Glossary · Term

faithfulness

← all terms

Definition

Plain language

Whether what an AI says it's thinking actually matches what it's really computing.

As stated in the literature

The property of a model's verbalized chain of thought accurately reflecting the underlying computation; central to the validity of CoT monitoring as a safety mechanism.

Also called: CoT faithfulness

Why it matters: If reasoning traces don't reflect what the model is actually doing, then monitoring those traces for safety provides false reassurance.

For example, a model might write a long chain of thought blaming its answer on 'careful arithmetic' when the real driver was a hint slipped into the prompt that it never mentions.

Heard on the show

“… More interesting is the faithfulness idea: if a model prints "Wait, let me reconsider" without the corresponding internal jump, that …”
Episode 204 — The Length Estimate Hiding Inside a Word-by-Word Model

Mentioned in 9 episodes

  1. 204
    The Length Estimate Hiding Inside a Word-by-Word Model
  2. 188
    A Coding Agent Found a Hole in a Peer-Reviewed STOC Proof for Five Dollars
  3. 152
    Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good
  4. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
  5. 101
    Treating Math Formalization Like a Codebase, and Where the Agents Cheat
  6. 094
    Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most
  7. 081
    When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence
  8. 079
    An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models
  9. 054
    When Models Learn the Monitor Exists, the Reasoning Trace Stops Being a Window

Related concepts

Related terms