Let me quickly contextualise this commentary by saying I’m no LLM booster. On the contrary, I see LLMs as a disaster for both our intellectual environment (e.g., Guest et al., Reference Guest, Suarez, Müller, van Meerkerk, Beverborg, de Haan, Reyes Elizondo, Blokpoel, Scharfenberg, Kleinherenbrink, Camerino, Woensdregt, Monett, Brown, Avraamidou, Alenda-Demoutiez, Hermans and van Rooij2025) and our environment (e.g., Luccioni, Viguier & Ligozat, Reference Luccioni, Viguier and Ligozat2023). Either AI fulfils its promise, and we’re all out of a job, or it doesn’t, and we’ll have burned academia, the global economy and a tonne of fossil fuels for nothing. If I could wave a magic wand and uninvent LLMs, I would.

But while LLMs are (still?) here, we may as well learn from them. And even as a committed non-fan, it seems to me that LLMs have already changed the game for child language acquisition research.

How so? Even today, my impression of the field is that most child language acquisition researchers are essentially essentialists. They see the child’s task as – in broad brush terms – acquiring abstract categories like VERB, NOUN, and DETMINER, and rules for combining them into sentences. This assumption is, of course, explicit in Chomskyan accounts of language development, which assume that these abstract categories are acquired very early, or (whisper it) perhaps even innate. But it is surprisingly common to see constructivist accounts of how particular abstract categories or constructions are “acquired,” and surprisingly rare to see ones that eschew these types of freestanding essentialist categories altogether.

Why? Humans are natural essentialists (e.g., Gelman, Reference Gelman2004; Hull, Reference Hull1976; Mayr, Reference Mayr1982). In a confusing world, it’s comforting to think in terms of categories with sharp boundaries – black or white; lion or tiger; good or evil – and when it turns out, as it always does, that “it’s actually a lot more complicated than that,” lots of people get very upset. Linguists and language acquisition researchers are by no means immune (e.g., Croft, Reference Croft, Arka, Barth and Bergqvistin press). It would be nice if we could understand language development as the acquisition of simple rules for manipulating discrete categories. And, as usual, attempts to say “it’s actually a lot more complicated than that” – i.e., it’s some kind of very complex functionally-based, input-based analogical generalisation – have historically been waved away (Adger, Reference Adger2020; Goodluck, Reference Goodluck2020; Hartshorne, Reference Hartshorne2020; Koring et al., Reference Koring, Giblin, Thornton and Crain2020; Pearl, Reference Pearl2014; Pinker & Ullman, Reference Pinker and Ullman2002), or even mocked (e.g., Pinker, Reference Pinker1979):

“Come on then, ” these papers said, “If it’s possible to learn grammar without rules and categories, show me a system that does so.” LLMs change the game for child language acquisition, then, because they do just this. They’re an existence proof that a system that doesn’t start out with abstract grammatical rules and categories – or even build them, except in the most distributed, emergent sense – can acquire abstract grammar. Yes, LLMs have all kinds of problems – they don’t have meanings grounded in the real world, they require implausible amounts of input, etc. (e.g., Frank, Reference Frank2023); but that doesn’t matter: they’re an existence proof. As many of us have been saying all along (e.g., Ambridge & Lieven, Reference Ambridge, Lieven, MacWhinney and O’Grady2015; Ambridge, Reference Ambridge2020a,b), you can learn grammar without abstract rules and categories; the question is no longer “whether” but “how.”

So that’s how LLMs change the game for child language acquisition research; what about the “vice versa”? This, I must admit, is more a pipe dream than a reality. But is it too vain to hope that we might be able to reduce both AI slop and its environmental impact by building into LLM-development insights from language-development research?

Children’s language learning is multimodal and embodied, containing rich real-world semantic representations (e.g., Reggin et al., Reference Reggin, Franco, Horchak, Labrecque, Lana, Rio and Vigliocco2023). Communicative function is key: children use language to get across what they want to say (e.g., Bannard, Rosner & Matthews, Reference Bannard, Rosner and Matthews2017). Children hear input forms not at random, but in bursts of identical/similar forms (e.g., Cychosz et al., Reference Cychosz, Romeo, Edwards and Newman2025). They are subject to developmentally decreasing memory and processing limitations that narrow the “context window” (Gathercole et al., Reference Gathercole, Pickering, Ambridge and Waring2004; Pine, Freudenthal, & Gobet, Reference Pine, Freudenthal and Gobet2023), and result in nonlinearities in learning (e.g., Brooks & Meltzoff, Reference Brooks and Meltzoff2008). They are overly sensitive to high-frequency chunks, in ways that yield characteristic errors (Ambridge et al., Reference Ambridge, Rowland, Theakston and Kidd2015). They learn, at least at first, not from tokenised text but from sounds better approximated as vectors of phonological features (Albright & Hayes, Reference Albright and Hayes2003; Rubehn et al., Reference Rubehn, Nieder, Forkel and List2024).

For these reasons, LLMs ignore developmental child language research at their peril, and vice versa. Disruption is mutually assured.