A Model for Using Ethical Theory to Specify Epistemic Goals for Explainable AI

This paper considers the normative requirements for “explainable AI” from the standpoint of moral accountability. If developers of sophisticated AI (or of the simple XAI designed to explain them) fail to accommodate these requirements, then it is likely that humanity will miss the crucial moment at which a sophisticated AI becomes the sort of entity that can be held accountable to its behavior. Existing AI already displays morally troubling behaviors that sometimes lead to the deaths of humans, and we have a great deal of difficulty assigning responsibility for the purpose of obviating these harms. By applying a well-regarded account of moral-worthiness (i.e., praise- and blame-worthiness), I provide a model for how ethical theories can be used to issue normative requirements for AI explainability.