Explainable Black Box Models

The idea that black box models are unexplainable has been elevated to an axiom. In contrast to explainable models, where the involved parameters have a certain meaning that is understandable, black box models rely on principles that have no other function than to produce sophisticated mappings. The lack of explainability is sometimes compensated by applying a so-called post-hoc explainability method. Such a method is supposed to extract some meaning from the separately constructed black box model, but this technique is criticized for several reasons. In this paper we argue that there is an alternative to explainable models and to post-hoc explainability, by representing explanations as variables at either the input side or the output side of a black box model. This results in explainable black box models, where explanations are unrelated to the working principles, but they are still part of that same black box model.