Per una genealogia dell'intelligenza artificiale

The article aims to perform a genealogy of the current context of success surrounding Artificial Intelligence. It offers a description of all the conceptual alternatives that animated its history over the past seventy years. From the logicist version of AI to cybernetics, from connectionism to integration with machines, it explores the most promising results that emerged over the last decade, derived from methods based on training with large data corpora, aimed at extracting patterns to project onto future series of the same data for prediction and intervention purposes. These strategies, known as machine learning and deep learning, have achieved significant results; however, their applications require a clear objective function. Without it, we cannot evaluate their effects due to the impossibility of controlling interpretative processes and the organization of probability calculations, given that the data volume is too large for manual control. Moreover, these data-driven learning and training methods lead to outcomes that tend toward standardization and normalization, amplifying existing stereotypes within the corpora. This effect is particularly pronounced in the realm of generative AI. The artificial production of content cannot guarantee its truthfulness, as this would imply curating the training data an impossible task, given their vast quantity. Users of these tools are compelled to verify results, checking responses to avoid hallucinations or systematic errors that can occur unexpectedly but consistently. This should prompt us to reflect on political choices regarding which tasks can be autonomously performed by these systems, which require close integration with human capabilities, and which should be exclusively undertaken by humans, especially in cases in which there are decision-making processes, that require a clear accountability structure.