Training a model against a written set of principles rather than case-by-case human judgements. The model critiques and revises its own outputs against the stated rules, and those revisions become training data. The advantage is that the values are written down and can be argued with.
Technical view
Combines a supervised phase, where the model self-critiques and revises against principles, with a preference phase where the principles guide AI-generated preference labels. Introduced by Anthropic. It makes the normative content explicit and auditable rather than implicit in annotator behaviour, and it scales oversight, though principles still require interpretation and can conflict, so the ordering and wording of the document does real work.
Related
Part of AI Terminology Index, section: Training and Post-training.