Futrell and Mahowald argue that because language models (LMs) can learn linguistic structure from text input without explicitly encoded, language-specific inductive biases, we need not suppose children have such biases either. This argument requires that LMs learn linguistic structure in the same way as children; the only difference Futrell and Mahowald consider is how much data each needs to reach mature competence (see also Frank, Reference Frank2023). Yet LMs and children also systematically diverge in how they acquire language, and even in what it is they acquire. If LMs cannot eventually model these aspects of acquisition, language-specific learning mechanisms may still be needed to explain them.

Language emergence. When children receive sparse linguistic input, they spontaneously generate linguistic structure not present in that input. Deaf children not exposed to sign language develop homesign systems whose structure and expressiveness outstrip the gestures of their hearing families. Homesign systems morphologically or syntactically distinguish parts of speech (Abner et al., Reference Abner, Flaherty, Stangl, Coppola, Brentari and Goldin-Meadow2019; Goldin-Meadow et al., Reference Goldin-Meadow, Butcher, Mylander and Dodge1994) and thematic roles (Goldin-Meadow & Mylander, Reference Goldin-Meadow and Mylander1998) and structurally mark grammatical subjects (Coppola & Newport, Reference Coppola and Newport2005) – distinctions absent or much reduced in homesigners’ hearing families’ gestures.

Further, when children’s input is limited to a rudimentary linguistic system like homesign, they quickly expand and regularize it. Famously, the conventionalized sign system generated de novo by children attending a Nicaraguan school for the deaf was elaborated by successive student cohorts, each imposing additional structure absent from their input. For instance, the second cohort, but not the first, uses spatial modulation to convey shared reference (Senghas & Coppola, Reference Senghas and Coppola2001; Senghas, Reference Senghas2003). Over a few cohorts, Nicaraguan Sign Language has come to exhibit discrete, compositional morphology (Senghas, Kita, & Özyürek, Reference Senghas, Kita and Özyürek2004) and recursion (Kocab et al., Reference Kocab, Senghas, Coppola and Snedeker2023). Much of the structure of emerging languages originates from the child, not their input.

In contrast, LMs learn only to recapitulate the statistical properties of their training data. To account for language emergence – that is, the addition of linguistic structure that is absent from the input – LMs would need to violate those statistical properties. If that is possible, LMs would then need to more specifically enrich the structure of a language in the ways children do when placed under similar communicative pressures. Although simple neural agents can create rudimentary compositional systems in cooperative reference games (Boldt & Mortensen, Reference Boldt and Mortensen2024; Lazaridou & Baroni, Reference Lazaridou and Baroni2020; Steinert-Threlkeld, Reference Steinert-Threlkeld2020), it is uncertain whether and under which conditions LMs could create systems with the rich structural complexity that emerges in the ad hoc communication system of even a single child.

Learning trajectories. Children’s word-learning trajectories exhibit characteristic patterns distinct from the frequency of words in their input. Concrete nouns are overrepresented at first, then verbs and adjectives, and only later do closed-class function words like “of” and “the” follow – though these are the most frequent words in the input (Frank et al., Reference Frank, Braginsky, Yurovsky and Marchman2016). Further, even the most basic syntactic productivity (e.g., combining two words) emerges only after a period when children communicate meaningfully using early words in isolation (Fenson et al., Reference Fenson, Marchman, Thal, Dale, Reznick and Bates2007). LMs show the opposite pattern: they produce correct complex multi-word syntax early in training, but semantically coherent productions only much later (Müller-Eberstein et al., Reference Müller-Eberstein, Van Der Goot, Plank and Titov2023; Xia et al., Reference Xia, Artetxe, Zhou, Lin, Pasunuru, Chen, Zettlemoyer and Stoyanov2023). And unlike children, LMs’ early productions reflect token input frequencies. “The,” periods, and commas dominate (Chang, Tu, & Bergen, Reference Chang, Tu and Bergen2024).

These learning trajectories reflect fundamental differences in how children and LMs use language. LMs learn patterns of co-occurrence between constituents of their text-based input. Children must additionally learn how words refer to the world. This is why children first learn nouns for concrete objects: those are the easiest to identify as referents in their environment (Gillette et al., Reference Gillette, Gleitman, Gleitman and Lederer1999; Snedeker, Geren, & Shafto, Reference Snedeker, Geren and Shafto2007). This gap might be bridged by multimodal LMs that learn both word-word and word-percept relations. Yet early attempts show that such models have severe limitations in even the simplest referential abilities, such as generalizing a learned label-object pairing to new objects of the same shape (Vong et al., Reference Vong, Wang, Orhan and Lake2024). While more training data may help, these limitations may reflect deeper differences in how LMs and children acquire the concepts underlying word meanings. Just as evidence from language emergence shows that children add morphosyntactic structure absent from their linguistic input, scholars from Kant to Spelke to Gleitman have adduced evidence that children’s preverbal conceptual and perceptual capacities structure thought beyond the combined associations of words with other words and with images. To the extent that learning language is the process of learning the expression of thought, LMs would need to model humans’ capacities for thought to fully model their language.

Critical periods. While general statistical learning abilities improve throughout development (Shufaniya & Arnon, Reference Shufaniya and Arnon2018), children’s ability to learn the grammar of their language declines around puberty (Hartshorne, Tenenbaum, & Pinker, Reference Hartshorne, Tenenbaum and Pinker2018; Johnson & Newport, Reference Johnson and Newport1989; Lenneberg, Reference Lenneberg1967), and the ability to learn native phonology declines in infancy (see Werker, Reference Werker2024, for review). This dissociation implies that learning language does not rely solely on general statistical mechanisms.

LMs might account for critical periods by demonstrating that declines in language learning ability can arise from language-specific experience that would not impede general statistical learning. However, recent work has found the opposite: learning a first language accelerates the LM’s rate of subsequently learning a second (Oba et al., Reference Oba, Kuribayashi, Ouchi and Watanabe2023). In fact, the way that a critical period can be induced in LMs is by imposing a regularizer partway through training – a process more analogous to a maturational decrease in language-specific plasticity than to any effect of experience (Constantinescu et al., Reference Constantinescu, Pimentel, Cotterell and Warstadt2025). By contrast, neural networks in other domains do show spontaneous critical periods. For instance, deep multisensor perceptual networks show a critical period during which improperly correlated data can permanently impair the model’s learning (Kleinman, Achille, & Soatto, Reference Kleinman, Achille and Soatto2023). Thus, while neural networks may be able to account for critical periods in some areas of development based on experience alone, the way LMs have been able to account for them in language specifically favors a maturational, language-specific explanation.