Several recent publications in cognitive science have made the suggestion that Large Language Models (LLMs) have mastered human linguistic competence and that their doing so challenges arguments that linguists use to support their theories (in particular, the so-called argument from the poverty of the stimulus). Some of this work goes so far as to suggest that LLMs constitute better theories of human linguistic cognition than anything coming out of generative linguistics. Such reactions are misguided. The architectures behind current LLMs lack the distinction between competence and performance and between correctness and probability, two fundamental distinctions of human cognition. Moreover, these architectures fail to acquire key aspects of human linguistic knowledge and do nothing to weaken the argument from the poverty of the stimulus. Given that LLMs cannot reach or even adequately approximate human linguistic competence they of course cannot serve to explain this competence. These conclusions could have been (and in fact have been) predicted on the basis of discoveries in linguistics and broader cognitive science over half a century ago, but the exercise of revisiting these conclusions with current models is constructive: it points at ways in which insights from cognitive science might lead to artificial neural networks that learn better and are closer to human linguistic cognition.

 

References

Lan, N., Geyer, M., Chemla, E., and Katzir, R. (2022). Minimum description length recurrent neural networks. Transactions of the Association for Computational Linguistics, 10:785–799.

Fox, D. and Katzir, R. (2024). Large language models and theoretical linguistics. To appear in Theoretical Linguistics.

Lan, N., Chemla, E., and Katzir, R. (2024). Large language models and the argument from the poverty of the stimulus. To appear in Linguistic Inquiry.