Explainable Artificial Intelligence (XAI) tools are often marketed as ways to boost trust in machine learning models by rendering aspects of their algorithms more understandable to humans. Yet a decade of evidence shows that this goal has largely failed, leading instead to overreliance on incorrect predictions and subsequent algorithmic aversion. In practice, XAI functions more as a debugging aid, offering anecdotal insights into features learned by the model. The core limitation stems from XAI methods oversimplifying the model’s inner workings, creating a mismatch between the explanation and the model itself—a phenomenon known as unfaithfulness or infidelity. Faithfulness can be defined as the degree of approximation between an explanation and the model; a rigorous computation of faithfulness would allow quantification of one aspect of an explanation’s trustworthiness. Existing faithfulness metrics, however, suffer from major drawbacks that render such quantification either infeasible or computationally prohibitive. This paper identifies five key shortcomings: inability to capture model non‑linearity, neglect of feature interactions, insufficient sensitivity to input perturbations, lack of reproducibility in evaluation, and poor generality across model architectures. To address these issues, we propose four research avenues: adopting locally linear approximation frameworks for interpretability, developing metrics that capture higher‑order feature interactions, constructing robustness‑focused faithfulness assessments under adversarial perturbations, and establishing standardized benchmark suites to improve reproducibility. Pursuing these directions could enhance the structural alignment of XAI explanations and genuinely strengthen trust in model predictions.
Review