The Conceptual Architecture of Artificial Intelligence: From Mechanical Computation to Statistical Prediction

Abstract

This article reconstructs the conceptual architecture of artificial intelligence by tracing its intellectual genealogy and the methodological shift that leads from early projects of mechanical and programmable computation to the contemporary paradigm of statistical prediction. The analysis begins with nineteenth-century antecedents associated with the Analytical Engine and Ada Lovelace’s “Notes”, and then situates the mid-twentieth-century foundational turn marked by the theory of computation and the research programme inaugurated at Dartmouth. On this historical basis, the article draws a structural distinction between symbolic artificial intelligence—grounded in explicit rules and inferential traceability—and statistical artificial intelligence, whose performance relies on the optimisation of mathematical functions over large volumes of data. The central section characterises large language models as systems of probabilistic inference over linguistic sequences and identifies three conceptually decisive features: the absence of semantic understanding and intentionality, dependence on statistical correlations, and structural opacity. The contribution is limited to the conceptual level. Its purpose is to establish a precise analytical framework that avoids conflating linguistic performance with reasoning and that enables subsequent normative—including constitutional—assessments of the use of artificial intelligence in decision-making functions.