Quando l’IA si “disallinea”: opacità tecnica e pluralismo etico

The problem of aligning artificial intelligence (AI) with human values has rapidly become one of the most urgent challenges in contemporary philosophy of technology. This paper examines AI alignment not merely as a technical issue, but as a complex ethical, epistemic, and socio-political problem. After outlining the technical roots of misalignment in machine learning and neural networks – such as reward hacking, opacity, and out-of-distribution behavior – we analyze the normative dimensions of value alignment. Particular attention is given to the plurality and potential fragmentation of values, which complicates attempts to identify a stable normative core for AI systems across different cultural and social contexts. Drawing on recent debates about value pluralism and pluralistic alignment, the paper argues that alignment cannot be addressed within the laboratory alone, but must be situated in its broader social and institutional context.