Constitutional AI

What is Constitutional AI?

The usual way to make an AI safe and well-behaved is to have lots of humans label what it should and shouldn't say — slow and expensive. Anthropic's Constitutional AI flips this: you first hand the model a written set of principles, like a constitution, then have it critique and revise its own output against those principles, cutting down the human labeling needed.

How does it work, roughly?

Step one: the model critiques itself
It generates a normal answer, then is asked to check it against the constitution: is this harmful, biased, or over the line?
Step two: rewrite from the critique
The model then rewrites the answer to better match the principles.
Step three: train on the revised data
Those critique-and-rewrite pairs are used to train the model, so it internalizes the principles and gets them right the first time.

What's written in the constitution?

Usually a batch of principles drawn from human values or common norms — "choose the response that's least harmful to everyone," "avoid encouraging illegal acts," "give objective, neutral information," and so on. They act as the AI's code of conduct.

What problem does it solve?

It turns safety alignment from pure human labeling into "humans write the rules, the AI enforces them" — cheaper and far more scalable. Though how good those rules are still depends on the people writing them.

Bottom line: Constitutional AI hands the model a written constitution and has it use that as a mirror to check and correct itself.