t'olofin AI

Constitutional AI: Harmlessness from AI Feedback presents a training method with supervised and reinforcement-learning phases. The constitution is written normative guidance used to evaluate responses, rather than a legal contract or the model’s architecture. In the supervised phase, the system generates responses, produces critiques under selected principles, and revises those responses. The model is then fine-tuned on the revised answers. Critiquing a draft and training from revisions are different operations: the former creates examples, while the latter changes model parameters. In the reinforcement-learning phase, a model compares response pairs according to principles and creates AI preference data. A learned preference model then supplies a reward signal for reinforcement learning. This is the RLAIF component of the stated method, not an instruction to replace every possible human training signal. The original experiments retain human helpfulness comparisons while replacing human harmlessness labels with AI-generated feedback. People also choose principles and assess results. Written guidance can make intended values easier to inspect, but cannot expose every learned internal preference or prove that every output follows those values. Editing the document alone does not retroactively retrain an already saved model; a relevant training or runtime process must actually use the changed guidance. Anthropic’s later specific-versus-general principles study compares broad and detailed guidance in particular dialogue-model experiments. Its findings do not imply that one short rule reliably resolves every normative conflict. Historical methods and principles should be identified by their version rather than treated as a complete description of every current Claude model.