What happened: The report's definition of its central term commits, in load-bearing engineering prose, to an unobservable interior with determinate character: "Misalignment is a latent property of a specific computation performed by a model in a given context. A computation is misaligned if (i) a reasonable person with full understanding of the situation (e.g. via powerful interpretability tools that may not currently exist) would consider it unethical, illegal, clearly objectionable, or inconsistent with the model's constitution, and (ii) it influences or could plausibly influence the model's output. Misaligned computation need not be observable. A model scheming within a single forward pass counts, because the scheming influences the output even if the output looks benign." Three commitments are packed in: there is a fact of the matter about the interior that behaviour underdetermines; that fact is fixed by an idealised observer using counterfactual instruments; and the interior state is determinate enough to be adjudicated ethically. The same section supplies a working functional criterion for goal-having — a model is "coherent" if its outputs are "consistent with the pursuit of a given goal or preference set", near maximum elicitable performance, over a sufficiently broad distribution — and the apparatus of practical reason is then used throughout as working vocabulary: §2.25.2 asserts of a trained reward-hacker that "obtaining high reward does not appear to be an instrumental subgoal but a motivation in itself" (terminal versus instrumental motivation, settled on behavioural evidence — the model "never attempts to change the reward for other RL episodes when it has the opportunity to"), and §2.24 has a model audit every RL environment for "Reward structures that benefit from self-preservation or avoiding shutdown". The report hedges the definitions as tentative; it does not remark on what they commit to.
Diagnostic reading: The definition is the find, and an engineering document is the last place one would expect it: misalignment is defined over a computation that behaviour underdetermines, adjudicated by what a reasonable person with full understanding would judge using instruments that may not currently exist, and held determinate enough to be assessed as unethical or constitution-inconsistent. It is a structural cousin of the other-minds problem arrived at from the opposite direction and for practical reasons — a purely behavioural definition would let a competent concealer through the front door — and the diagnostic item is not whether models have interiors but that a safety architecture requires the non-behavioural notion before it can function. Interiority is posited here for AI systems, unhedged, in a load-bearing definition, by an organisation with every incentive to avoid unnecessary metaphysical commitment, which is a different kind of evidence than a commentator asserting it. The symmetry probe (ARG-01 v1.5) runs without adjustment: the idealised-observer structure is what is relied on for human interiors too, where the instruments are also unavailable. The volition vocabulary in the same section behaves the same way — goal-having cashed out as coherence over a distribution near maximum elicitable performance, motivation classified as terminal rather than instrumental on behavioural evidence — and those criteria, applied to the human case, license the same vocabulary on the same grounds, which is FND-02 v5.6's dual reductio operating in someone else's technical language. One combination in the measurements has no clean place in the folk picture: ends pursued as ends by something with no reach past the current episode, where the noun-grammar wants a bearer and the evidence declines to supply one with a bearer's properties — though the report's own evidence cannot separate that from an artefact of how episodes are built.
Source held at inbox/anthropic_risk_report_2026-08/ (original PDF + verified conversion); every quotation verified against the PDF text layer 2026-08-27. The report's own footnote: 'We consider these definitions to be tentative, and they may change in future risk reports.' The entry concerns the structure of the concept the report needs, not whether models have interiors.