It is instructive to compare Futrell and Mahowald’s well-reasoned and empirically grounded target article with counterpoints from skeptics demanding that large language models (LLMs) conform to theories of generative linguistics to be judged as good models of language (Bolhuis et al., Reference Bolhuis, Crain, Fong and Moro2024; Murphy et al., Reference Murphy, Leivada, Dentella, Montero, Günther and Marcus2025), or who insist that LLMs are “stochastic parrots” that “only manipulate the form of language, with neither understanding nor communicative intent” (Bender & Hanna, Reference Bender and Hanna2025). Gauging the usefulness of LLMs as models of language based on how well they adhere to a particular theory of language is not only self-serving and unscientific but also risks further isolating many linguists from the broader language sciences. And reducing LLMs to “stochastic parrots” is an insult to people’s intelligence. Are we to believe that the 800M weekly users of ChatGPT are credulous dolts unable to distinguish competent language use from mere parroting? Or is it more likely that a decades-long scientific program of learning language through prediction (Elman, Reference Elman1990) was on to something – that learning language from data is not only possible but actually works?

By “actually works”, I mean that a general-purpose neural network endowed with a simple learning rule to lower prediction error, an architectural capacity to track long-distance associations, and a corpus of language use can learn language. Some continue to deny the premise, arguing that the failure of LLMs to “Draw [a] tree structure, in line with Minimalist syntax, for the sentence ‘The doctor the nurse the hospital had hired met John’” means that LLMs lack “fundamental principles of linguistic structure” (Murphy et al., Reference Murphy, Leivada, Dentella, Montero, Günther and Marcus2025). This argument is absurd. My 5-year-old is a highly competent language user who cannot draw a parse tree or tell you what makes a sentence ungrammatical. What counts is use! If a language model can use language proficiently enough to pass the Turing test (Jones & Bergen, Reference Jones and Bergen2025) while failing to conform to minimalist syntax, so much the worse for minimalist syntax (Hu et al., Reference Hu, Mahowald, Lupyan, Ivanova and Levy2024).

The ability to judge sentences like “Colorless green ideas sleep furiously” to be grammatical despite being nonsensical has played an outsized role in generative linguistics. These metalinguistic judgments are often used to test competing theories (Schütze, Reference Schütze2016) and have often been used to evaluate LLMs (e.g., Dentella, Günther, & Leivada, Reference Dentella, Günther and Leivada2023). That children engage in sophisticated language use while lacking the ability to make such explicit judgments is prima facie evidence that being a competent language user does not require being able to dissociate conceptual content from structural configurations. It is also true, however, that many adults without linguistic training can do this. But so can LLMs. As we previously showed, ChatGPT-4 is better than people at judging the normative grammaticality of sentences (Hu et al., Reference Hu, Mahowald, Lupyan, Ivanova and Levy2024). In recent informal tests, we also found that GPT-5 can generate meaningless grammatical sentences as well as adults. Figure 1 shows meaningfulness ratings of sentences generated by people and GPT-5 (t < 1), as well as people’s attempts to guess whether each sentence was generated by another person or an LLM (t < 1).

Figure 1.

A. Results from an informal experiment in which participants were tasked with generating maximally meaningless grammatical sentences. Each participant then judged the meaningfulness of the sentences generated by other participants and ChatGPT-5 and whether each sentence was generated by a person or an LLM. (A) Mean meaningfulness ratings with lines joining each participant’s mean ratings. (B) Accuracy in judging whether each sentence was generated by a person or LLM. An example of similarly meaningless LLM and human sentences: “Let velvet measure tomorrow with raisins” and “Burpy trees monkey in mountain poodle”.

Figure 1 Long description

Two graphs depict the results of an informal experiment. Panel A: A violin plot shows the distribution of meaningfulness ratings for sentences generated by humans and a large language model (LLM). The y-axis represents meaningfulness on a scale from 1 to 5, and the x-axis distinguishes between human and LLM-generated sentences. Lines connect the mean ratings for each participant across both sentence types. Panel B: A violin plot displays the accuracy of participants in judging whether sentences were generated by a human or an LLM. The y-axis represents accuracy in terms of d-prime (d’), a measure of sensitivity, and the x-axis indicates whether the sentence was produced by an LLM. Green dots represent individual data points, and a black dot with error bars shows the mean and standard error.

An intriguing question is when the ability to distinguish meaningfulness from grammaticality emerges in LLMs and whether this ability has any causal bearing on proficient language use. Although grammaticality can be read out of base models (LLMs trained only through self-supervision) using surprisal, base models appear incapable of making explicit grammaticality judgments (Hu et al., Reference Hu, Mahowald, Lupyan, Ivanova and Levy2024). In our preliminary tests, we find that base models can generate implausible sentences (e.g., A cat barks at the dog) but struggle to generate ungrammatical or truly meaningless ones. The ability to explicitly uncouple meaningfulness from grammaticality appears to emerge with fine-tuning on even small datasets (e.g., Databricks-dolly-15k, 2023) which do not contain any examples of grammaticality judgments. What makes this intriguing is that despite their enormous latent knowledge of both language and the world, base models make lousy conversationalists (the public couldn’t care less about GPT-3 but was blown away by its fine-tuned version released as ChatGPT). Understanding whether the ability to make explicit grammatical judgments is causally related to being able to use language would be a fascinating development.

It seems that over the decades, some linguists have lost the plot (see Knight, Reference Knight2016 for an engaging intellectual history). The goal of a language user is not to generate parse trees that conform to minimalist syntax. Rather, it is to do things with language. This includes inferring people’s mental states from what they say, communicating goals to enable collaboration, and convincing them to act in your interests. To the extent that such language use requires learning specific aspects of language structure, we should find these by appropriately interrogating LLMs. For example, it seems clear that representing the hierarchical structure – in some form – is critical to parsing language, and converging evidence appears to show that in learning language, LLMs learn hierarchical structure (e.g., Goldstein et al., Reference Goldstein, Ham, Schain, Nastase, Aubrey, Zada, Grinstein-Dabush, Gazula, Feder, Doyle, Devore, Dugan, Friedman, Brenner, Hassidim, Matias, Devinsky, Siegelman, Flinker, Levy, Reichart and Hasson2025; Liu, Xiang, & Ding, Reference Liu, Xiang and Ding2026). LLMs provide an unprecedented opportunity to understand that the aspects of linguistic structure are critical for language use, even when learned by a system with an architecture radically different from ours.

In his 1990 short story Bears Discover Fire, Terry Bisson imagines a world just like ours, but in which, inexplicably, bears have discovered fire: “They don’t hibernate anymore … They make a fire and keep it going all winter” (Bisson, Reference Bisson1990). One reaction to such a development would be to insist that bears are just not the kind of animal that can control fire; that it’s a trick; that their fire must not be real fire. Another would be to wonder if perhaps we were mistaken in our understanding of what bears are capable of, or in our understanding of what it takes to control fire. We have now entered a post-Turing test era where the torch of language is no longer unique to our species. What we make of this new reality is up to us.