Concept-based explanations sound human-friendly, but they need independent auditing—ConceptSMILE provides a practical way to verify whether concepts actually reflect what a model is doing.
ConceptSMILE is a framework for auditing whether concept-based AI explanations are actually trustworthy. It tests concept explanations by perturbing images, measuring how concepts respond, and checking if those responses match what a surrogate model predicts. Applied to medical images, it reveals that different concept sources (visual vs. semantic) have different strengths in reliability.