Eliminating I-Language in Large Language Models

Chomsky’s (1986) notion of I-language holds that human linguistic competence is grounded in an internal, intensional system of rules and representations. Large language models (LLMs) now produce syntactically rich and context-sensitive language without any evident symbolic grammar, which raises a question of explanatory necessity: when a system performs linguistically, must an I-language be posited to explain that performance, or does the posit merely redescribe the very behavior it was meant to explain? This paper argues for the latter and develops a formal eliminative argument. Drawing on Lohndal’s minimalist syntax, Glanzberg’s contextual semantics, and Field’s paracomplete logic, it first shows that transformer parameter spaces fail the structural criteria of an I-language, namely discrete, asymmetric, recursive operations such as Merge. It then shows that continuous vector geometry, through multi-head attention and paracomplete negation, accounts for syntactic, semantic, and pragmatic behavior without discrete rules or modular boundaries. A further discussion of Chomsky’s internalist rejection of semantic values, and its Moorean critique, clarifies what is at stake for reference and meaning. The conclusion is methodological: internal computational structure does not amount to an internal language in the theoretically relevant sense, and fluent output alone therefore provides no ground for attributing an I-language, or richer mental states, to such systems.