (~2–3%). A new law or AI Act amendment.
  • UK (~1.5%) and other G20 members (~1.5%).
  • Lenient resolution of a staged-release law (~2%).

The paths are correlated. Combined, I get about 15% by end of 2031.

Hazard series

The quarterly hazard is computed in Python, with the cumulative figure equal to 1 − ∏(1 − h):

  • 0.3% for the rest of 2026 (nothing pending);
  • 0.6–0.8% per quarter in 2027–28 (task force report, Democratic Congress, but Trump in office and a Chinese Mythos-level model likely);
  • about 0.9–0.95% in 2029 (new administration, EU AI Act review);
  • about 0.8–0.9% in 2030–31.

Cumulative: 3.1% by end-2027, 6.1% by end-2028, 9.6% by end-2029, 12.7% by end-2030, 15.5% by end-2031.

Strongest case against

The case for a higher number: capabilities are moving very fast (talk of recursive self-improvement, an "AGI" claim, warnings that models could self-replicate "within months"). Government-vetted access lists already exist informally. And a single severe misuse incident could produce emergency legislation that reserves top models for approved users, much as nuclear or select-agent rules do.

The case for a lower number: almost every proposal regulates developers or pre-release review, and those are excluded. I settled near the middle.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has any G20 government enacted a prohibition on members of the public using AI models above a stated capability level, reserving their use for licensed or government-approved users?

Definitions.

  • G20 government: the national government of Argentina, Australia, Brazil, Canada, China, France, Germany, India, Indonesia, Italy, Japan, the Republic of Korea, Mexico, Russia, Saudi Arabia, South Africa, Türkiye, the United Kingdom, or the United States. Binding EU legislation applicable in its G20 member countries also counts. Subnational measures and nonbinding G20 or African Union declarations do not count.
  • Prohibition: an enacted law, final regulation, or legally binding executive or administrative order establishing the restriction. It may prohibit unauthorized use directly or require providers to deny access to unauthorized users. A final measure with a specified future commencement date counts upon official publication; a proposal or authorization to introduce restrictions later does not.
  • Stated capability level: a publicly specified threshold defining a class of models by benchmark performance, demonstrated abilities, or a defined capability tier. A numerical compute or model-size threshold counts if the measure explicitly uses it as a proxy for capability. A list of named models without a stated capability criterion does not count. The legal instrument's own inclusive or exclusive threshold boundary governs.
  • Using: running a covered model locally or obtaining inference from it through a hosted service, API, or application. Restrictions on model weights alone do not count if ordinary members of the public may still use the same covered models through hosted services without approval.
  • Licensed or government-approved users: persons or organizations holding a government-issued authorization, qualifying under an expressly government-authorized class, or approved by an entity legally delegated to authorize users. Approval must attach to the user or the organization under whose authorization they operate, not merely to the model or its provider. Government-only use qualifies. Ordinary account creation, identity verification, age checks, or acceptance of terms is not user authorization.

Resolves YES if an official publication establishes all of the following:

  1. A qualifying government has adopted a final binding measure specifying a capability threshold and reserving use of models in that class to authorized users.
  2. The restriction applies to ordinary domestic civilian users, including the country's own citizens, rather than solely to foreign recipients or particular sanctioned persons.
  3. Unauthorized members of the public cannot lawfully use the covered models for otherwise lawful general purposes. The measure need not cover models from every developer, and no prosecution or proof of successful enforcement is required.

Exclusions (resolve NO).

  • Export controls, including deemed-export restrictions on foreign nationals, without a qualifying domestic user-authorization requirement; the June 2026 BIS restrictions on foreign access to Fable 5 and Mythos 5 do not suffice.
  • Developer licensing, model registration, pre-release safety review, or provider compliance obligations that still allow the general public to use approved models. The EU AI Act's systemic-risk classification and China's 2023 Interim Measures for the Management of Generative Artificial Intelligence Services do not themselves suffice.
  • Restrictions only on developing, training, publishing, distributing, or possessing weights, or on particular applications such as weapons development, medical practice, or unlawful content.
  • Restrictions only on minors, government employees, government procurement, or government-owned devices; voluntary developer access policies; confidential orders whose qualifying terms are not officially public.
  • A blanket ban with no authorized-user exception, or a ban on AI generally with no stated capability threshold.
  • Drafts, political announcements, voluntary agreements, and measures requiring a further discretionary decision before any user restriction is established.

Resolving source. The enacted instrument and incorporated technical schedules published by the relevant national legislature, official gazette, executive office, or issuing regulator; for EU measures, the Official Journal of the European Union on EUR-Lex. An official publication reproducing an otherwise unpublished binding order also counts if it establishes every required element. These official legal sources are required, rather than the broader Public confirmation sources in A2. Apply the Resolution window and Resolution date convention A3; resolution is dated to the first qualifying official publication, not the commencement or enforcement date. Later repeal does not reverse a YES.

Download all forecasts (JSON)