Alignment · Established · Advanced

Constitutional AI

An alignment approach where a model critiques and revises its own outputs against a written set of principles, reducing reliance on human harm labelling.

What Constitutional AI is

Instead of humans labelling every harmful response, an explicit constitution states the principles, and the model applies them to generate its own preference data.

How it works

The model produces a response, critiques it against the principles, revises it, and the revision pairs become training data for supervised and preference tuning.

Why it matters

It makes the value specification explicit and inspectable, which is easier to govern than values implicit in a crowd of annotators.

Common uses

→Harm reduction training
→Transparent behaviour policy
→Scaling alignment data

Strengths

✓Written, auditable principles
✓Less exposure of humans to harmful content

Watch for

✓Principles still reflect their authors
✓Model must interpret them correctly

Continue exploring

Reinforcement Learning from Human Feedback AI Alignment AI Ethics