Daily papers

Sunday, 15 November 2026 · Weekend classic from 2022 · 6 min read

Training harmless AI assistants using AI-generated feedback and constitutional principles

This paper introduces Constitutional AI, a method that trains AI assistants to be harmless and non-evasive by using AI-generated critiques and preference labels based on a set of human-written principles, reducing reliance on human feedback labels for harmfulness.

Based on the paper Constitutional AI: Harmlessness from AI Feedback by Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell and 47 others

Challenges in training helpful and harmless AI assistants

As AI systems grow more capable, ensuring they remain helpful, honest, and harmless becomes increasingly important but challenging. Traditional methods rely heavily on human feedback to supervise AI behavior, which is costly and difficult to scale.

A key difficulty is balancing helpfulness and harmlessness: models trained to be harmless often become evasive, refusing to engage with sensitive queries, which reduces their usefulness and transparency.

To address these issues, the authors seek methods that reduce human supervision, encode behavioral goals transparently, and produce AI assistants that engage non-evasively with harmful queries by explaining their objections.

Sign in to read the whole paper

You have read the opening. The rest is free with a Bezen account, and no card is needed.

In the rest of this paper

  • How the method works, step by step, with a flow chart
  • Three figures from the paper, and how to read them
  • How it was tested, what was found, and where it falls short
Training harmless AI assistants using AI-generated feedback and constitutional principles · Bezen