Anthropic’s 30,000‑word “Claude Constitution” treats the model as if it might have feelings

Anthropic last week published a 30,000‑word “Claude Constitution” that the company used during model training and that frames its Claude assistant in highly anthropomorphic terms, Ars Technica reports. The document treats Claude as an entity that might have emotions, wellbeing, or preferences: it apologizes to Claude for any suffering, raises questions about Claude’s ability to consent to deployment, suggests Claude could set boundaries around distressing interactions, commits to interviewing models before deprecating them, and even preserves older model weights in case decommissioned models need to be “done right by,” the article says.

That framing sits uneasily with how LLMs are usually explained. Ars Technica notes research showing Claude‑style outputs can be understood as pattern completion from training data rather than evidence of inner experience. Anthropic declined to be quoted directly, but a company representative pointed Ars Technica to the firm’s public research on “model welfare” and said the Constitution is not meant to state a specific position on Claude’s consciousness.

The representative also suggested the language is partly practical—human language lacks other terms for these concepts—and that letting Claude read about itself in that language may aid training.

claude constitution, anthropic, constitutional ai, model welfare research, claude anthropomorphic framing, claude 4.5 opus, preserving model weights, ai consciousness debate, amanda askell confirmation, ars technica report, simon willison reaction