arXiv 2610.12493 Details Five Structural Flaws in XAI Faithfulness Metrics

The September 2026 submission argues that XAI tools have produced overreliance on incorrect predictions rather than calibrated trust. It frames faithfulness as approximation quality between explanation and model internals. Existing metrics rely on anchor points that introduce structural bias, rendering quantification either infeasible or prohibitively expensive for high-dimensional inputs.

Empirical patterns across LIME, SHAP, and integrated gradients show consistent divergence when ground-truth feature contributions are recovered from synthetic models. The paper isolates anchor dependency, perturbation instability, and monotonicity violations as primary drivers. These gaps were understated in prior surveys that treated metric scores as direct proxies for trustworthiness without testing structural assumptions.

Operationally the position reframes XAI as a debugging instrument rather than a trust mechanism. Adoption of the proposed avenues would require new benchmarks that decouple explanation fidelity from human-interpretable anchors. Teams building regulated systems would need to replace current pipelines with metrics that scale linearly with model width.

Research follow-up will test whether removing anchors reduces infidelity scores below 0.15 on tabular and vision benchmarks within 12 months.