A government accepts a safety case for a superhuman model

Where things stand (16 Oct 2026)

A YES needs two things, both published on or after 16 Sep 2026 and both about the same model: (A) a government body's own official record accepting a developer's safety case, and (B) the developer explicitly claiming the model beats the leading human experts in every field it was tested on. Neither has ever happened for any AI model.

Element A: government acceptance. No regime does this today.

• US. Executive Order 14409 (2 Jun 2026) sets up voluntary pre-release cyber testing of up to 30 days. It says outright that it does not create "a mandatory governmental licensing, preclearance, or permitting requirement" (Foley Hoag, 29 Jun 2026; tech-insider). Lawfare (13 Oct 2026) calls the result a de facto licensing system, but the framework is secret; its text is due out by 30 Oct under the Protect Democracy lawsuit. Lawfare also notes "no official comment from the White House about the safety or security" of GPT-6 Astra. Sen. Budd (30 Jun 2026) said CAISI, now renamed CAISSI, had been told to stop publishing its frontier-model findings. The 29 Sep "White House Accord on Super Intelligence" relies on companies policing themselves, with outside auditors and no government enforcement role (presidency.ucsb.edu; Al Jazeera). Lifting the export controls on Anthropic's Fable 5 and Mythos 5 in June was effectively a permit, and it came before the window opened.
• UK. AISI says it is "not seeking to mandate requirements" (AISI blog, Oct 2026). It contributes to safety-case reviews with the labs, but not publicly.
• EU. The AI Office has had enforcement powers since August. It has sent its first formal requests for information to OpenAI and Anthropic (technology.org, 15 Oct 2026), but it has no pre-approval step and does not publish acceptances.
• Others. California SB 53/SB 813, New York's RAISE Act and Korea's AI Basic Act rely on transparency, filings or private verifiers, none of which counts. The Great American AI Act draft would route audits through private verification organisations.

Element B: the capability claim is far from current wording. Anthropic calls Fable 5 "state-of-the-art on nearly all tested benchmarks" and says Mythos 5 "matches or beats skilled human operators." OpenAI's GPT-6 Astra launch page says "most intelligent and aligned model." Independent tests show it trailing rivals in places (Kingy AI). Brockman's "AGI era" remark was an executive interview, which doesn't count. Progress is very fast, though. An internal OpenAI model "significantly more capable than GPT-6 Astra" solved the Navier–Stokes Millennium Prize problem (openai.com, Sep 2026). Pachocki writes that machine intelligence is "starting to exceed that of humans in transformative ways." OpenAI is aiming for an automated AI researcher by March 2028, and Anthropic says large gaps in judgement remain. Trump's renaming of AI to "Super Intelligence" (The Verge) is a label, not a developer's performance claim.

Reference class and model

There is no direct base rate: zero events so far. Regulators in nuclear, offshore and aviation accept safety cases routinely, but only under licensing laws that took years to build. My breakdown to end-2031:

• About 30% that some government body publishes a model-specific acceptance of a frontier safety case. Paths: a post-2029 US administration or Congress, the EU, the UK, or an international body.
• About 25–30% that some developer makes the explicit universal superhuman claim (mostly fast-takeoff worlds, 2028–30).
• About 50% that both attach to the same, published model.
• About 0.5% for narrow models, such as a medical AI approved by a regulator.

That gives roughly 5–6%. Quarterly hazards rise from 0.05% now to about 0.38% per quarter in 2030–31, reflecting growing capability and the chance of a new US administration in 2029. The cumulative series is computed from those hazards and reaches 5.6% at end-2031.

Strongest case against my estimate

Higher: recursive self-improvement is visibly under way. Superhuman-across-the-board models could arrive by 2028–29, and incidents are driving strong public pressure for licensing. Lower: developers have reasons not to claim superintelligence, with bills to ban it in the US and UK. Governments shy away from publicly endorsing safety arguments, and US reviews are secret. On balance I stay near 5%.

Ambiguities

I read "accept" strictly. Clearances, lifted export controls and the government's own evaluations don't count unless they expressly accept the developer's argument. "Exceeds expert baselines" doesn't meet the "leading experts" bar.

Question. By the last day of each quarter, has a government body accepted a developer's safety case for a model that the developer describes as exceeding the best human experts in every field it was tested on?

Definitions.

• Developer: any AI model developer, not limited to covered developers.
• Government body: a national, subnational, or supranational public authority, including a regulator, government department, or government AI safety or security institute. A private evaluator, industry association, or government-funded organization that is not itself a public authority does not qualify.
• Safety case: a developer's structured argument, supported by evidence, that the identified model is acceptably safe for a specified training, deployment, or use context. It need not bear the title “safety case”; an equivalent affirmative safety argument counts. A model card, benchmark report, or compliance checklist alone does not count unless it contains such an argument.
• Accept: an official, model-specific determination that the safety case is adequate, satisfactory, or otherwise establishes acceptable safety for the specified context. Acceptance can be restricted to a particular use or subject to continuing safeguards. Conditions requiring a further safety determination before acceptance do not count until that determination is published. Legal authorization to deploy is not required.
• Exceeding the best human experts in every field tested: the developer explicitly claims that the identified model performs strictly better than the leading human experts in all fields in which it has been tested. Equivalent wording counts, but “human-level,” “expert-level,” “better than most experts,” “most fields,” and higher aggregate performance do not. A field is a subject or activity domain, such as mathematics, biology, or programming, not an individual benchmark item. No minimum number or breadth of tested fields is imposed. A claim restricted to selected successful fields does not establish the universal claim. Independent verification of the claimed superiority is not required.

Resolves YES if qualifying public documents establish both the government body's acceptance and the developer's capability description for the same identifiable model/version. Both must be published within the Resolution window and by the relevant quarter-end; they may appear in separate documents and either order. A government statement quoting and expressly attributing the capability description to the developer can establish both in one document. The complete safety case and its confidential evidence need not be public if the official determination identifies the developer, model, and safety case sufficiently to establish acceptance.

Exclusions (resolve NO). Receipt or filing of a safety case; acceptance for review; participation in an evaluation; endorsement of a template or hypothetical safety-case sketch; a finding that only one component of a case is adequate; and a permit, registration, procurement decision, or finding of regulatory compliance that does not expressly accept the safety argument do not count. Predictions about a future model, claims concerning a different model version, and superiority claims made only by third parties do not count. Acceptance need not prove that the model is actually safe; subsequent incidents or withdrawal do not undo a qualifying YES. Otherwise resolve NO at a horizon if the qualifying record is absent.

Resolving source. Acceptance must be documented in the accepting government body's own official statement, assessment, decision, or published official record, naming the model. This is narrower than Public confirmation (A2); developer claims or press reports of government acceptance alone do not suffice. The capability description must be in a developer-authored publication under A2 or expressly quoted as the developer's description in the accepting body's official record. The safety-case definition follows the structured-argument usage in the UK AI Security Institute's “Safety cases at AISI”, not a requirement for acceptance by that institute specifically. Apply the Resolution window and Resolution date (A3); the qualifying date is publication of the last document needed to establish both elements. Once YES, always YES.