The speed with which infants learn languages is a core challenge for models of human language acquisition. Chomsky’s hypothesis is that humans have an inductive bias, a “Universal Grammar” (UG), that not only constrains the linguistic structures of all attested languages, but also allows children to learn quickly. By contrast, language models (LMs), such as ChatGPT, lack human-like priors, and consequently, they display just the opposite profile, namely, they learn “impossible languages” that violate UG just as easily as attested ones (Mitchell & Bowers, Reference Mitchell and Bowers2020), and at the same time, require training on many orders of magnitude of more language data compared to children (Bowers, Reference Bowers2025a). According to Chomsky, the ease with which LMs learn impossible languages undermines LMs as theories of human languages. For instance, Chomsky,Roberts, and Watumul (Reference Chomsky, Roberts and Watumull2023) write:

Since Mitchell and Bowers (Reference Mitchell and Bowers2020), there have been several studies that have compared how easily LMs learn attested and impossible languages. For instance, Kallini et al. (Reference Kallini, Papadimitriou, Futrell, Mahowald and Potts2024) reported two impossible languages that were more difficult to learn than English and concluded that LMs (and humans) do not need any linguistic priors to account for the possible-impossible learning gap. F&M describe this study and cite several others (including Mitchell and Bowers) in a way that suggests that LMs consistently show a human-like difficulty in learning impossible languages, writing:

But this characterization of the research is mistaken. First, consider the Kallini et al.’s study. Although the authors reported two impossible languages that were more difficult to learn, they reported that several additional impossible languages were learned almost as easily as English, including one of the impossible languages that Mitchell and Bowers (Reference Mitchell and Bowers2020) devised. Furthermore, the impossible language that was the most difficult to learn was composed of random shuffles of words. That is, there was no structure to learn. The other difficult-to-learn impossible language was composed of deterministic random shuffles, with sentences of different lengths shuffled in different deterministic ways (e.g., all sequences of length 5 were shuffled in one order, sequences of length 6 were shuffled in another order, etc.). In this case, there was something to learn, namely, different random languages for the different sentence lengths. The observation that it is more difficult to learn multiple languages compared to a single random language does not challenge Chomsky’s point. More generally, a learning algorithm that performs well on one data type must perform badly on others (“no free lunch theorems”; Wolpert & Macready, Reference Wolpert and Macready2002). Accordingly, the observation that LMs perform badly on some languages is an unnecessary empirical confirmation of these theorems. In contrast, the success in learning some impossible languages is a falsification of the claim that LMs are an adequate model of human language learning.

The same issue applies to Yang et al. (Reference Yang, Aoyama, Yao and Wilcox2025) article that F&M cited. The authors again assessed LMs on a range of impossible languages, including the same deterministic shuffled languages used by Kallini et al. And again, the authors found that several impossible languages were easily learned, whereas the random shuffle language was difficult. And again, the slow learning of the random shuffled language was taken as problematic for Chomsky:

But again, the observation that it is more difficult to learn multiple different impossible languages (for different sentence lengths) is not surprising on any theory.

The Xu et al. (Reference Xu, Kuribayashi, Oseki, Cotterell and Warstadt2025) article that F&M cited varied the plausibility of languages (rather than the impossibility of languages). In this case, the authors found that LMs struggled with some implausible languages, but again, in some cases, the LM found the implausible languages easy to learn. And with regard to conditions in which LMs struggled, the authors note that there may have been errors in the construction of the materials that weaken the conclusions that can be drawn:

It is difficult to take these findings as problematic for Chomsky’s thesis.

Finally, the Ziv et al. (Reference Ziv, Lan, Chemla and Katzir2025) article that F&M cited reports several impossible languages that LMs learn just fine, including partially reversed languages (replicating Mitchell & Bowers, Reference Mitchell and Bowers2020). In addition, in another study not cited, Lou, Ramscar, and Love (2024) report that LMs can happily learn reversed languages.

So, contrary to the claims of F&M, LMs often learn impossible languages as easily as attested languages. At the same time, LMs require many orders of magnitude more training than infants (e.g., Bowers, Reference Bowers2025a), and a recent attempt to induce a universal grammar in LMs (McCoy & Griffiths, Reference McCoy and Griffiths2025) does not make LMs much more data efficient (Bowers, Reference Bowers2025b). Given all this, the appropriate conclusion is that current LMs learn in just the way one would expect of models that are missing human-like innate biases for language learning.