Constitutional AI

Anthropic's technique of training models to critique and revise their own outputs against a written set of principles, reducing dependence on human labels for harmlessness.