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

XAI tools have largely failed their primary goal of increasing trust in AI systems. Rather than building appropriate confidence in model decisions, these explanation methods often promote "overreliance on incorrect predictions" and can lead to "algorithmic aversion," suggesting that current explainability approaches may actually harm human-AI collaboration rather than improve it.

Explanations frequently oversimplify how models actually work, creating a dangerous gap between what the explanation claims and what the model actually does. This "unfaithfulness" means users may be getting misleading insights into model behavior, which is particularly concerning when these explanations are used to make high-stakes decisions.

Current methods for measuring whether explanations are actually faithful to their models have significant limitations that make rigorous evaluation nearly impossible or computationally prohibitive. Without reliable ways to measure faithfulness, there's no systematic way to know which explanations can be trusted and which cannot.

XAI functions better as a debugging tool for developers than as a trust-building mechanism for end users. The paper suggests that explanations are more useful for engineers to gain "anecdotal insights" during model development rather than for helping non-technical stakeholders understand or appropriately trust AI system decisions.

The research community needs to fundamentally rethink how faithfulness metrics are designed. The authors propose four new research directions to address the structural problems with current metrics, indicating that incremental improvements won't solve the core issues with how we measure explanation quality.