AI Glossary Term

What does “Constitutional AI” mean?

Definition

A training and behavior-shaping approach where model outputs are guided by a fixed set of written principles.

Browse the full AI Glossary Explore all free AI guides

How Constitutional AI works

Example

A model is given a short list of written principles, such as avoiding harmful advice. It critiques and rewrites its own draft replies according to them, and those improved replies are used as training data.

Why it matters

Writing down principles makes the intended behavior inspectable. It reduces the amount of human labeling needed, though the principles themselves involve choices.

Common questions

Who developed it?

Anthropic described the approach in a 2022 paper.

Is it replacing human feedback?

It supplements it. Parts of the process use AI-generated feedback guided by the written principles.

Who chooses the principles?

The developer, which is why publishing them helps outside scrutiny.

Evaluation SetA held-out dataset used to measure model quality after training.GeneralizationHow well a model performs on new, unseen data outside the training set.Human-in-the-LoopA workflow where humans review, guide, or override AI outputs.HyperparameterA configuration value set before training, such as learning rate, batch size, or depth.InferenceThe runtime phase where a trained model generates predictions or outputs.Knowledge DistillationTraining a smaller model to imitate the outputs of a larger model.

Learn more in our free guides

Constitutional AI

Test yourself

Constitutional AI Quiz