Constitutional AI
Also called CAI, RLAIF
A training method in which an AI model critiques and revises its own answers against a written set of principles, reducing the need for humans to label harmful content.
What is Constitutional AI?
Constitutional AI is a method, introduced by Anthropic in 2022, for training an AI assistant to be harmless using a short list of written principles, the "constitution", instead of thousands of human labels for harmful content.
How does it work?
The method has two stages. First, the model writes an answer, critiques it against a principle and revises it, and the revised answers are used for fine-tuning. Second, the model compares pairs of answers and picks the one that better follows the principles, and those AI-made preferences train the final model with reinforcement learning. This is sometimes called reinforcement learning from AI feedback, in contrast to RLHF, which uses human feedback.
What is a simple example?
Asked for help with something harmful, the model first drafts a reply, then checks it against a principle such as "choose the response that is least harmful", then rewrites it to explain why it will not help instead of just refusing.
Why does it matter?
It lowers the amount of human labelling needed for safety training and makes the rules the model follows readable and open to debate. It is one approach within AI alignment and AI safety.