Constitutional AI (CAI) — AI Glossary

Reference details

Related terms

• AI Alignment
• RLHF (Reinforcement Learning from Human Feedback)
• Red Teaming

How it works

1. A set of written principles is defined, and those principles — rather than case-by-case human labels — are what the training optimises against.
2. The model generates a response, then critiques it against a principle, then revises it, producing a corrected pair from its own output.
3. Those revisions become supervised training data, which is how harmlessness is taught without a human writing each example.
4. A preference stage then uses model-generated comparisons judged against the principles, replacing the human preference labels used for harmlessness.
5. Because the principles are written down, the target is auditable and changeable in a way that an aggregate of human labels is not.

Example: Why the written principle is the auditable artefact

Human preference labelling for harmlessness produces a model whose boundary is implicit in thousands of individual judgements — to ask why it refused a particular class of request, you can only inspect behaviour. A written constitution makes the target explicit: the principle can be read, disagreed with, and revised, and a behaviour change can be traced to a principle change. The capability is not that it refuses better; it is that the basis for refusal is a document rather than a dataset.

Commonly confused with

Read as a guardrail system, which it is not. Guardrails are runtime filters that inspect inputs and outputs and block them; this is a training method that shapes the model's own behaviour before deployment, so there is nothing at inference to bypass or to configure. It also differs from RLHF in where the supervision comes from rather than in mechanism: both end in preference optimisation, and this substitutes model-generated critiques judged against written principles for human comparison labels on the harmlessness axis.

When to use it

Reach for it when:

• Explaining how harmlessness training can scale beyond human labelling capacity.
• Discussing alignment approaches whose target is auditable rather than implicit.
• Distinguishing training-time behaviour shaping from runtime filtering.

Reach for something else when:

• Describing runtime input or output filtering, which is guardrails.
• As a guarantee of safety — it shapes a tendency and does not enforce a rule.
• As an alternative to RLHF wholesale, since it replaces one axis of its supervision.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Constitutional AI (CAI) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path