Examples
Worked examples
- Is an instance
A model trained with a 75-item constitution covering helpfulness, harmlessness, and honesty.
- Is an instance
A research replication of Constitutional AI on a smaller open-weights base.
Counter-examples
Looks similar, but isn't
- Not an instance
Pure RLHF with human-only preference data.
- Not an instance
Rule-based output filtering after generation (a different mitigation layer).
Editorial commentary
Constitutional AI (Bai et al., Anthropic, 2022) is a training method for aligning a language model’s behaviour to a stated set of principles — a “constitution,” typically a list of several dozen natural-language rules — using the model’s own outputs and self-critique rather than relying solely on large volumes of continuously collected human labels. It runs in two phases: a supervised phase, where the model critiques and revises its own responses against the constitution and is fine-tuned on the revised versions; and a reinforcement-learning phase that uses AI-generated preference judgments, rather than human preference judgments, as the training signal — a variant sometimes labelled RLAIF (RL from AI Feedback) to distinguish it from standard RLHF.
Constitutional AI is best understood as a complement to RLHF, not a full replacement for it: it reduces how much of the alignment signal needs to come from continuous large-scale human labelling of harmful/harmless comparisons, but it does not eliminate human judgment from the process — humans still author the constitution’s principles and evaluate the resulting model’s behaviour. This entry covers the methodological concept published in the original paper, not any single product’s specific, evolving implementation of it (Anthropic has published and revised its own constitution used for Claude models separately from the original research paper).
What it is not
Constitutional AI is not the same thing as a model’s system prompt or usage policy — a constitution is a training-time input that shapes the model’s learned behaviour before deployment, whereas a system prompt is a run-time instruction applied to an already-trained model. Nor is it a documentation artefact like a model card; it is a training technique that a model card or system card would disclose as part of describing how the model was aligned.
References
- Bai et al., ‘Constitutional AI: Harmlessness from AI Feedback’ (arXiv, 2022)
Also known as
CAI · RLAIF (related concept)
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Constitutional AI (concept)"
vocab-term-identifier="https://casrai.org/dictionary/term/constitutional-ai-concept" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/constitutional-ai-concept",
"name": "Constitutional AI (concept)",
"identifier": "https://casrai.org/dictionary/term/constitutional-ai-concept",
"description": "A training methodology in which a model is trained to align its outputs with a written set of principles ('a constitution'), with the model itself used to critique and revise candidate responses against those principles in place of direct human feedback at scale.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/constitutional-ai-concept",
"sameAs": [
"CAI",
"RLAIF (related concept)"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-08-22T15:54:53",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
