Futrell and Mahowald (F&M) present themselves as offering a measured middle ground between the extreme claims that large language models (LLMs) obviate the need for linguistic theory and the (in their estimation) equally extreme claim that humans possess language-specific innate inductive bias. However, despite this posture, they align themselves aggressively with the former, without fully engaging the empirical discoveries that have propelled the latter forward. Here, we limit our discussion to the structure of filler-gap dependencies and their emergence during language development.

F&M present these models as the central path forward for linguistic theory. But despite their optimism, LLMs are simply not as good at language as they claim. LLMs appear to succeed at recognizing not only simple filler-gap dependencies but also the environments in which they cannot occur – syntactic islands (Wilcox et al., Reference Wilcox, Levy, Morita, Futrell, Linzen, Chrupała and Alishahi2018, Reference Wilcox, Futrell and Levy2024). Islands represent prototypical evidence for innate bias in generative linguistics because the patterns are semantically unpredictable and share structure only at a high level of abstraction (Boeckx, Reference Boeckx2008; Chomsky, Reference Chomsky, Culicover, Akmajian and Wasow1977; Ross, Reference Ross1967), so LLM success matters. However, while LLMs discover island sensitivity for embedded wh-movement (Wilcox et al., Reference Wilcox, Futrell and Levy2024), their behavior is less robust than that found in human speakers (Sprouse Reference Sprouse2007; Sprouse et al., Reference Sprouse, Caponigro, Greco and Cecchetto2016, i.a.). Yet despite these limitations, LLMs’ successes are touted as powerful evidence against nativist theorizing: islands do not require innate bias. This claim misses the broader observation that motivates islands’ centrality: islands do not just constrain wh-questions (1), but also surface-dissimilar dependencies such as clefting (2), topicalization (3), tough movement (4), and comparative VP-ellipsis (5), as well as in filler-gap dependencies across unrelated languages.

1. (1) * What did the engineers ignore the claim that humans need?

2. (2) * It is syntax that engineers ignore the claim that humans need.

3. (3) * Syntax, engineers ignore the claim that humans need.

4. (4) * Syntax is easy for the engineers to ignore the claim that humans need.

5. (5) * Engineers need less syntactic bias than I believe the claim that humans do.

Syntacticians posit a representational format for filler-gap dependencies that gives rise to parallel behavior across phenomena both within and across languages, including wh-agreement in Irish (McCloskey, Reference McCloskey2001), wh-agreement in Chamorro (Chung, Reference Chung, Gerdts, Moore and Polinsky2009), quantifier float in Irish-English (McCloskey, Reference McCloskey2000), scope islands in Mandarin (Huang, Reference Huang1982), comparatives in Greek (Merchant, Reference Merchant2009), the interaction of scope, binding, and movement in English (Fox, Reference Fox1999), and even in children’s nontarget productions in English (Liter, Grolla & Lidz, Reference Liter, Grolla and Lidz2022; Lutken, Legendre, & Omaki, Reference Lutken, Legendre and Omaki2020; Thornton, Reference Thornton1990), Brazilian Portuguese (Grolla, Reference Grolla, Gong and Kpogo2022), and French (Oiry & Demirdache, Reference Oiry, Demirdache, Torrens and Escobar2006).

LLMs instantiate an alternative: constraints on filler-gap dependencies can be extracted from the surface distributions of strings in the language. However, to cover the same range of phenomena as humans (i.e., (1)–(5)), they must either access enough data to arrive at consistent hypotheses individually for each construction, or recognize a fundamental similarity across constructions based on surface distributions, and then use the discovered island sensitivity in one construction to generalize broadly.

Whether human children see enough data to arrive at such hypotheses independently is an open empirical question, but LLMs, given magnitudes more input than human children, do not (Howitt et al., Reference Howitt, Nair, Dods, Hopkins, Barak and Alikhani2024; Ozaki et al., Reference Ozaki, Yurovsky, Levin, Ettinger, Hunter and Prickett2022). Further, LLMs extract contradictory hypotheses where evidence for one construction’s island sensitivity is taken as evidence against another’s. Howitt et al. (Reference Howitt, Nair, Dods, Hopkins, Barak and Alikhani2024) showed that training an LLM on grammatical instances of clefting (e.g., It was syntax that the humans needed) allowed the model to correctly posit islands for clefts, but also led them to lose island sensitivity in tough constructions (4). And while some work suggests that LLMs sometimes use the same representational networks to inform their representations of some filler-gap dependencies (Boguraev et al., Reference Boguraev, Potts, Mahowald, Christodoulopoulos, Chakraborty, Rose and Peng2025), these findings and, in fact, the methods that reveal them are dependent on the highly similar surface forms of this subset of constructions.

While F&M’s research program invites us to expand our search for inductive biases away from grammatical architecture, the support for this invitation is based on hope, not evidence. LLMs might capture surprising distributional facts in one language, but in the absence of architectural constraints, observed cross-linguistic and cross-constructional consistency would have to depend on an extralinguistic source. Critics of an architectural explanation point to communicative pressures as this extralinguistic source (Abeillé et al., Reference Abeillé, Hemforth, Winckel and Gibson2020; Cuneo & Goldberg, Reference Cuneo and Goldberg2023; Erteschik-Shir, Reference Erteschik-Shir1973; Goldberg, Reference Goldberg2006), but given the range of phenomena that display island sensitivity and the diversity of the uses to which these phenomena are put, such hypotheses fall short of the explanatory mark (Lidz & Williams, Reference Lidz and Williams2009; Cartner et al., Reference Cartner, Kogan, Webster, Wagers and Sichel2025).

Finally, F&M’s research program, like many classic works (Chomsky, Reference Chomsky1975; Gold, Reference Gold1967; Wexler & Culicover, Reference Wexler and Culicover1980), addresses learnability abstractly but provides little insight into how knowledge grows. We now know that local argument structure dependencies are acquired by 15–16 months (Fisher, Jin, & Scott, Reference Fisher, Jin and Scott2020; Lidz, White, & Baier, Reference Lidz, White and Baier2017; Perkins & Lidz, Reference Perkins and Lidz2019), a few months prior to nonlocal dependencies in wh-questions (Perkins & Lidz, Reference Perkins and Lidz2021). Explicit, interpretable models of language acquisition reveal (a) how learners could actively select a subset of data to learn from to arrive at argument structure representations first (Perkins, Feldman, & Lidz, Reference Perkins, Feldman and Lidz2022) and (b) that this knowledge is necessary to identify the form of filler-gap dependencies (Perkins, Feldman, & Lidz, Reference Perkins, Feldman and Lidz2025). These models of acquisition allow us to explain not just acquired knowledge but specific developmental steps that learners take to acquire that knowledge. LLMs, by their very nature, do not afford developmental analysis.

It is no surprise that an LLM trained on 450 million years of input has a pretty good idea about what is likely to occur in a language. But an 18-month-old with just 450 days of language experience also has a pretty good idea, and we understand both the steps they take to get there and the reasons those steps occur in the order that they do. The partial successes of LLMs are impressive. But our goal should not be to make these models better, but to make better models that respect the empirical landscape identified by theoretical and developmental linguists.