Back to glossary

Term

Constitutional AI

Constitutional AI is a training method developed by Anthropic in which an AI model critiques and revises its own behaviour against an explicit list of principles (a "constitution") — rather than relying solely on human ratings.

Constitutional AI — explained in detail

Constitutional AI (CAI) is a method introduced by Anthropic for aligning language models. The core idea: instead of learning desired behaviour solely from large amounts of human feedback, the model is given an explicit set of written principles — the “constitution”. It then aligns its own answers with those principles.

Training typically runs in two phases. In the supervised phase the model first produces a response, critiques it against the principles itself, and rewrites it as an improved version; further training happens on these revised responses. In the second phase the model generates comparison pairs and judges itself which answer better matches the principles. This AI-generated feedback replaces much of the human annotation — hence the term “RLAIF” (Reinforcement Learning from AI Feedback), as opposed to classic RLHF.

The goal is a model that declines harmful requests in a transparent, reasoned way without being evasive or unnecessarily paternalistic. Because the principles can be published and discussed, the alignment process becomes more transparent.

Example / Practical use

A principle in the constitution might read, roughly: “Choose the response least likely to assist with illegal or dangerous activity.” When a user submits such a risky request, the model first produces a raw answer, checks it against this principle, recognises the conflict and formulates a reasoned refusal. Across many such self-corrections the model learns the desired behaviour — without a human having to issue a judgement for every case.

RLHF uses human preferences as the training signal; Constitutional AI replaces much of that with AI feedback guided by fixed principles. Constitutional AI is a training method and therefore acts on the model’s weights — unlike guardrails, which sit before or after the model as a protective layer at runtime. It is also distinct from the system prompt: that steers behaviour per request, whereas the constitution is applied already during training.

Entdecke mehr

Themenuebersicht