Do language models model the language(s) of humans, and do they have anything to contribute to linguistic theory? This question has been debated by many (cf. Piantadosi, Reference Piantadosi, Gibson and Poliak2024 and its numerous replies), and Futrell and Mahowald offer one of the most optimistic answers: yes, language models do learn non-trivial linguistic structure, and linguists should therefore recognize their value. An essential part of their argument draws on statistical and usage-based traditions in linguistics, which highlight that human language learning relies heavily on distributional patterns. On this view, grammar in the human mind is best understood not as a system of purely symbolic rules, but as a repertoire of structures shaped by probabilities and frequencies. Put simply, Futrell and Mahowald take modern LMs as vindicating these “statistical approaches” to language, since such models learn language strikingly well from raw text alone, without being endowed with strong inductive biases or constraints. However, in emphasizing how language models serve as a proof of concept for many of their core claims, Futrell and Mahowald paint a reductionist picture of usage-based approaches. It is certainly true that current (L)LMs are compatible with usage-based ideas, and we concur with their idea that LMs may point to “real patterns” in the sense of Dennett (Reference Dennett1989) that can also prove insightful for usage-based linguistics. Yet, at the same time, they entirely lack the socio-cognitive grounding that usage-based linguistics considers constitutive of language. Therefore, the implications of current language models for usage-based linguistics are severely limited, despite widespread claims to the contrary (e.g., Ambridge & Blything, Reference Ambridge and Blything2024, Goldberg, Reference Goldberg2024, Piantadosi, Reference Piantadosi, Gibson and Poliak2024).

Like human learners, LMs excel at finding patterns in data and learn to produce fluent language through domain-general statistical mechanisms. But Futrell and Mahowald’s focus on frequencies and distributions sidelines much of what usage-based linguists are actually interested in: the ways in which cognitive, social, and interactive pressures shape the linguistic system and how these factors enable humans to use language as fluently as they do. Language models learn and process language input unlike humans. They learn from statistical associations in corpora, whereas humans learn from associations of linguistic forms and their meanings in social contexts and their lived experience. From a usage-based point of view, learning means “constructing new representations via interaction with environmental stimuli” (Rowland et al., Reference Rowland, Westermann, Theakston, Pine, Monaghan and Lieven2025). This learning process rests on a number of domain-general cognitive capacities, including the registration of frequencies and transitional probabilities. But crucially, it also relies on social-interactional environmental contexts (Enfield, Reference Enfield and Dancygier2017) to provide the mapping of meaning to form. Intention reading and, in many contexts, joint attention are crucial to identifying the meaning of linguistic forms and to learning their mappings (Ibbotson, Reference Ibbotson2011; Tomasello, Reference Tomasello2003).

Hence, for human language, the relevance of interaction, grounding in the real world, and the functional dimension of language can hardly be underestimated. Putting these processes into the center of linguistic inquiry is a core tenet of usage-based linguistics, and certain statistical patterns might only emerge as side effects of these processes. However, they are not represented in the lifecycle of current, widely used language models, which are typically trained on gigantic corpora, much larger than the linguistic input that humans typically receive over their lifespan (cf. Brysbaert et al., Reference Brysbaert, Stevens, Mandera and Keuleers2016). Initial pre-training on unstructured data is complemented by mid-training on reasoning data as well as post-training/preference tuning on task-oriented dialogs. Moreover, current evaluation methods probing the linguistic capabilities of language models remain anchored in formalizations of language perception. For example, the BabyLM test suite (Charpentier et al., Reference Charpentier, Choshen, Cotterell, Gul, Hu, Liu, Jumelet, Linzen, Mueller, Ross, Shah, Warstadt, Wilcox and Williams2025) contains, inter alia, minimal pair tests where preference for supposedly “grammatical” sentences is measured through string probabilities (BLiMP, Warstadt et al., Reference Warstadt, Parrish, Liu, Mohananey, Peng, Wang and Bowman2020), and correlation tests for word-level surprisal and human acquisition norms (Chang & Bergen, Reference Chang and Bergen2022). These tests correspond to deliberate experimental interventions in humans, not their natural language use. Recent benchmarks embrace fewer binary assumptions of language structure, for example, by comparing language models’ preferences for different nonce word inflections with instances of human usage (Hofmann et al., Reference Hofmann, Weissweiler, Mortensen, Schütze and Pierrehumbert2025). Still, these tests detach human language from its socio-cognitive context. Evaluation methods that test for functional aspects of language, like social and real-world knowledge (e.g., SiQA, Sap et al., Reference Sap, Rashkin, Chen, Le Bras and Choi2019; EWoK, Ivanova et al., Reference Ivanova, Sathe, Lipkin, Kumar, Radkani, Clark, Kauf, Hu, Pramod, Grand, Paulun, Ryskina, Akyürek, Wilcox, Rashid, Choshen, Levy, Fedorenko, Tenenbaum and Andreas2025), necessarily cast them as text-based problems, on which language models regularly underperform (Mahowald et al., Reference Mahowald, Ivanova, Blank, Kanwisher, Tenenbaum and Fedorenko2024).

In fact, even the very notion of language (use) is fundamentally different between language models and humans. For a language model, language is a probability distribution over sets of strings. In usage-based linguistics, using language is a way of acting in the social world, something that language models cannot do. A nice example is provided by Engelmann et al. (Reference Engelmann, Granlund, Kolak, Szreder, Ambridge, Pine, Theakston and Lieven2019), who compared the learning of person-number agreement by Finnish and Polish children with that of a computational model. While the order of acquisition was similar for both, there was one exception. Despite the very high frequency of the second-person singular in the input, children produced the self-referring first-person singular earlier and more accurately, while, by contrast, the model simply reflected the frequencies in the input and produced the second-person singular well before the first-person singular.

In light of these crucial differences between language (use) in humans and language models, we argue that (L)LMs remain poor objects of linguistic inquiry from a usage-based perspective. As of now, they cannot model “the real thing” (p. 13), simply because the real thing is not just a (highly apt) text-generation mechanism, but a complex, embodied, context-sensitive process shaped by interaction and social experiences. This problem has been partially recognized by the functionally inspired language modeling community. There, reinforcement learning is being applied with more developmentally plausible objectives, such as communicative success in interaction (e.g., Mayer Martins et al., Reference Mayer Martins, Bashir, Khalid and Beinborn2025; Padovani et al., Reference Padovani, Bunzeck, Ali, Momen, Bisazza, Buschmeier and Zarrieß2025; Salhan et al., Reference Salhan, Gu, Rooein, Galvan-Sosa, Gaudeau, Caines, Yuan and Buttery2025). Still, most of these experiments show that optimization for such goals leads to deteriorating scores on linguistic benchmarks and impairs text generation capabilities. The aim of modeling more realistic language use seems to be broadly at odds with current LM architectures, and it remains unclear if (and how) these aims could be reconciled.