Constitutional AI (CAI) is Anthropic's technique for training AI models to be helpful, harmless, and honest — using a set of explicit principles rather than purely human feedback.

How It Works

1. Define a "constitution" — a set of ethical principles
2. Use AI feedback to evaluate responses against the constitution
3. Fine-tune the model to prefer constitutional responses
4. Reduces reliance on human labelers for safety

Why It Matters

CAI makes the values baked into an AI model explicit and auditable — rather than implicit and opaque. It's a significant advance in making AI safety transparent.

Reference:

Constitutional AI paper

https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback