Bai et al. 2022: Constitutional AI — harmlessness from AI feedback

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Constitutional AI can train a harmless but non-evasive AI assistant without any human harmfulness labels.claim0.844
  The paper's central claim, supported by findings that RL-CAI outperforms HH RLHF in harmlessness while being non-evasive.
• Constitutional AI methods can be applied broadly to steer model behavior, e.g., writing style, tone, persona, not just harmlessness.claim0.841
  Discussion section suggests generalizability beyond harmlessness.
• The constitutional approach makes it easier to control AI behavior precisely and with far fewer human labels.claim0.813
  Explicit principles replace large datasets of preference labels, enabling faster iteration.
• The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'.quote0.811
  Defines the core concept of the paper.
• AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al. 2024)concept0.804
  Related work studying capability of LLMs to subvert safety measures if severely misaligned
• Reinforcement Learning Constitutional AIframework0.803
  The RL stage of CAI using AI feedback to train a preference model, then RL, resulting in a policy trained by RLAIF.
• Constitutional AIframework0.800
  Alignment approach by Anthropic that explicitly trains self-observation; predicts highest baseline and lowest prompt lift.
• Constitutional AI explicitly trains self-observation-like behavior, which is why CAI models score highest and show lowest contemplative lift.claim0.791
  Interpretive claim connecting the battery's circularity to the empirical finding