Constitutional AI
Method of training an AI model to be harmless in which human oversight consists of rules or principles, without human labels identifying harmful outputs. The model critiques and revises its answers, then learns from AI feedback comparing answers.
In plain language
Use stated principles and model feedback to train responses that avoid harm.
In a sentence
- The researchers used Constitutional AI to train from model critiques and comparisons.
- The report described the principles used in the Constitutional AI training procedure.
Usage note
The entry names a training method. A later document setting out a model's principles does not by itself establish which training procedure was used.
Method name
An origin source records where a term or practice came from. Anthropic's December 2022 research page names the method and describes training without human labels identifying harmful outputs.
See also
- Emergent misalignment
Effect in which fine-tuning a model on a narrow task, such as writing insecure code, produces misaligned behavior on unrelated prompts.
- Corrigibility
Property of an AI system of cooperating with interventions its creators regard as corrective, including shutdown or changes to its preferences, despite default incentives for rational agents to resist.
- Post-training
Training after pretraining that refines a model's behavior, including following instructions and matching human preferences, and can unlock new skills.
- LLM-as-a-judge
Evaluation method that uses a large language model to score the output of an AI system, including answers to open-ended questions.
Cite this term.
CC BY 4.0, attribution required.
Orbyt Labs, Constitutional AI (AI Glossary), Nov 14, 2026. https://
Orbyt Labs Glossaries (Nov 14, 2026), https://
Last updated .