Constitutional AI: Definition and Examples

Constitutional AI (CAI) is an alignment approach for language models introduced by Anthropic in 2022. Its fundamental principle is to equip an AI model with a set of explicit rules—called a 'constitution'—that guide its behavior. These principles cover values such as honesty, helpfulness, harmlessness, and respect for fundamental rights. Concretely, the process takes place in two phases. In the first phase (critique and revision), the model generates responses and then self-evaluates by referring to the constitutional principles. It identifies potential violations and produces a revised version of its response. This critique-revision cycle can be repeated several times to refine quality. In the second phase, the response pairs (original vs. revised) are used to train a reward model via RLAIF (Reinforcement Learning from AI Feedback). This reward model partially replaces direct human feedback, making the process more scalable while maintaining a high level of alignment. The major advantage of Constitutional AI is transparency: the rules are explicit and auditable, unlike the implicit preferences captured by classical RLHF. It also allows for public debate on the values encoded in the system and modification without fully retraining the model.