Do Large Language Models Refute or Reaffirm Chomsky’s Poverty of the Stimulus?

Abstract

The Poverty of the Stimulus (POS) argument holds that core properties of human grammars-such as structure-dependence, binding constraints, and island effects-are underdetermined by the primary linguistic data available to children, implying the existence of domain-specific innate knowledge (Chomsky, 1965, 1980, 1986). The spectacular linguistic performance of contemporary large language models (LLMs), trained without explicit supervision on internet-scale text, has been widely interpreted as an empirical refutation of the POS: if statistical learning at scale suffices for machines, perhaps it also suffices for children.

This article argues that LLMs neither refute nor resolve the classic POS argument. Instead, they circumvent it through three decisive advantages unavailable to human learners: (1) exposure to five orders of magnitude more data than any child receives; (2) training corpora that are covertly enriched with metalinguistic information, parsed examples, and pedagogical artefacts; and (3) transformer architectures whose inductive biases (attention, positional encodings, residual connections) constitute a powerful, hard-wired theory of linguistic structure. When these confounding factors are controlled for, systematic failures of generalisation persist in exactly the domains where the POS predicts difficulty. Far from dissolving the logical problem of language acquisition, current models reaffirm its force while highlighting how far computational systems remain from genuine infant-like learning.