Are Explanations Helpful? A Comparative Study of the Effects of Explanations in AI-Assisted ...
summary
Randomized controlled experiments on MTurk (782 recidivism, 561 forest cover) compare five conditions: feature importance, feature contribution, nearest neighbors, counterfactuals, and no explanation. Three desiderata are tested: model understanding, uncertainty awareness, and trust calibration. Feature importance and feature contribution increase appropriate trust and reduce undertrust on recidivism tasks, especially when model confidence ≥0.7, though feature importance can raise overtrust on low-confidence cases. Nearest neighbors and counterfactuals show no reliable effect, and almost no explanation helps tasks where participants lack domain knowledge.
AI-written summary · about
related talks
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability
45mAI-404: Explanation Not Found
41mFostering Trust via Explainable ML Inferences
32mOn the Importance of Explainability
24mToward Trustworthy AI: Explainability and Robustness
22mPersonalized explanations for hybrid recommender systems
19mMoving Forward with Explainable AI: Adoption Rate & Essential Dimensions for a Reliable Use
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python