Constitutional AI
MethodAnthropic's method for training a harmless assistant from a short list of written principles rather than human labels of harmful outputs. In a supervised phase the model critiques and revises its own responses against the principles and is fine-tuned on the revisions; in a reinforcement phase, AI-generated preference labels guided by the principles train a reward model.
Read
Papers
- Bai et al.: Constitutional AI: Harmlessness from AI Feedback (opens in a new tab)arXiv / paper authors · 2022
About this resource
The Constitutional AI paper: a supervised critique-and-revise phase followed by reinforcement learning from AI feedback produces a harmless assistant that explains its objections instead of evading.
Where to read: Abstract
Checked 2026-10-31
Authority for this purpose: arXiv / paper authors publishes this original research record.
Destination inspection: Bounded public-address-pinned GET followed by destination text and concept-scope inspection
Useful destination; no explanatory claim is asserted from this link.
- arXiv record is a preprint. Verified the abstract and version record; no claim of full-paper review or venue status from the arXiv URL.
Reviewed authoritative resource
Where to read: Abstract
A first-party destination supports its documented scope; it does not establish independent performance validation or historical priority.
- Lee et al.: RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback (opens in a new tab)arXiv / paper authors · 2023
About this resource
Google compares RL from AI feedback with RL from human feedback on summarization and dialogue and finds human raters prefer them about equally.
Where to read: Abstract
Checked 2026-10-31
Authority for this purpose: arXiv / paper authors publishes this original research record.
Destination inspection: Bounded public-address-pinned GET followed by destination text and concept-scope inspection
Useful destination; no explanatory claim is asserted from this link.
- arXiv record is a preprint. Verified the abstract and version record; no claim of full-paper review or venue status from the arXiv URL.
Reviewed authoritative resource
Where to read: Abstract
A first-party destination supports its documented scope; it does not establish independent performance validation or historical priority.
Engineering blogs
- Constitutional AI: Harmlessness from AI Feedback (opens in a new tab)Anthropic · 2022
About this resource
Anthropic's summary of the method and its motivation: less human labeling of harmful content and more transparent training objectives.
Where to read: Research summary
Checked 2026-10-31
Authority for this purpose: Anthropic publishes this first-party source.
Destination inspection: Bounded public-address-pinned GET followed by destination text and concept-scope inspection
Useful destination; no explanatory claim is asserted from this link.
- First-party account of the publisher's own system or practice; reported results are self-reported, not independently validated.
Reviewed authoritative resource
Where to read: Research summary
A first-party destination supports its documented scope; it does not establish independent performance validation or historical priority.
- Claude's Constitution (opens in a new tab)Anthropic · 2023
About this resource
Describes the principles Anthropic wrote for Claude, where they came from, and how they are used in Constitutional AI training.
Where to read: Full article
Checked 2026-10-31
Authority for this purpose: Anthropic publishes this first-party source.
Destination inspection: Bounded public-address-pinned GET followed by destination text and concept-scope inspection
Useful destination; no explanatory claim is asserted from this link.
- Body is client-rendered; only the destination identity was inspected automatically. Anthropic has since published revised versions.
- First-party account of the publisher's own system or practice; reported results are self-reported, not independently validated.
Reviewed authoritative resource
Where to read: Full article
A first-party destination supports its documented scope; it does not establish independent performance validation or historical priority.
Related concepts
Prerequisites
Broader
- Preference optimizationKnowledge area
Related
Loading discussion…
