The companion entry showed that AI governance mandates the basis reach the receiver but does not adjudicate it. A natural rejoinder is that the field's explanation methods fill the gap: if a system can explain a decision, surely the receiver has the basis. This entry reads that literature on its own terms and finds the opposite. Even when an explanation is produced, the prevailing methods yield a plausible approximation, not evidence that a particular decision's basis was sufficient. The receiver cannot recover admissibility from a post-hoc explanation, and the field's own most careful voices say so.
Description is not justification
The sharpest framing of why explanation falls short of the basis comes from the legal literature, which has the clearest reason to demand justifiable decisions. Selbst and Barocas draw a distinction the technical debate tends to run together: a machine learning model can be inscrutable, hard to describe, and separately nonintuitive, describable but strange, and these demand different responses. Dealing with inscrutability requires a sensible description of the rules; addressing nonintuitiveness requires a satisfying explanation for why the rules are what they are (Selbst and Barocas, 2018). The methods surveyed below almost all attack the first problem, and their authors observe that doing so may not help if the goal is to assess whether the basis for decision-making is normatively defensible. The reason is put plainly: that decisions based on machine learning reflect the particular patterns in the training data cannot be a sufficient explanation for why a decision is made the way it is (Selbst and Barocas, 2018). A description of what the model does is not a justification that what it does is warranted. That gap, between describing the rules and justifying them, is the gap between an explanation and a basis, and it is the gap every method in this entry leaves open.
The field admits its measures do not certify soundness
The foundational call for rigour in the area is candid about what interpretability is for. It is a means to auxiliary goals, trust, fairness, causality, safety, that the system's own evaluation metrics cannot capture, which is precisely why a separate notion of interpretability is invoked at all (Doshi-Velez and Kim, 2017). That is an admission worth dwelling on: the field reaches for interpretability exactly because accuracy and its cousins do not certify that a decision is sound. Interpretability is offered as a proxy for the thing that cannot be measured directly. A proxy for soundness is not a certification of it, and interpretability so defined is not admissibility.
The dominant methods are approximations, not evidence
The sharpest critique comes from within explainable AI. The dominant post-hoc methods, local linear and gradient-based approximations such as the widely used technique that fits a simple model around a single prediction, are best understood as simplified models that approximate the true decision criteria, in the exact sense in which a scientific model approximates a physical system (Mittelstadt et al., 2018). The authors invoke Box's maxim, that all models are wrong but some are useful, and draw the consequence precisely: such a local approximation is an accurate representation only of a specific slice of the model's behaviour, it offers false assurances when a recipient does not understand where it breaks down, and it does not provide evidence of the trustworthiness or acceptability of the model overall. This is the fluency-disguises-basis point made against explainable AI's own flagship methods. A plausible-looking explanation of a decision is not a warrant that the decision's basis sufficed; it is a useful approximation that can, at worst, mislead the receiver into treating it as one.
In deployment, explanations do not even reach the receiver
If the theory says explanations approximate rather than certify, the practice says they often do not reach the affected party at all. The one substantial field study of how explainability is actually deployed, drawn from interviews with around fifty people across roughly thirty organisations, found that explanation techniques are consumed overwhelmingly by internal machine-learning engineers to debug and sanity-check models, not delivered to the end users affected by the decisions (Bhatt et al., 2020). The authors state it plainly: there is a gap between explainability in practice and the goal of transparency, because explanations primarily serve internal stakeholders rather than external ones. They further observe that organisations lack frameworks for deciding why they want an explanation, and that current research fails to capture the objective of an explanation. So the governance mandate from the companion entry, that basis reach the receiver, is failing empirically as well as conceptually: the explanation, such as it is, stays with the engineer.
The flagship legal remedy gives up on the basis
The clearest concession comes from the field's most influential legal-technical proposal. Faced with the difficulty of opening the black box under a putative right to explanation, the proposal for counterfactual explanations deliberately does not try. Counterfactuals, its authors write, do not attempt to convey the logic involved, and instead bypass the substantial challenge of explaining the internal workings of complex systems (Wachter et al., 2018). They go further, doubting whether human-comprehensible meaningful information about the logic involved in a particular decision can ever exist for modern models composing functions millions of times over. What they offer instead is a plausible, actionable surrogate that serves three receiver-side purposes: to understand a decision, to contest it, and to alter future behaviour to obtain a better outcome. These are exactly the receiver's needs at the handover, and the proposal meets them without ever representing the specific decision's basis. It is the field's canonical remedy conceding that the basis is not recoverable and substituting something useful in its place. Notably, the same authors who wrote the critique of approximation-as-explanation wrote this remedy; the concession is consistent, not incidental.
What the block establishes
Taken with its companion, this half of AI governance closes the rejoinder. The frameworks require that basis reach the receiver and even name the failure when it does not; the explanation methods meant to satisfy that requirement produce approximations, not evidence of a specific decision's basis, frequently do not reach the affected party, and in their most careful form explicitly decline to convey the logic at all. So a receiver at a machine handover cannot recover admissibility from a post-hoc explanation. That is the precise space this work occupies: not producing a better approximation of how the model behaved, but carrying, with a particular output, a representation of whether its basis suffices for the action, so the elevation of that output into an acted-upon fact can be refused when it does not.
References
Selbst, A. D. and Barocas, S. (2018). The Intuitive Appeal of Explainable Machines. Fordham Law Review, 87(3). ir.lawnet.fordham.edu/flr/vol87/iss3/11/
Doshi-Velez, F. and Kim, B. (2017). Towards A Rigorous Science of Interpretable Machine Learning. arXiv:1702.08608. arxiv.org/abs/1702.08608
Mittelstadt, B., Russell, C. and Wachter, S. (2018). Explaining Explanations in AI. FAT* 2019. arXiv:1811.01439. arxiv.org/abs/1811.01439
Bhatt, U., Xiang, A., Sharma, S., Weller, A., Taly, A., Jia, Y., Ghosh, J., Puri, R., Moura, J. M. F. and Eckersley, P. (2020). Explainable Machine Learning in Deployment. FAT* 2020. arXiv:1909.06342. arxiv.org/abs/1909.06342
Wachter, S., Mittelstadt, B. and Russell, C. (2018). Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR. Harvard Journal of Law & Technology, 31(2). arXiv:1711.00399. arxiv.org/abs/1711.00399