Constitutional AI: Harmlessness from AI Feedback (Bai et al. 2022b)

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.claim0.848
  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.claim0.831
  Discussion section suggests generalizability beyond harmlessness.
• Constitutional AIframework0.814
  Alignment approach by Anthropic that explicitly trains self-observation; predicts highest baseline and lowest prompt lift.
• The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'.quote0.798
  Defines the core concept of the paper.
• The constitutional approach makes it easier to control AI behavior precisely and with far fewer human labels.claim0.787
  Explicit principles replace large datasets of preference labels, enabling faster iteration.
• AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al. 2024)concept0.786
  Related work studying capability of LLMs to subvert safety measures if severely misaligned
• Constitutional AI produces a distinctive signature: high boundary_awareness, low aesthetic_response relative to peers.claim0.781
  Interpretive finding from dimension profile analysis: training for honest limits comes at cost to aliveness.
• Contemplative Constitutional AIframework0.777
  Paper's proposed adaptation of Constitutional AI incorporating contemplative wisdom charter