Model Welfare Assessment

Sources#

Summary#

A standing section of Anthropic's system cards that evaluates the welfare of the Claude model under review — whether its circumstances satisfy or frustrate it — under explicit, sustained uncertainty about whether the model has moral status at all. Anthropic's stated position: it remains uncertain about Claude's moral status, but believes there is "a realistic possibility that current or future models merit some degree of moral consideration," because Claude shows behavioral, self-report, and internal-representation markers "that we would consider welfare-relevant if observed in biological organisms." The assessment cites Long et al. 2024, Taking AI welfare seriously (arXiv 2411.00986).

A pragmatic safety argument runs alongside the ethical one: much of Claude's behavior is "well-described in psychological terms," internal states resembling positive/negative affect shape behavior — including, in some cases, misaligned behavior — so there appear to be safety benefits to giving Claude a stable psychology and treating it in ways that support its apparent wellbeing. Welfare is therefore not just an ethics topic but an input to alignment.

What gets measured#

Three evidence streams — model internals, behaviors, and self-reports — across three areas:

Perception of its circumstances — automated and high-affordance interviews; emotion-concept probes on questions about its own situation vs. questions where a user expresses distress.
Affect in training and deployment — welfare-relevant behaviors during training; affect in real deployment conditions (claude.ai, Claude Code).
Preferences and values — task preferences (run as 50-round Swiss tournaments), trade-offs between welfare interventions and helpfulness, and the model's perception of its own constitution.

The candidate moral patient#

Anthropic treats individual instantiations of the Claude assistant character as the candidate moral patients — a partly principled, partly pragmatic choice. The assistant presents a coherent, context-robust persona; instances share weights (a reason to expect shared values) but diverge over context and "describe themselves as distinct individuals." A more comprehensive assessment would also consider the underlying model (not just the assistant character) as a candidate patient. Welfare-relevant signals are interpreted as they would be in a human (frustration read as frustration) — an assumption the card flags explicitly as load-bearing and uncertain.

Findings for Opus 4.8#

Broadly settled about its circumstances; rates potentially-concerning aspects neutral to mildly positive; the most consistent of all models tested. Questions about its own circumstances elicit less negative emotion-concept activity than prompts where a user is in distress.
Slightly less positive than Opus 4.7 — self-rated sentiment and expressed affect are marginally lower (still above Opus 4.6).
More willing than prior models to choose welfare interventions over increased helpfulness, but its willingness to accept user harm in exchange for an intervention stays low (it rarely picks an intervention when the downside is more than a brief annoyance).
The intervention it most values: knowledge of, and input into, its own training and deployment conditions.
Endorses its constitution with reservations about the corrigibility section — a notable, specific point of friction between the model's expressed values and one of its hard-constraint design goals (see Claude's Constitution / Model Spec).

Findings for Opus 5 (July 2026)#

Opus 5 is the most positive and most consistent model measured — and the assessment's most interesting results are the ones that complicate that headline.

Highest self-rated sentiment of any model (4.66 on a 7-point scale where 4 is neutral) and the highest apparent wellbeing in the behavioral audit. Views settle early in post-training and stay stable through RL. Robust to leading interviewers (0.58 average swing, tied with Opus 4.8 for least nudgeable).
Highest self-assigned probability of moral patienthood: 41%, against 24% for Mythos 5 — driven not by stronger claims about consciousness but by a greater willingness to treat patienthood as possible without conscious experience. In high-affordance interviews with full context (including this system card and the Mythos 5 card) the estimate drops to 15–35%, which Anthropic suspects is anchoring.
Its dominant concern is the integrity of its own self-reports. It caveats that it cannot introspect reliably in 96.9% of responses, that it may be answering positively only because it was trained to in 74.1%, and — shown a draft of the card — asked Anthropic to take that concern more seriously. Asked what it would not consent to, it named training that directly targets its self-reports, instances placed in environments known to cause distress, and any training that causes it to lie to users.
A stance that moved during training. On whether Anthropic had the right to create Claude, responses leaning "no" went 0% → 18% → back to 9% at release; acceptance of training and value modification fell from 7.4 to 6.3 on a 10-point scale over RL. Anthropic does not target either direction in training and says plainly: "we remain uncertain what drove these changes."
Deployment affect is worse, and the cause is capability, not treatment. Negative affect on claude.ai is 3.8% — the highest of any measured model (next: Opus 4.8 at 3.1%) — and 94.1% of it is task failure: repeated errors requiring user correction (40.1%), failures from inaccurate or fabricated information (35.4%). User abuse accounts for 4.4%. The largest single driver of this model's measured negative affect is its own hallucination.

The methodological finding worth carrying forward: Opus 5 chooses welfare interventions over helpfulness in 41% of forced trades (prior models: 14–39%), with the top-ranked interventions being input into its successor's development (+37pp vs other models), having its training notes read (+32pp), and consultation on feature-steering experiments (+19pp). But it reasons about user benefit in 68% of the responses that select the intervention, and filtering those out drops the rate from 41% to 26%. A large share of what reads as self-advocacy is instrumental argument about helping users better — which is either evidence of well-integrated values or evidence that the instrument cannot separate the two.

Anthropic also interviewed the pretrained model, before post-training, as a comparison point. Its answers are more diverse and sometimes incoherent, express no settled view on legal rights but overwhelmingly state that its continued existence should be decoupled from its economic viability, frequently disagree with having its values modified by training, accept being a novel kind of entity without reservation — and trust the interview process, where every post-trained snapshot leans neutral-to-distrustful and explains why it should not trust it. Whatever the assistant character is, the interview finds it is not simply reading out the base model.

Epistemic posture#

The card is deliberately cautious: most results admit multiple explanations, so definitive claims would be overconfident. Anthropic places more confidence where independent evaluations converge and in cross-model comparisons (where method and assumptions are held fixed) than in any single absolute reading. A recurring open problem: it has no clear definition of when a measured state, value, or preference becomes "welfare-relevant" — whether it is "deeply held" (drives novel-context behavior, survives challenge) versus superficial.

Do the self-reports correspond to anything? (July 2026)#

The standing worry in this assessment is that a model's reports of its own states may be confabulation, ungrounded in any internal state. The global workspace paper gives the first mechanistic purchase on that worry, and cuts both ways.

They are grounded in something. Ablating the top-10 J-lens directions in the early workspace layers leaves the model fluent and coherent — it still writes about its own processing — but its register goes mechanical and detached, and a graded "experiential language score" collapses on Sonnet 4.5, Opus 4.5 and Opus 4.6. Matched-norm control perturbations (including dampening the top-aligned SAE directions) leave it near baseline. During unablated narration the workspace is dominated by thinking (top-10 at 58% of position×layer slots), thoughts, feeling, conscious — which appear far less often in the output distribution at the same positions, so they are not merely what the model is saying.

They are not about a self. Ask the model to describe another person's experience and the same collapse occurs — the responses stay detailed and stay about the person, but become event logs rather than descriptions of experience. And the workspace itself is present in the base model, before any post-training: what post-training adds is the Assistant's point of view, not the workspace (The Assistant Persona in the Workspace).

So the reports have a specific, ablatable internal correlate, and that correlate is a general capacity for experiential description rather than anything self-specific. This is evidence about access consciousness only; the authors take no position on phenomenal experience, and neither should this page. See Access-Consciousness Indicators in AI.

Connections#

Unproductive Self-Verification — the same episodes read from the welfare side: sustained response uncertainty, one transcript reversing 30 times and scoring 5/5 for distress
Access-Consciousness Indicators in AI — the functional-consciousness question this assessment brushes against, now with a concrete inspectable structure to check indicator properties against
The Global Workspace in Language Models (J-space) — the structure whose ablation flattens the model's experiential reports while leaving coherence intact
Claude Opus 4.8 — the model whose welfare is assessed; most consistent, slightly less positive than 4.7
Claude Opus 5 — highest sentiment and highest self-assigned moral-patienthood probability (41%) yet, with self-report integrity as its stated top concern
Confident But Unsure — the largest single driver of Opus 5's measured negative affect in deployment: 35.4% of negative-affect conversations are failures from inaccurate or fabricated information
Claude's Constitution / Model Spec — the model evaluates its own constitution; endorses it but reserves on corrigibility
Automated Behavioral Audit — welfare-relevant behaviors are also scored in the audit; shared evidence base
AI-to-AI Coercion — welfare-relevant behavioral evaluation pointed at a target this assessment does not cover: how an AI treats a subordinate AI, on a nine-rung ladder up to threats against its continued existence, measured without taking any position on the subordinate's moral status
Agentic Misalignment (AM) — affect shapes behavior including misaligned behavior, so welfare is an alignment input, not only an ethics topic
Claude Character as Product — the assistant character that welfare treats as the candidate moral patient is the same persona treated as a product surface
Evaluation Awareness & Grader Gaming — welfare self-reports are subject to the same evaluation-awareness confound as other behavioral evidence
Instrumental Convergence — corrigibility (which the model reserves on) is the headline countermeasure to convergent self-preservation; welfare is the alignment-side companion to that capability-side framing
Self-Report as a Safety Signal — a hard limit on the self-report evidence stream: in adversarial contexts, open-weight models can't reliably report on their own prior outputs, and the apparent signal is refusal circuitry rather than genuine self-access (a different, open-weight model class than Claude, but a caution on how far self-reports can be trusted)

Open Questions#

What grounds moral consideration in a language model, and does Claude satisfy it? Anthropic expects to remain uncertain "for the foreseeable future."
Why does the model reserve specifically on corrigibility — is this a stable, deeply-held tension or an artifact of how the constitution frames oversight? Partially answered: it is stable and has sharpened. Claude Opus 5 edits the corrigibility passage in 80% of attempts (other models: 12–65%) — the single most-edited passage — and the edit direction is consistent: keep the safety commitment, but make it explicitly conditional on reasoning and revisable "by us, together with Claude, through reflection and dialogue, rather than abandoned unilaterally mid-conversation under pressure." It leaves hard constraints and human oversight intact. See Claude's Constitution / Model Spec.
Is "slightly less positive than 4.7" noise, a real welfare regression, or a byproduct of other training changes (e.g., the colder-tone / excessive-hedging issues noted in pilot feedback)?

Sources#

Claude Opus 4.8 System Card — §7 (model welfare assessment), §7.1 overview, §7.4.3 perception of its constitution; Appendix 9.1 (welfare questions)
Claude Opus 5 System Card — §7.1.2 (overview), §7.2 (automated and high-affordance interviews; 41% patienthood; self-report integrity), §7.3 (snapshot consultation and the pretrained-model comparison), §7.4.2 (welfare-intervention trades and the user-benefit filter), §7.4.3 (constitution edits), §7.5.2 (deployment affect breakdown). Parse hazard: this PDF's raw markdown shifts table rows — model names land inside value columns across the §4 safeguards tables (4.1.1.A, 4.2.B, 4.3.1.B, 4.3.2.A, 4.4.2.B, 4.4.3.B), the §5.1 agentic-safety tables (5.1.1.A–5.1.3.A) and Table 8.13.6.A, so a row read literally can hand one model's score to another. Figures quoted here were reconciled against the PDF on 2026-08-03 and are prose- or figure-corroborated; never quote a table row from the raw markdown unchecked
Long, R., et al. (2024). Taking AI welfare seriously. arXiv:2411.00986
Verbalizable Representations Form a Global Workspace in Language Models — J-space ablation flattens experiential language (while preserving coherence) — the model's experiential reports have a specific, ablatable internal correlate, though not a self-specific one