Large Language Models are useless for linguistics, as they are probabilistic models that require a vast amount of data to analyze externalized strings of words. In contrast, human language is underpinned by a mind-internal computational system that recursively generates hierarchical thought structures. The language system grows with minimal external input and can readily distinguish between real language and impossible languages.

According to F&M, both infants and language models (LMs) find attested languages easier to learn than “impossible languages” that have unnatural structures. We review the literature and show that LMs often learn attested and many impossible languages equally well. Difficult to learn impossible languages are simply more complex (or random). LMs are missing human inductive biases that support language acquisition.

Language models excel at finding patterns in linguistic data, and can therefore prove insightful for statistical approaches to linguistics in that they provide further evidence for the strong reliance of natural languages on recurrent, fixed patterns. Nevertheless, regarding actual usage-based language processing, their implications are severely limited as they lack a crucial aspect of language use: interaction.

Language models (LMs) call for a theoretical rethinking grounded within linguistics itself. Rather than signalling the “end of linguistics” or merely encouraging interdisciplinarity, LMs function as an empirical testing ground for formal linguistic concepts. They prompt a renewed examination of form–meaning mapping, theoretical autonomy, and the conditions under which computational systems can genuinely inform linguistic explanation.

In this commentary, we challenge the idea that Language Models (LMs) provide explanatorily adequate models of human language. Findings from language evolution show that both humans and LMs fail to produce human-like language via inductive biases alone; the communicative function of language is crucial. More generally, both experimental and computational work underscore the fact that language is more than just incremental prediction.

Futrell and Mahowald present a useful framework bridging technology-oriented deep learning systems and explanation-oriented linguistic theories. Unfortunately, the target article’s focus on generative text-based Large Language Models (LLMs) fundamentally limits fruitful interactions with linguistics, as many interesting questions on human language fall outside what is captured by written text. We argue that audio-based deep learning models can and should play a crucial role.

Large language models are interesting, but linguistics and cognitive science should be cautious about centering any new technology as a magic bullet. Doing so reinforces the historically “narrow focus” of linguistics identified in the target article, rather than expanding our understanding of human language. We argue that centering values instead of technology is the best way to future-proof scholarly work and to keep finding out new things about language.

By “linguistics,” the target article means usage-based linguistics, which, we agree, plays very well with language models. But the article summarily dismisses important work on meaning from the now disreputable tradition of generative linguistics. We refute the authors’ arguments from distributional semantics and “green tea,” and highlight the importance of formal compositional semantics for the study of thought.

The commentary argues the authors employ misdirection and strawmanning to cast others as polarized extremes and themselves as the reasonable centrists. We argue that these patterns of misrepresentation ultimately damage any consensus and middle ground they claim to hope to reach.

Futrell and Mahowald argue that the success of large language models should move the field away from the formal structures of generative linguistic theory. The limited success of these models falls short of formal linguistic theory in explaining both the character of human languages and understanding the trajectory of child language acquisition.

We argue that language models (LMs) have strong potential as investigative tools for probing the distinction between possible and impossible natural languages and thus uncovering the inductive biases that support human language learning. We outline a phased research program in which LM architectures are iteratively refined to better discriminate between possible and impossible languages, supporting linking hypotheses to human cognition.

The diversity of variation captured in data can strongly affect the generalization of a learning system – even when that variation occurs along axes orthogonal to the generalization in question. Thus, I argue that data richness both distinguishes current language models from prior linguistic models and may still underlie their remaining linguistic data inefficiency.

Acknowledging that large language models have learned to use language can open doors to breakthrough language science. Achieving these breakthroughs may require abandoning some long-held ideas about how language knowledge is evaluated and reckoning with the difficult fact that we have entered a post-Turing test era.

Futrell and Mahowald frame the success of neural language models (LMs) as supporting gradient, usage-based linguistic theories. I argue that LMs can also instantiate theories based on formal structures – the types of theories seen in the generative tradition. This argument expands the space of theories that can be tested with LMs, potentially enabling reconciliations between usage-based and generative accounts.

Futrell and Mahowald argue language model (LM) success suggests humans may learn language entirely through domain-general statistical mechanisms. However, children differ crucially from LMs in their ability to surpass their input, their language learning trajectory, and the presence of a critical period. Until LMs account for these phenomena, it remains possible that human language acquisition is supported by innate, language-specific learning mechanisms.

Futrell & Mahowald argue that language models (LMs) discover linguistic structure as “real patterns.” I contend this framing underplays what mechanistic interpretability uncovers: LMs implement specific algorithms to solve linguistic tasks. Under the framework of causal abstraction, we can rigorously test whether LMs converge on algorithms posited by linguistic theory, which further supports the authors’ conciliatory proposal.

Futrell and Mahowald argue that neural networks have learned non-trivial aspects of language. We argue that these systems have not in fact demonstrated “mastery” of syntax, marshalling recent evidence, and that they further obscure explanatory insights with respect to topics in the cognitive neuroscience of language.

Under the lens of Marr’s (levels of analysis, we critique and extend the authors’ two points about language models (LMs) and language processing: first, predicting upcoming linguistic information based on context is key to language processing, and second, that many advances in psycholinguistics would be impossible without LLMs. We also outline directions combining LLMs’ strengths with psycholinguistic models.

In this response, we outline the central idea of a real pattern as applied in the philosophy of science as well as its relationship to the modal structure of models. We suggest that there may be additional concerns with linguistically real patterns in terms of convergence between natural language processing and formal linguistic theory.

Futrell and Mahowald claim language models (LMs) “serve as model systems,” but an assessment at each of Marr’s three levels suggests the claim is clearly not true at the implementation level, poorly motivated at the algorithmic-representational level, and problematic at the computational theory level. LMs are good candidates as tools; calling them cognitive models overstates the case and unnecessarily feeds large language model hype.