A Formal Methodological Framework for Auditing Robustness and Fidelity in Explainable AI: From Application to Trust Certification

Popular explanation tools like SHAP and LIME can produce unreliable outputs: The paper found that these widely-used AI interpretation methods can give substantially different explanations when input data is slightly altered by noise, raising fundamental questions about whether we can trust their results.

High model accuracy does not guarantee trustworthy explanations: The researchers discovered that models achieving over 99% accuracy (AUC > 0.99) can still produce explanations that are "numerically degenerate or flatly uninformative," meaning a model's strong predictive performance tells us nothing about whether its explanations are actually meaningful.

A new auditing framework is needed to verify AI explanations: The paper proposes measuring two key properties—robustness (stability of explanations under small data changes) and fidelity (whether identified important features actually influence predictions)—combined into a "Trust Score" to systematically evaluate whether explanations can be relied upon.

Overfitting undermines explanation quality in ways standard metrics miss: When models are overfit to training data, their fidelity scores become unreliable, meaning traditional performance metrics fail to catch problems with explanation trustworthiness—a critical gap for high-stakes applications.

Auditing explanations is essential for sensitive real-world applications: The authors emphasize that verifying XAI outputs is "not optional but necessary," particularly in domains like food security and malnutrition prediction where AI-informed decisions directly affect human welfare.