Explanation for Whom? Hospitable Interpretability for Machine Learning

A developer and a regulator can carefully read the same feature-attribution plot and still disagree about whether anything has been explained. The properties by which interpretability evaluates its artifacts---faithfulness, sparsity, stability---describe how an artifact relates to a model, and say little about its relation to the question someone has come to ask. The deductive-nomological account of explanation exposed the same problem: explanatory force depends not only on the adequacy of an answer, but on the why-question to which it is addressed. For systems making consequential decisions, this shifts attention from the search for a single better artifact, or for separate artifacts tailored to separate audiences, toward hospitality: a property of the record in which several legitimate questions about a case each receive a marked answer drawn from the same decision, with the places they diverge kept visible rather than smoothed into a single verdict.