Futrell and Mahowald argue convincingly that LLMs can expand linguistics, opening the field to neglected questions and creating bridges across disciplines. Their focus is on how LLMs can (potentially) explain the acquisition and implementation of syntactic regularities – a traditional focus of linguistics, especially in the Chomskian tradition. But why stop at syntax? We suggest a bolder aim in the same general spirit. LLMs can illuminate perhaps the deepest enigma about language: what Chomsky called the creative aspect of language use (CALU) (Chomsky, Reference Chomsky2009).

CALU refers to speakers’ ability to freely select, from the vast array of grammatically possible sentences, utterances that are appropriate to the situation and linguistic context. This capacity, which transcends stimulus–response associations (“stimulus independence”), has long been central to accounts of human linguistic capacities (e.g., Descartes). Early in his career, Chomsky envisioned an intimate connection between generative linguistics and CALU (e.g., Chomsky, Reference Chomsky2011, p. 8, Reference Chomsky1965, p. 6), especially insofar as recursive syntax provides a foundation for the unboundedness of linguistic competence. But he later came to endorse a mysterian stance regarding CALU as a whole: “Honesty forces us to admit that we are as far today as Descartes was three centuries ago from understanding just what enables a human to speak in a way that is innovative, free from stimulus control, and also appropriate and coherent” (Chomsky, Reference Chomsky2006, p. 11). He added that CALU will likely always remain beyond the reach of scientific inquiry (Chomsky, Reference Chomsky1982, Reference Chomsky2006; see Reference DrachDrach, 1981, for further discussion).

LLMs pose a stark challenge to mysterianism. As human-built systems that generate fluent, contextually appropriate responses, LLMs instantiate many features of CALU. For example, LLMs exhibit humanlike levels of novelty (McCoy et al., Reference McCoy, Smolensky, Linzen, Gao and Celikyilmaz2023), pragmatic reasoning (Lipkin et al., Reference Lipkin, Wong, Grand and Tenenbaum2023), and stylistic adaptation (Kandra et al., Reference Kandra, Demberg and Koller2025), and humans cannot reliably distinguish human versus LLM conversation partners (Jones et al., Reference Jones, Rathi, Taylor and Bergen2025). Thus, LLMs are positioned to be the first tangible scientific models of key aspects of CALU, a promising step in moving beyond mysterianism toward mechanistic understanding.

Several LLM mechanisms are likely central to a future “science of CALU.” First, LLMs operate with massive feature spaces, with recent work suggesting that small models (17 billion parameters) encode millions of features in dense superposition (Templeton, Reference Templeton2024), and that large models with trillions of parameters likely contain dramatically more features. The vastness of LLM representational regimes allows encoding of countless shades of meaning and extraordinarily fine-grained distinctions among contexts, enabling the system to retrieve and combine just those specific features relevant to producing situationally apt utterances. Second, attention mechanisms selectively amplify subtle aspects of the context, supporting fine-grained sensitivity to discourse and ongoing temporally evolving inter-speaker dynamics. Third, autoregressive generation factors predictive distributions over extended sequences into tractable components (involving predictions of next tokens conditional on context), enabling the open-ended production of never-before-seen utterances and the emergence of genuine novelty. Fourth, reinforcement learning supports policy learning from iterative feedback, aligning language generation with situational goals and pragmatic appropriateness. Together, these mechanisms constitute an initial toolkit for a future science of CALU, and the next generation of work will elaborate them in richer detail and add other mechanisms.

Importantly, LLMs also clarify why earlier attempts to build computational models of language generation and use – whether in cognitive science or artificial intelligence – were destined to fall short of an adequate account of CALU. LLMs suggest that CALU is an emergent property of scale: it arises only when an architecture has access to extraordinarily rich representational resources and correspondingly powerful mechanisms for manipulating them (Sutton, Reference Sutton2019). By contrast, prior approaches to modeling language use – e.g., assistive dialog systems or question-answering systems – were heavily engineered and tailored for very limited domains, and relied on correspondingly limited representational resources. The gap between even the best of these pre-LLM models and human-level CALU was so large that Chomsky’s mysterianism remained an appealing stance. LLMs embody a sharply contrasting approach: the ability to flexibly tailor linguistic production to nuanced contexts requires vast numbers of features and the capacity to deploy them through similarly vast numbers of tightly interlocked operations. LLMs provide, for the first time, a viable account of CALU as an emergent ability of high-dimensional, high-capacity models – a decisive shift in the explanatory paradigm.

To be sure, there are also critical divergences between human CALU and the analog observed in LLMs, but they need not be viewed as decisive impediments to the overall approach. Instead, these divergences represent promising research directions within a new framework. For example, LLMs often exhibit pragmatic brittleness, such as difficulty with irony, indirect speech acts, or norm-governed conversational implicatures (Hu et al., Reference Hu, Floyd, Jouravlev, Fedorenko and Gibson2023). Each of these divergences highlights a concrete locus where mechanistic hypotheses can be sharpened, and understandings of new, heretofore poorly appreciated mechanisms might emerge. And of course, the most impressive capacities for CALU reside in models trained on internet-scale data. Rigorous poverty-of-stimulus arguments might one day motivate specialized innate mechanisms behind CALU in humans. But we are only in the earliest stages of understanding what is possible in building models with more limited data, and more importantly, what is possible with data derived from grounded, agentic experience (Vong et al., Reference Vong, Wang, Orhan and Lake2024).

One reason generative linguistics has historically focused narrowly on grammatical competence is that questions like CALU seemed intractable to formal analysis; indeed, Chomsky began sketching arguments along these lines as early as his 1959 review of Skinner’s Verbal Behavior. But LLMs shift the boundary of tractability. They supply not only successful demonstrations of CALU-like behavior but also concrete mechanistic hypotheses for how such behavior might arise. By treating LLMs as artificial model systems, linguistics gains powerful tools to study the “inner springs” of creative language use at levels once dismissed as scientifically inaccessible.