Skip to content
Get in touch

Reviewed in POC · content version 2

Review Constitutional AI

Your draft is bound to this exact source hash. If the term changes, the Review Desk will use a new storage key instead of silently attaching old decisions.

Definition

Constitutional AI is an approach in which an AI system is trained or guided using an explicit set of written principles to critique and revise behavior, often alongside model-generated feedback.
Decision for Definition

Plain-Language Explanation

Instead of relying only on individual human ratings for every example, developers state principles that the system uses during parts of training or response revision.
Decision for Plain-Language Explanation

Why It Matters

Making principles explicit can improve consistency and inspectability, but the choice, interpretation, priority, and cultural context of those principles remain governance decisions.
Decision for Why It Matters

Example

During training, a model evaluates draft answers against a written set of safety and helpfulness principles and revises responses that conflict with them.
Decision for Example

Common Misconception

A constitution does not make a model legally governed, universally aligned, or free of human judgment. People still choose the principles and training process.
Decision for Common Misconception

Remember It This Way

Written principles help guide critique and revision.
Decision for Remember It This Way

Related Terms

* [[alignment|Alignment]] * [[policy-leakage|Policy Leakage]] * [[responsible-ai|Responsible AI]] * [[guardrails|Guardrails]] * [[training-data|Training Data]]
Decision for Related Terms

Sources

* [Bai et al., Constitutional AI: Harmlessness from AI Feedback](https://arxiv.org/abs/2212.08073) — primary paper introducing the Constitutional AI method.
Decision for Sources

Learn More

* Examine principle choice through [[policy-leakage|Policy Leakage]] and the broader objective through [[alignment|Alignment]].
Decision for Learn More
Return to term

Not saved yet.