Constitutional AI (CAI) trains a model to be helpful and harmless by having it critique and revise its own outputs against a written set of principles (a "constitution") — reducing reliance on humans labeling every harmful example.
How it works
1. Model answers a prompt
2. Model critiques its own answer against the constitution ("was that harmful? why?")
3. Model revises the answer
4. Train on the improved answers → "RLAIF" (RL from AI Feedback)It's a twist on RLHF: the feedback comes from the model judging itself by stated rules, not only from human raters.
Why it matters
- Scalable — humans can't review billions of outputs; a constitution can.
- Transparent — the values are written down and debatable, instead of buried in opaque human labels (The Black Box Problem).
- A partial answer to whose values: at least they're explicit.
It doesn't solve alignment — a model can still be jailbroken or harbor backdoors — but it's a practical step, and one defensive layer among many.
Related: Reinforcement Learning from Human Feedback (RLHF) · The Value Loading Problem · Guardrails and Defense in Depth for LLM Apps