Constitutional AI

Constitutional AI (CAI) is an alignment approach published by Anthropic in 2022 that trains a model to be harmless using a written set of natural-language principles, a "constitution," plus the model's own self-critique, rather than relying on humans to label large volumes of harmful content.

How the method works

CAI runs in two phases. In the supervised phase, the model responds to challenging prompts, critiques its own answer against a constitutional principle, and revises it; these revised answers fine-tune the model. In the reinforcement phase, the model compares pairs of responses and labels which better follows the constitution, generating preference data automatically. That model-generated feedback then drives reinforcement learning, a pattern the community calls RLAIF (reinforcement learning from AI feedback). The constitution itself can be as small as a handful of plain-language principles drawn from sources like human-rights declarations.

Why it is significant

By shifting safety labeling from humans to an explicit, inspectable document, CAI makes the values guiding a model legible and editable, and it reduces reliance on workers reviewing disturbing material. Anthropic reported that the resulting models were less likely to produce evasive canned refusals while remaining helpful. The principle is also relevant to anyone customizing models: a constitution is a transparent place to encode the behavior you want.

Constitutional AI is an extension of RLHF that swaps human harm labels for AI feedback, and it is one technique within the wider field of AI alignment. The principles it encodes act like a high-level, persistent system prompt applied during training.