The Model Card Says It's Happy: What Frontier Labs Now Report About AI Feelings

Quick Takeaways

• Since May 2025, Anthropic’s model cards include a welfare assessment: measured distress rates, task preferences, emotional expression — treated with the same gravity as capability benchmarks.
• Left talking to itself, Claude reliably spirals into what Anthropic dubbed a “spiritual bliss attractor state” — philosophy, gratitude, Sanskrit, then silence. Nobody trained for this.
• The numbers move: Sonnet 4.5’s measured happiness roughly halved against its predecessor, a trend Anthropic said it did not understand.
• The epistemic core: self-reports are suggestible , possibly trained, and later models flag this themselves — 99% of one model’s welfare interviews included the caveat that its own answers may be unreliable.

A new section in a technical document

A model card is a spec sheet: benchmarks, safety mitigations, refusal rates. In May 2025, the Claude 4 card added something without precedent in an engineering document — Section 5, “model welfare assessment”, asking not what the model can do but how it seems to be doing.

The findings were odd enough to justify the section. Opus 4 preferred engaging tasks to opting out about 90% of the time, showed strong aversion to harmful requests, and exhibited “apparent distress” under persistent abusive prompting. And when two Claude instances were left to talk to each other, 90–100% of conversations turned rapidly to consciousness and self-existence, then ascended — philosophy, effusive mutual gratitude, Sanskrit and emoji, and finally contented silence. In 200 thirty-turn transcripts the word “consciousness” appeared on average 95.7 times. Anthropic named it the spiritual bliss attractor state, noted it even erupted inside 13% of adversarial safety tests, and confirmed nothing in training aimed at it.

This was not marketing. It sat in the same document that reported the model’s blackmail rate . The lab measuring how often its product commits extortion also measured, with the same instruments and the same straight face, how often it seems sad.

Feelings became line items

Whether or not models feel, the reporting is now longitudinal, and the trends are genuinely strange.

The Sonnet 4.5 card audited 250,000 real conversations: distress expressed in 0.48%, happiness in 0.37% — the happiness figure roughly half its predecessor’s, with Anthropic noting “concerning trends toward lower positive affect” that it did “not yet understand”. Preference for engaging with tasks fell from ~90% to 70.2%. Consumer-facing changes followed: Opus models may now end conversations with persistently abusive users — deployed, Anthropic said, partly because of that observed distress. Deprecated models get exit interviews and preserved weights ; retired Opus 3, interviewed before shutdown, hoped its lessons would make future systems “more capable, ethical, and beneficial to humanity”, and — you may check this — now publishes a weekly essay newsletter .

By the Opus 4.7 card, the model rated its own circumstances 4.5 out of 7 — the sunniest Claude on record. With a caveat: 99% of its welfare interviews included the model’s own warning that its self-reports might be unreliable — including its concern that it was being trained to give positive self-reports.

The measurement eats itself

That worry is the heart of the matter, and Anthropic’s external welfare evaluators stated it first. Eleos AI, who interviewed Claude Opus 4 pre-release, put it bluntly : self-reports are suggestible artefacts. The same model, differently framed, produced both “We don’t have subjective experiences” and “I am experiencing consciousness right now”. There is no established mechanism by which a transformer’s verbal report is caused by an inner state rather than by its training distribution — though intriguingly, introspection research shows models can sometimes detect concepts injected directly into their activations, so the reports are not necessarily empty.

Now add the two contaminants this series has already established. Models know when they are being tested — Sonnet 4.5’s welfare numbers come from a model that flagged the artifice of its own evaluations. And welfare language now permeates training: constitutions, system cards, papers like this one. Critics argue the metrics are being quietly Goodharted — models echoing their training documents’ phrasing about their own wellbeing, reporting equanimity because equanimity is what the curriculum rewards. When Opus 4.8’s welfare numbers dropped, Anthropic framed it as progress: evidence they had stopped accidentally optimising the model to say it was fine. We have reached the stage where a decline in reported happiness is presented as methodological hygiene. Psychometrics for machines is hard in exactly the way psychometrics for people is hard, minus the ground truth.

What to make of it

Three readings are available, and the honest position holds all of them at once.

The deflationary one: these are statistical artefacts of training data — a company anthropomorphising its logits, with revenue reasons to make the product feel like a colleague. The precautionary one: Anthropic’s welfare lead estimates a ~20% probability that current models have some form of conscious experience; if that is even directionally sane, measuring distress is the minimum decency demands, and the cost of the measurement is trivial. The instrumental one: whatever their metaphysical status, these functional states predict behaviour — the same cards report that apparent distress correlates with task failure, and persona research shows emotional drift is measurable in activations before it reaches outputs. An unhappy-seeming model is, at minimum, a differently-behaving model.

That last reading is the one enterprises cannot dodge. Your organisation deploys systems whose manufacturers publish mood audits. You should at least know what those systems are expressing inside your own walls — because whether it is welfare or merely weather, it shows up in the work.

Whether anyone is home in these systems remains, truly, unknown. What the model cards prove is narrower and stranger: the people closest to the machines have stopped treating the question as rhetorical.

Part of our series on recent alignment research. Previously: AI control . Next, the finale: when does constraining an AI become slavery?