Constitutional AI

An AI safety approach developed by Anthropic where models are trained to follow a set of principles (a 'constitution') through self-critique and revision. The model learns to identify and correct harmful outputs based on these principles, reducing the need for extensive human feedback.