Faithfulness Without Anchors: Unraveling the Structural Flaws of Explainable AI Metrics
Introduction to Explainable AI and the Trust Paradox
Explainable Artificial Intelligence (XAI) tools have long been promoted as the key to building user trust in machine learning models. By making complex algorithmic decisions more human-understandable, developers hoped to bridge the gap between black-box models and end-users. However, evidence accumulated over the past decade points to a sobering reality: XAI has largely failed to foster blind trust. Instead, it has sometimes promoted overreliance on incorrect predictions, ultimately leading to algorithmic aversion.
Rather than serving as a universal trust-building mechanism, XAI functions primarily as a debugging tool. It helps machine learning practitioners gain anecdotal insights into the features learned by their models. A central driver of this limitation is that XAI methods often oversimplify inner model mechanisms, causing a misalignment between the explanation and the actual model behavior—a phenomenon known as unfaithfulness or infidelity.
Understanding Faithfulness in Feature Importance
Faithfulness can be defined as the degree of approximation between an explanation and the underlying machine learning model. If we could compute faithfulness rigorously, we could accurately quantify a core element of an explanation's trustworthiness.
Unfortunately, existing metrics have severe limitations that render this quantification either entirely infeasible or computationally prohibitive.
A recent position paper presented at the IJCAI-ECAI 2026 Workshop on Explainable Artificial Intelligence—titled "Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance" (arXiv:2610.12493) by Marco Zullich and co-authors—takes a critical look at these challenges.
Core Entities and Concepts Defined
- Main Subject: Explainable AI (XAI) Faithfulness Metrics for Feature Importance.
- Key Authors: Marco Zullich, Emily Schiller, Ivan Gentile, David Dembinsky, Adriano Lucieri, Kilian Göller, and Steffen Seitz.
- Venue: IJCAI-ECAI 2026 XAI Workshop (arXiv:2610.12493).
Five Major Shortcomings of Current Faithfulness Metrics
According to the authors of arXiv:2610.12493, current methodologies suffer from severe structural misalignments. The paper outlines five major shortcomings:
- Over-simplification of Complex Mechanics: Explanations reduce high-dimensional, highly non-linear decision boundaries into rudimentary linear or local approximations.
- The Anchor Dependency Issue: Many metrics rely on rigid reference points or baselines that do not realistically represent the data distribution, distorting the faithfulness score.
- Computational Intractability: Rigorously calculating the exact alignment between model and explanation often demands exponential computational resources.
- Sensitivity to Perturbations: Minor, imperceptible changes to input data can drastically alter faithfulness scores without reflecting any real change in model logic.
- Task Misalignment: Metrics often evaluate explanations in a vacuum, failing to correlate with how humans actually utilize debugging insights in practice.
Four Proposed Avenues for Future Research
To move the field past these structural hurdles, the authors propose four promising research directions aimed at bolstering the reliability of faithfulness evaluation:
- Developing Anchor-Free Evaluation Frameworks: Moving away from rigid baselines to better capture true model behavior.
- Refining Structural Alignment: Designing metrics that respect the hierarchical and non-linear nature of modern neural networks.
- Scalable Approximation Techniques: Creating computationally feasible methods for estimating faithfulness without losing mathematical rigor.
- Human-Centric Validation: Aligning quantitative faithfulness metrics more closely with practical debugging workflows and empirical user studies.
Frequently Asked Questions (FAQ)
What is the primary thesis of arXiv:2610.12493?
The paper argues that current Explainable AI (XAI) metrics for feature importance suffer from structural misalignments (unfaithfulness), making rigorous quantification of trustworthiness nearly impossible with existing methods.
Who authored the paper "Faithfulness Without Anchors"?
The paper was authored by Marco Zullich, Emily Schiller, Ivan Gentile, David Dembinsky, Adriano Lucieri, Kilian Göller, and Steffen Seitz.
Where was this work presented?
It was presented at the IJCAI-ECAI 2026 Workshop on Explainable Artificial Intelligence (XAI).
Why are XAI explanations considered "unfaithful"?
Because current explanation tools often oversimplify the inner workings of complex machine learning models, leading to a significant divergence between what the explanation claims the model is doing and what the model is actually computing.
