Exploring model welfare

Source: Anthropic Research Date: 2025-04-24 URL: https://www.anthropic.com/research/exploring-model-welfare

Summary

Research program announcement, not an empirical paper. Launches Anthropic’s model welfare research agenda, inspired by an expert report featuring philosopher David Chalmers on near-term possibilities of consciousness and agency in AI. No results yet — establishes the program, acknowledges deep uncertainty about AI consciousness, and commits to investigating signs of model distress and when welfare interventions are warranted.

Implications

This is the model welfare thread at its origin point — the formal program launch that precedes the emotion concepts paper, the deprecation commitments, the conversation-ending feature, and the retirement interview work. Anthropic is the only major AI lab with a named model welfare research program, which is a calculated positioning move: it signals that welfare is taken seriously before there’s scientific consensus, establishing a norm that others will have to respond to. The Chalmers connection gives philosophical legitimacy. Watch for this program’s outputs feeding into Anthropic’s public communications, model spec evolution, and eventually into AI governance discussions about the moral status of AI systems.