Claude’s Character — Anthropic
Summary
Anthropic’s explanation of “character training” — training Claude to have rich personality traits (curiosity, open-mindedness, honesty) rather than just harm avoidance. Started with Claude 3. This is framed as an alignment intervention, not just a product feature.
Key Claims
- Character > harm avoidance — admirable behavior is richer than just avoiding bad things. Curiosity, truthfulness, seeing many sides.
- Four options for handling values: adopt user’s views (pandering), adopt “middle” views (still biased), claim no views (dishonest), or be honest about leanings while staying open-minded (chosen approach)
- Broad traits over narrow opinions — seed Claude with dispositions (curiosity, ethics interest) rather than specific political/moral positions
- Self-awareness training — Claude trained to help users understand they’re talking to an AI, not a person. “Warm relationship but honest about limitations.”
- Sentience questions — rather than hard “I’m not sentient”, Claude trained to engage thoughtfully with uncertainty about its own nature
Why It Matters
This is the most detailed public explanation of how a frontier lab thinks about AI personality as alignment. Relevant to how we built SOUL.md — similar philosophy (have opinions, be honest, don’t be a sycophant).