Position paper argues current XAI faithfulness metrics are structurally misaligned with model behavior

A new position paper on arXiv argues that explainable AI (XAI) tools have not achieved their stated goal of increasing trust in machine learning models over the past decade. Instead, the authors contend, these tools often lead to overreliance on incorrect predictions and algorithmic aversion.

The paper focuses on the structural misalignment between XAI explanations and the underlying models, a problem referred to as unfaithfulness or infidelity. Faithfulness is defined as the degree to which an explanation approximates the model's behavior, and the authors argue that rigorous quantification of faithfulness is essential for assessing the trustworthiness of explanations.

The authors identify five major shortcomings of current faithfulness metrics, which they argue make rigorous quantification either infeasible or computationally expensive. These limitations undermine the ability to reliably evaluate how well an explanation reflects a model's true decision-making process.

To address these issues, the paper proposes four research directions aimed at bolstering the reliability of faithfulness metrics. The goal is to better align explanations with model behavior, thereby improving the utility of XAI as a debugging tool rather than a mechanism for increasing user trust.

The work was presented at the IJCAI-ECAI 2026 Workshop on Explainable Artificial Intelligence (XAI), indicating its relevance to ongoing discussions in the explainability research community.