ConceptSMILE: Auditing the Trustworthiness of Concept-Based Explainable AI

Concept-based explanations need independent verification: The paper demonstrates that making AI explanations more human-understandable through concepts doesn't automatically make them trustworthy—they require separate auditing to verify their reliability, similar to how a financial audit checks accounting claims.

Different explanation methods have different strengths: Testing two approaches to generate concepts (MedSAM visual concepts vs. VLM semantic concepts) revealed trade-offs: MedSAM performed better at spatial accuracy (R² = 0.8503), while VLM concepts showed stronger stability when tested against image artifacts, suggesting no single method is universally superior.

Reliability varies across individual concepts: Rather than declaring an entire explanation system trustworthy or not, ConceptSMILE evaluates each concept separately, finding that some concepts are more reliable than others—an important distinction for practitioners deciding which explanations to trust in high-stakes decisions.

A practical auditing framework is now available: ConceptSMILE provides a model-agnostic tool that organizations can use to independently test concept-based explanations across multiple reliability dimensions (accuracy, fidelity, faithfulness, stability, and consistency), enabling more rigorous governance of explainable AI systems.

Medical AI explanations can be systematically validated: The framework was tested on retinal fundus images, demonstrating that concept-based explanations in healthcare—where trustworthiness is critical—can now be audited using standardized, measurable criteria rather than relying on subjective assessment.