SESKStuff Everyone Should Know
SESK: Stuff Everyone Should Know · sesk.info/ai-constitutional-ai-rlaif

In Constitutional AI, what is the only human oversight used to judge harmfulness?

A written list of rules or principles, the constitution.

Start playing free

  1. A written list of rules or principles ✔
  2. Thousands of human harm labels
  3. A human reading every answer

Why: Anthropic's 2022 method has a model critique and revise its own answers against written principles, then a model picks the better of two answers. That AI feedback trains a preference model for reinforcement learning, called RLAIF.

Huh, didn't know that: RLAIF stands for reinforcement learning from AI feedback.

The dad joke
My kids have a constitution too. It says bedtime is a suggestion.
Like this card?
Play the game

References

Answer: A written list of rules or principles. Anthropic's 2022 method has a model critique and revise its own answers against written principles, then a model picks the better of two answers. That AI feedback trains a preference model for reinforcement learning, called RLAIF.

  1. Bai et al., Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073) arxiv.org/abs/2212.08073
    The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'.
    Checked 2026-10-10.

Think this answer is wrong? Tap "Challenge this answer" on the card in the game and tell us why.

Test yourself in the game

Artificial IntelligenceIn the InstructGPT study, which answers did people prefer: a 1.3B RLHF model's or the 175B GPT-3's?Artificial IntelligenceWhat causes specification gaming in an AI system?Artificial IntelligenceWhat are the two broad targets of reward tampering?

That's stuff everyone should know.