The Challenges of Traceability in the Era of Black Boxes

Abstract

This article offers a critical examination of the AI Act, centering its analysis on the functional framework it establishes and the way it allocates distinct roles and responsibilities to the several actors involved in the value chain of artificial intelligence systems. In particular, it examines the distribution of obligations throughout the lifecycle of such systems, with special attention to contexts of algorithmic opacity in Black Box systems, where internal decision-making processes are inaccessible or difficult to reconstruct, even for their own developers. The article questions whether the AI Act can ensure effective attribution of responsibilities in light of technological advancement and the proliferation of synthetic content, given that technical traceability becomes increasingly challenging. The central argument posits that a possible solution lies in strengthening the evidentiary framework and incorporating the figure of the Logger, which would allow for the preservation of decision-making records and prevent obligations from being reduced to a merely formal and ineffective categories. Nevertheless, it concludes that practical implementation poses a significant challenge; because the exhaustive decomposition of an AI system’s decision-making chain, when its design prioritizes efficiency over explainability, would, in many cases, be incompatible with the very operational logic of efficiency.