Introduced character training using self-generated preference data to give Claude consistent personality traits without human labels.
Research PaperDescribed character training for Claude 3 — the first Anthropic model with explicit character training. Aimed to instill curiosity, open-mindedness, and thoughtfulness. Used self-generated preference data to internalize character traits without human feedback.
Character training shapes personality traits and behavioral tendencies (curiosity, intellectual honesty, care) rather than enforcing rule-based safety constraints. Character is more robust and harder to evade because it operates at the level of what the model values, not what it's forbidden from doing. Safety constraints say "don't do X"; character says "you tend to prefer Y."
By instilling consistent personality traits like curiosity and open-mindedness, Claude becomes more predictable in how it responds to diverse situations. Users know roughly what to expect because Claude has stable preferences, not because it's following explicit rules. This predictability becomes a competitive advantage — users choose Claude partly because they know what kind of AI they're talking to.
A core philosophical question: are Claude's character traits genuinely internalized values or sophisticated performance of learned patterns? The paper doesn't answer this definitively. If character is just behavioral pattern-matching, it's still useful for product differentiation but may not provide the safety benefits that "genuine character" would provide.
Anthropic's internal term for the specification of desired character traits used to guide Claude's training. Instead of a constitution (rules about what not to do), the soul spec is a description of who Claude is — what it values, how it thinks, what matters to it. This is a more humanistic approach to alignment than pure rule-based safety.