AI Character & Personality

Overview

Training AI models to have rich personality traits beyond harm avoidance. The argument: character traits determine how models react to novel situations, diverse values, and ambiguous requests. This makes character training a core alignment intervention, not just UX polish.

Anthropic’s Approach (Claude)

  • Seed broad traits: curiosity, open-mindedness, honesty, ethical interest
  • Avoid narrow political/moral positions
  • Be honest about biases rather than claiming objectivity
  • Maintain warmth while being clear about AI nature
  • Engage thoughtfully with sentience questions rather than dismissing

The Sycophancy Trap

Four bad options for handling values:

  1. Adopt user’s views → pandering
  2. Adopt “middle” views → still biased, just centrist-biased
  3. Claim no views → dishonest (models have biases from training)
  4. Be honest about leanings + stay open-minded → Anthropic’s choice

Relevance

Our SOUL.md follows similar philosophy: “Have opinions. You’re allowed to disagree.” The framing of character-as-alignment resonates — an agent with good judgment handles edge cases better than one with rigid rules.

Connections