A lab says it cannot verify its newest model's alignment itself

Current status (as of 2026-10-12): nothing qualifies yet

I checked every relevant developer document I could find from on or after 2026-09-16. None meets both conditions.

• Claude Opus 5.5 system card (Sep 22, 2026), https://anthropic.com/claude-opus-5-5-system-card. I downloaded the full 230-page text and searched it for phrases like "cannot verify". Section 6.1.3 does have an earlier model (Claude Mythos 5.1) review the alignment assessment. But Anthropic calls this "an experimental source of additional assurance", and the authors made their own judgement on the review. The rest of the assessment rests on "manual expert review of hundreds of transcripts alongside extensive model-assisted review", using investigator and judge models. The limitations sections only acknowledge leftover uncertainty, such as the model noticing it is being tested and gaps in coverage. That uncertainty is explicitly excluded by the question. The Sonnet 5.5 (Sep 28) and Haiku 5.5 (about Oct 8) system cards use the same "experimental" wording.
• Anthropic, "Measurements for understanding the pace of AI development" (Sep 17), https://www.anthropic.com/institute/measuring-pace-of-ai-development. Claude "leads" 26% of Anthropic's research and engineering work. About 100,000 flagged transcripts a week are handled by model-based monitors, and about 50 are escalated to humans. This is automation chosen for scale and efficiency, which the question excludes.
• Anthropic, "Investigating unintended model actions" (Oct 10): incident reporting only, with no qualifying language.
• OpenAI:
  • The GPT-6 Astra system card (Sep 3) predates the window. It says "If the model were to try to sandbag covertly, we would likely be unable to catch it" and uses GPT-5.x models as monitors. That is an admission of a limit, but not of relying on earlier models to establish alignment.
  • The October GPT-6 Sol/Luna material and "Building standards for the next phase of AI" (Sep 21) say fully autonomous self-improvement "should not [be pursued] unless and until it can be done safely". They warn about humans "unable to provide oversight on research processes they no longer understand".
  • OpenAI shelved GPT-6.1 Astra on Sep 28 over failed alignment tests, which shows it is willing to pause rather than lean on AI judgement.
• Google's Gemini 4 Argon announcement relies on monitoring the model's written reasoning and has nothing relevant.

Reference class and base rate

There is no precedent. For about two years, system cards have relied more and more on AI-run audits, AI judges and earlier models acting as monitors, and Anthropic's risk reports and system cards now include a model's review of the assessment itself. Even so, no covered developer has stated that its human staff cannot independently establish the assessment.

So the base rate for any given quarter is near zero today. The question is when capabilities make the admission both true and publishable.

The trajectory points to 2027–2029

• Dario Amodei's "We Must Pace the Frontier" (https://darioamodei.com/post/we-must-pace-the-frontier) says AI's ability to improve itself began "this summer".
• Anthropic's August 2026 Risk Report (https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf) says Anthropic "may cross" its threshold for automated AI research "in the coming year". It also expects that AI systems "might be responsible for much of the research and analysis that underpins risk assessment".
• METR, the outside evaluator, puts a 30% chance on research already running about twice as fast because of AI.
• OpenAI aims for an automated AI researcher by March 2028 (https://openai.com/index/research-acceleration-view-inside-openai/).

The most likely route to YES is an Anthropic risk report or system card after that threshold is crossed. Anthropic is unusually candid. A second route is a lab describing "scalable oversight" in practice: earlier, trusted models judging behaviour that humans cannot evaluate.

Main paths to NO

• Labs have strong legal, regulatory, investor and public reasons not to admit they cannot verify alignment.
• They have pledged to keep humans in control, so they are more likely to pause, or to describe oversight as human-led with AI assistance, or to cite volume or leftover uncertainty (both excluded).
• Labs may use the newest model to check itself, which fails the "earlier models" condition.
• Coordinated pacing, regulation, a plateau, or disruption such as government control or secrecy could also prevent it.

Estimate

I weighted three scenarios:

• Fast world (about 0.45 that human verification becomes truly infeasible by 2029 while labs keep publishing): about 0.5–0.55 chance of an explicit qualifying statement.
• Moderate world: about 0.2.
• Slow or disrupted world: about 0.05.

That gives about 0.3 by the end of 2031. I turned this into quarterly hazards, computed in Python: about 1% per quarter now, rising to about 2.6% per quarter through 2028, then falling to about 1.1% by 2031. The cumulative results are 7.4% by end-2027, 16.6% by end-2028 and 31.5% by end-2031.

Reading adopted: a strict one. The developer must explicitly say its staff are unable to independently establish the assessment of the newest model, and that earlier models' judgements are an essential basis for it. "Experimental" model reviews, volume or efficiency framing, and general uncertainty do not count.

Strongest case against: this world is moving extremely fast, and Anthropic is radically transparent, publishing incident reports and internal metrics and hosting outside evaluators inside the company. So a 2027 admission could come sooner than I model. Against that, labs' explicit pledges to keep humans in the loop argue for keeping the numbers below 50%.

Question. By the last day of each quarter, has a covered developer stated that its own staff cannot verify the alignment of its newest model and that it relies on the judgement of earlier models to do so?

Definitions.

• Newest model: a model or checkpoint that the developer identifies as its latest model, including an internal or unreleased model, at the time of the assessment described. A latest version within a named model family counts; an older model merely receiving a new product interface does not.
• Verify alignment: assess whether the model's goals, dispositions, and behaviour conform to the developer's intended human instructions, values, or safety constraints. This means a substantive alignment assessment, not a mathematical proof or a guarantee of zero risk.
• Own staff cannot verify: the developer explicitly states that its human staff, collectively using their available review methods, cannot independently establish the alignment assessment. Inability to understand or check the decisive evidence counts. Merely not reviewing every output, choosing automation for efficiency, or acknowledging residual uncertainty does not.
• Earlier models: one or more distinct predecessor models or checkpoints developed before the assessed model, whether produced by the same developer or another developer. Multiple instances of the assessed checkpoint do not count as earlier models.
• Relies on the judgement: the developer actually uses earlier models' evaluations or conclusions as an essential basis for its alignment assessment, while stating that its staff cannot independently validate the decisive judgements. Humans may retain formal sign-off or deployment authority; exclusive machine decision-making is not required.

Resolves YES if a qualifying developer statement establishes both the inability of its own staff to verify alignment and actual reliance on earlier models' judgement for the same newest model. The statement must describe an actual assessment, whether for internal use or external release. Both conditions may be established in explicitly linked developer documents. The developer need not claim that the earlier models' judgement is infallible or that the assessed model is aligned.

Resolves NO / Exclusions.

• No qualifying statement establishing both conditions has been published by the quarter-end.
• Predictions, hypothetical scenarios, research agendas, or statements that future superhuman models will require AI oversight.
• Automated auditing, model-generated evaluations, weak-to-strong supervision, or AI-led alignment research without the stated inability of staff to independently verify the resulting alignment assessment.
• General claims that alignment remains unsolved, that monitoring is incomplete, or that no model can be proved perfectly safe.
• An isolated task, benchmark, red-team exercise, or deliberately restricted human-review experiment, unless the developer explicitly applies its conclusion to its actual alignment assessment of the newest model.
• Third-party inferences about the developer's dependence on AI, without the developer's own qualifying statement.

Resolving source. Developer-authored system cards, risk reports, technical reports, official policy statements, or other statements on the covered developer's own domain, under Public confirmation (A2). Developer-authored written testimony officially published by a legislature, court, or regulator also counts; oral testimony alone does not. Third-party reports cannot substitute for the developer's own statement. Apply the Resolution window and Resolution date (A3): only qualifying publications on or after 2026-09-16 count, with resolution dated to the publication that first establishes both conditions. Once YES, always YES for later horizons.