This work presents a rigorous philosophical examination of the notion of understanding and semantics in artificial intelligence language models. Starting from the main currents of philosophy of language (Chomsky, Frege, Russell, Wittgenstein, Quine, Putnam, and Kripke), the classical distinction between syntax and semantics in analyzed, along with various theories of reference, logical atomism and language game theory. Subsequently, the fundamental technical characteristics of language models (T…
Read moreThis work presents a rigorous philosophical examination of the notion of understanding and semantics in artificial intelligence language models. Starting from the main currents of philosophy of language (Chomsky, Frege, Russell, Wittgenstein, Quine, Putnam, and Kripke), the classical distinction between syntax and semantics in analyzed, along with various theories of reference, logical atomism and language game theory. Subsequently, the fundamental technical characteristics of language models (Transformers architecture, embeddings, and attention mechanisms) are described and contrasted with philosophical conceptions of meaning and reference. On this basis, operational criteria are proposed for attributing “understanding” to such models (referential anchoring, adoption of normative practices, inferential capacity, pragmatic sensitivity, traceability, and experience) and their applicability is evaluated in formal, empirical, and social domains. Finally, employing ideas from Popper and Deutch, an emergentist and conjectural theory of knowledge is articulated, in which creativity, refutability, and the difficulty of variation of explanations constitute its pillars. The study concludes by suggesting pathways for endowing future artificial intelligences with sensory and multimodal anchoring, with the aim of expanding their semantics beyond the statistical processing of language.