What is Constitutional AI and how does it let AI supervise itself?

The answer

Introduced by Anthropic in December 2022, Constitutional AI (CAI) is an alignment approach where an AI model critiques and revises its own outputs based on a set of written principles - its 'constitution.' The process has two phases: first, the model generates responses, then critiques and revises them according to constitutional principles (RL from AI Feedback); second, a preference model is trained on these self-revised comparisons and used for reinforcement learning. This reduces reliance on human labelers for identifying harmful outputs and makes the AI's values transparent and auditable through the explicit principles it follows.