Termine

Abstract: Developers, users, and regulators of modern AI systems demand that AI should be, e.g., reliable, transparent, trustworthy, and conform with moral and legal rights. In order to ensure this, it is frequently claimed, we must be able to explain how the system in question works. In this talk, I shall discuss how relevant explanatory information may be gathered from a philosophy of science point of view, specifically by looking at recent debates about mechanisms. I shall contrast a glassboxing-approach to explainable AI (XAI) with one relying on interpretability methods. Building on the pattern account of mechanistic inquiry (Kästner and Haueis 2019), I shed light on what norms guide discovery and explanation in XAI research and how different contexts affect what norms are given priority. I close by comparing the epistemic situation in XAI with that in other disciplines. [Poster]