Constitutional AI

Constitutional AI (CAI) is an AI alignment technique from Anthropic that uses a set of principles (a "constitution") to guide AI self-improvement — the AI critiques and revises its own outputs according to the constitution, then trains on the revised outputs, reducing the need for human feedback.

CAI Pipeline

• Constitution: A set of principles (e.g., "be helpful, harmless, and honest") written in natural language.
• Critique: The AI generates a response, then critiques it against each principle.
• Revision: The AI revises its response based on the critique — producing a constitutionally aligned output.
• RLAIF Training: Train a preference model on (original, revised) pairs — the revised version is preferred.

Why It Matters

• Scalable Alignment: Reduces dependence on expensive human feedback — the constitution encodes values.
• Transparent: The constitution is an explicit, readable specification of AI behavior standards.
• Harmlessness: CAI is particularly effective at reducing harmful outputs — the constitution explicitly forbids harm.

CAI is teaching AI values through principles — using a written constitution to guide AI self-critique and revision for scalable alignment.