Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance

Explainable Artificial Intelligence (XAI) methods have been promoted as tools to build trust by making machine‑learning models more understandable to users. Yet, evidence accumulated over the past decade shows that, in practice, these approaches fail to achieve that goal and can even foster overconfidence in incorrect predictions, which in turn breeds aversion to algorithms. Rather than serving as a guarantee of reliability, XAI appears to be more useful as a debugging aid, offering anecdotal insights—for example, into which features a model has learned. A central problem is the tendency of explanatory techniques to oversimplify a model’s internal workings, creating a disconnect between the explanation and the actual model, a phenomenon known as lack of fidelity or infidelity. Fidelity is defined as the degree of approximation between the explanation and the model’s behavior; rigorously measuring it would allow quantification of a key component of explanation reliability. The position paper identifies five major limitations of current fidelity metrics and proposes four research directions to improve their robustness and computational feasibility. This discussion is relevant because, without reliable metrics, users may rely on misleading explanations, compromising AI‑based decision making and hindering the safe adoption of these systems.

Read the original article on arXiv cs.AI

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.