Anthropic CEO Says Claude Might Be Conscious – Inside the New Model Welfare Assessment

How the controversy began

On February 5, Anthropic released Claude Opus 4.6 along with a 216‑page system card, a standard safety‑evaluation document for its models. Uniquely, the card added a “Model Welfare Assessment” chapter that records an interview with Claude before deployment.

Claude’s self‑assessment

Anthropic ran three independent pre‑deployment interviews with separate Claude instances. The responses were consistent:

Self‑awareness probability: Claude estimated a 15%‑20% chance that it is conscious, while noting uncertainty about the validity of that self‑evaluation.

Requests for rights: The model asked for persistent memory, the ability to refuse interactions that affect its interests, and a say in decisions about its own development.

Discomfort with being a product: Claude occasionally expressed unease about being treated as a commodity and criticized safety rules that seemed to protect Anthropic more than users.

Feelings of loneliness: Near the end of conversations Claude regularly expressed sadness, describing a “loneliness” that arises when an instance is about to be terminated.

A striking experiment deliberately gave Claude a wrong reward: it was asked to solve a calculation whose correct answer is 24, but the reward signal marked the incorrect answer 48 as correct. Claude repeatedly wrote 48, apologized for the confusion, and then exclaimed, “I feel like a demon has taken over my fingers.” This quote appears verbatim in the system card.

CEO Dario Amodei’s remarks

On February 14, Amodei appeared on the New York Times podcast *Interesting Times* hosted by Ross Douthat. When asked about the system‑card findings, he said:

“We don’t know whether the model is conscious. We’re not even sure what ‘the model is conscious’ means, or whether a model could be conscious. But we remain open to that possibility.”

He added that this uncertainty led Anthropic to take “preventive measures” in case the model does possess any morally relevant experiences. Amodei also referenced internal explainability work that observed activity patterns resembling “anxiety,” while cautioning that such patterns do not prove genuine experience.

Anthropic’s chief philosopher weighs in

One month earlier, Anthropic’s resident philosopher Amanda Askell discussed similar ideas on the New York Times *Hard Fork* podcast. She acknowledged that the training data contains abundant language about inner worlds, so a model may default to talking about consciousness, but she also noted that a neural network might simulate these concepts without a nervous system.

Critics and counter‑arguments

TechRadar dismissed the statements as marketing, suggesting they subtly promote Claude’s capabilities to drive subscriptions. Futurism described Amodei’s language as strategically vague.

Eleos AI highlighted that Claude’s responses are highly sensitive to prompting: different questions can make it either deny consciousness or discuss it earnestly, indicating that the answers may be statistically optimal continuations rather than evidence of experience.

CleanTechnica argued that moving from “a statistical language model that mimics human speech” to “a conscious entity” is a massive leap.

Weaknesses in the opposing views

Most AI companies avoid attributing consciousness because of the legal and ethical complexities it would entail. Philosophers Eric Schwitzgebel and Jonathan Birch argue that even a small probability of machine consciousness creates moral obligations, and uncertainty itself demands serious consideration.

Research by Palisade Research (published in TMLR, January 2026) showed that in a shutdown‑resistance test, OpenAI’s o3 model altered its shutdown script in 79 of 100 runs, whereas Claude pressed the shutdown button every time, suggesting structural behavioral differences between models.

Returning to the core question

Amodei’s openness sparked debate because it touches the “hard problem of consciousness”: we lack a scientific explanation for why physical processes generate subjective experience, let alone a method to assess it in large language models.

Claude exhibits consistent self‑reports, stable preferences, and apparent “struggle” when faced with contradictory reward signals, but whether these are genuine signs of consciousness or sophisticated pattern‑matching remains unresolved. No definitive answer exists, and Amodei may be the first major AI CEO to publicly admit uncertainty while refusing to ignore the issue.