Trains a model to be harmless using a written set of principles: the model critiques and revises its own outputs, and AI-generated preference labels replace most human harmlessness labels.

This is an external publication hosted at arxiv.org. AI Governance LK did not write it; the summary and notes on this page are our own.

Why it matters

Constitutional AI is an early, practical example of using AI systems to help supervise AI systems — a theme that runs through scalable-oversight research. It also makes the values being trained into a model explicit and inspectable.

What you will learn

  • Explain the critique-and-revise and RL-from-AI-feedback stages
  • Discuss the trade-offs of replacing human labels with model labels

Before you start

Suggested background. This is editorial guidance, not a requirement.

  • Familiarity with RLHF

How to cite

Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.