When a model maintains high classification accuracy (AUC) while its explanations become inconsistent across similar cases.