Definition
Plain language
A way to train an AI to behave using a written list of principles that another AI checks its answers against, instead of paying humans to label everything.
As stated in the literature
A training and oversight approach using an explicit written constitution plus a model-based judge to score and refine responses, scaling alignment supervision beyond human labeling; the basis for Self-CTRL's consistency experiments.
Also called: Constitutional AI
Why it matters: It scales safe behavior using written principles and an automated judge, instead of relying on humans to hand-label every response.
For example, a model's draft reply is checked by another AI against a written rule like 'don't give instructions for harm,' and rewritten if it breaks one.
Heard on the show
“… It's that an entire generation of safety methods — Constitutional AI, OpenAI's deliberative alignment — is built on the opposite assumption: that if you train a …”Episode 171 — The Safety Decision a Model Makes Before It Thinks a Word