Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence

In 30 Sekunden

Was passiert ist

Researchers tested whether LLM explanations of decisions actually match the factors that drive those decisions using controlled interventions across Claude, GPT, and Gemini models.

Warum es zählt

Engineers building systems where users rely on LLM explanations to monitor, debug, or override decisions need to know if those explanations are trustworthy.

Achtung

Cited factors often do not match measured influence; uncited factors sometimes score higher than cited ones, suggesting explanations may mislead operators about true decision drivers.