Asked about Studying behaviour: honesty, faithfulness and sandbagging
How would you test whether a model's chain of thought is faithful?
An answer that works
Intervene on something that should affect the answer, such as a biasing hint, and check whether the answer changes and whether the reasoning acknowledges the cause. Answers that move with the hint while the reasoning never mentions it are evidence of unfaithful reasoning.
Then they push further
Why does this matter for safety?
Monitoring a model's reasoning is one proposed way to catch problems. If the stated reasoning does not reflect what drives behaviour, that monitoring gives false confidence.
What gets marked up
Marked up for an intervention with a control, and for connecting faithfulness to monitoring.
Reading an answer is not being able to give one. The section that teaches this is Studying behaviour: honesty, faithfulness and sandbagging, 4 hours of Safety & evaluation research.
Learn this properlyAlso asked about Safety & evaluation research