Instead of paying humans to rank every harmful output, the model critiques and revises its own answers against an explicit list of principles, then trains on those self-improvements — a step called RLAIF (RL from AI Feedback). This scales alignment cheaply and makes the values auditable: you can read the constitution rather than guess what the labelers preferred. It's a big part of why Claude refuses and reasons about safety the way it does, and it reduces reliance on humans reviewing toxic content.
ELI5
Think of Constitutional AI as a simple recipe for doing the work better. Anthropic's method for aligning a model using a written set of principles (a 'constitution') instead of relying mostly on human labels.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.