In the current literature on AI, at least two notions of “explainability” occur, one ‘extrinsic’ and one ‘intrinsic’: extrinsic explanations, on the one hand, occur when AI systems produce outputs which are taken by human agents as ‘explanations’ of certain states of affairs in a given sociotechnical environment (e.g. decision support systems in law or medicine). Intrinsic explanations, on the other hand, call for ‘causal explanations’ of certain outputs of a given system, taking into account the formal structure of the system as a technological artifact. In practical contexts, it turns out that the boundaries between both of these dimensions of explanation often seem to be blurred, while their underlying objectives vary significantly – this fact becomes especially problematic if ethical considerations are involved.

Based on a short review of recent philosophical literature on explanations, the talk provides an overview over the explainability debate both in ethics and mathematics/technology and places it within the broader context of an empirically informed philosophy of technology and mathematics. It then argues for an integrated account of AI explainability.