Can explanations be audited legally?

Lens adapted to this topic: Limits of explanation-based legal audits

A skeptical reading argues that auditing an explanation is not the same as auditing the decision or its legal justification. A system can produce a plausible explanation that is post-hoc, insincere, or too local to reveal how the model behaves more broadly. Under this view, explanations may assist investigations but cannot by themselves establish accountability or legality.

• Post-hoc explanations may be plausible without faithfully representing the model’s actual reasoning, making them difficult to use as reliable legal evidence.
• Third-party auditors may need an exorbitant number of queries to verify local explanations in complex, high-dimensional settings.
• A pointwise prediction and explanation may be insufficient for users to verify a system’s behavior in complex settings.
• Explanation rights may fail to bridge the gap between statistical inference and the causal or normative justification expected by law.
• Explanations can contribute to accountability, but they are insufficient on their own as a legal foundation for algorithmic accountability.

Featured sourceAuditing Local Explanations is Hardproceedings.neurips.cc3 sourcesView sources