Anthropic Guide Sets 14 Criteria for AI Wellbeing Evaluations

Anthropic’s wellbeing-evaluation guide sets out 14 design qualities and sample-size targets, giving researchers a framework for assessing how AI conversations affect users over multiple turns.

The guidance suggests at least 100 distinct items for complex evaluations involving multiple turns and at least 1,000 for simpler tests covering a single turn. Its recommendations include clear measurement criteria, realistic conversation scenarios, expert input and checks that graders are valid.

The guide says risks to wellbeing may emerge as a conversation develops, so evaluating a single response may miss relevant context. Anthropic announced a $5 million grant program Aug. 25, 2026, to support independent research and open-source evaluations of AI’s effects on wellbeing.