Constitutional AI

Constitutional AI is how Anthropic trains Claude’s behaviour: instead of relying only on human labels for what is harmful, the model is trained against an explicit written constitution — a set of principles it critiques and revises its own outputs against.

Why it matters practically, not just philosophically

Two consequences you can actually observe:

The principles are published. You can read what Claude is trained to weigh. That is unusual, and it makes model behaviour arguable from a document rather than from guesswork.

Behaviour is a stated design target, not an emergent accident. When Claude declines something, there is a written basis for it. That does not make every decision correct, but it makes it inspectable.

What it is not

• Not the Usage Policy. The constitution shapes how the model behaves. The usage policy governs what you are permitted to do with it. Different documents, different enforcement.
• Not a guarantee. System cards exist precisely because trained behaviour has to be measured, not assumed.
• Not static. The constitution has been revised and republished.

What this page does not claim

This entry does not reproduce the constitution’s contents or summarise its principles — it is a substantive document that deserves reading in full rather than in paraphrase.