A logical approach to algorithmic opacity

In [ 1 ], we introduced a novel definition for the epistemic opacity of AI systems. Building on this, we proposed a framework for reasoning about an agent’s epistemic attitudes toward a possibly opaque algorithm, investigating the necessary conditions for achieving epistemic transparency. Unfortunately, this logical framework faced several limitations, primarily due to its overly idealized nature and the absence of a formal representation of the inner structure of AI systems. In the present work, we address these limitations by providing a more in-depth analysis of classifiers using first-order evidence logic. This step significantly enhances the applicability of our definitions of epistemic opacity and transparency to machine learning systems.