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

URL SCAN: Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance
FIRST LINE: Computer Science > Artificial Intelligence

The Dissection

This paper is an autopsy of XAI’s credibility problem. It argues that faithfulness metrics often fail to measure whether an explanation actually tracks the model’s behavior, while still being treated as evidence of trustworthiness. Its proposed research agenda attempts to repair the measurement layer: better anchors, more reliable evaluations, and lower computational cost.

The paper’s sharper finding is that XAI has not delivered the promised trust. It has often produced explanations that make incorrect predictions easier to accept. But the paper remains inside the governance frame it is criticizing. It treats the crisis as a defective instrument problem rather than a structural conflict between opaque optimization systems and human demands for intelligibility, accountability, and control.

The Core Fallacy

The central error is conflating explanatory faithfulness with trustworthiness. An explanation can be perfectly faithful to a model that is biased, strategically manipulated, brittle, or operating on illegitimate proxies. Accurate disclosure of a bad mechanism does not make the mechanism safe. It merely makes the failure more legible.

The paper also treats better measurement as if it could close the gap between machine behavior and human oversight. It cannot. Faithfulness metrics evaluate an approximation under selected perturbations, baselines, features, and model interfaces. Those choices are themselves anchors imposed by the evaluator. There is no neutral measurement window into a high-dimensional system—only a controlled slice that may or may not matter operationally.

Under the Discontinuity Thesis, this is a secondary technical dispute. Faithfulness metrics do not weaken cognitive automation. If anything, they can strengthen it by making deployed systems easier to audit, defend, and institutionalize. The paper is trying to improve the dashboard on a machine whose ownership and productive consequences it does not examine.

Hidden Assumptions

• That model behavior can be meaningfully approximated by explanations at a cost institutions will accept.
• That the relevant model is stable enough for a faithfulness score to remain valid after deployment, retraining, distribution shift, or adversarial adaptation.
• That perturbing or removing features reveals causal importance rather than artifacts of the chosen intervention scheme.
• That feature importance is an adequate unit of explanation for models whose behavior depends on interactions, representations, temporal context, and hidden state.
• That improved explanation reliability will produce calibrated human trust rather than more sophisticated overreliance.
• That institutions possess the incentives and authority to act on explanations once failures become visible.
• That the central problem is epistemic—knowing what the model used—rather than political and economic—deciding who controls the model and who absorbs its errors.
• That making automated decisions more defensible will not accelerate their deployment across cognitive work.

Social Function

Classification: partial truth, transition management, and ideological anesthetic.

The partial truth is substantial: current faithfulness metrics are structurally weak, expensive, and vulnerable to misleading interpretations. The transition-management function is equally clear. By improving explanation audits, the paper helps institutions normalize increasingly capable automation without resolving the legitimacy crisis created by delegating judgment to machines.

Its ideological anesthetic is the implied belief that better explanations can restore a human-centered relationship with AI. They cannot. They may reduce debugging uncertainty and improve selected forms of model oversight, but they do not restore productive participation, distribute ownership, or prevent automated systems from replacing cognitive labor. They make the replacement apparatus more inspectable, not less replacing.

The Verdict

This is a useful technical warning trapped inside a narrow reformist frame. It correctly identifies that many faithfulness scores are unanchored approximations masquerading as rigorous evidence. But it mistakes measurement repair for structural repair.

The paper can improve XAI as a debugging and governance instrument. It cannot make explanations trustworthy by metric alone, and it cannot reverse the Discontinuity Thesis. At best, it supplies better instrumentation for the transition into machine-dominated cognition. The dashboard may become more accurate. The operators still do not control the ownership structure, and the human labor circuit remains marked for deletion.