Skip to content
Safety

Constitutional AI

A training approach where AI models are given a set of principles (a 'constitution') and learn to self-critique and revise their outputs to comply with those principles. Reduces reliance on human feedback for safety alignment.

Why it matters

Teaching AI to behave well usually means paying people to rate thousands of responses, which is slow and hard to scale. Constitutional AI gives the model a written set of principles and lets it critique and fix its own answers against them. This makes safety training more consistent and transparent, since the guiding rules are actually written down rather than buried in scattered human judgments.

A concrete example

A model drafts a reply that's technically correct but rude. Under Constitutional AI, it checks its draft against a principle like "be helpful and respectful," notices the tone problem, and rewrites the answer to be polite while keeping the useful content. This self-review happens during training, so by the time you use the model, it has already learned to lean toward responses that follow those principles.

How to use it

As a user, this explains why a model refuses some requests and how the refusal is shaped. As a builder, the transferable idea is worth borrowing: a written set of explicit principles that a system checks its own output against catches more than a list of banned phrases, because it generalises to cases you did not enumerate. The method is described in Constitutional AI: Harmlessness from AI Feedback.

The common mistake

Assuming a principle-trained model has consistent judgement. Principles are applied through the same probabilistic process as everything else, so refusals and permissions vary with phrasing rather than following a rule you could state.

Related terms

Put Constitutional AI into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.