This is [Anthropic’s LLM] CLAUDE’s summary of our exchange about some of the implications of Judith Gervain’s work presented in her Inaugural Address to the Hungarian Academy of Sciences on the occasion of her induction as External Member.
Our exchange started with Judit Gervain’s research on what babies learn about language before they are born. Because the womb acts as a low-pass filter — muffling high frequencies while letting through the melody and rhythm of speech — the fetus hears prosody long before it hears anything segmental. Gervain’s work shows that this prenatal exposure is not passive: newborns already come to prefer the rhythmic patterns of their native language in utero; their brains show lasting changes specific to the language they heard in the womb. The prosody, Gervain argues, is not just a curiosity but a scaffold — it helps the infant later carve speech into words and phrases, bootstrapping into grammar.
A natural comparison is with songbirds, where a lively literature shows something strikingly similar: embryos inside the egg respond differentially to their mother’s calls, and what they hear before hatching shapes their later vocal learning and even their tutor preferences. This parallel is real but has limits. For birds, the functional endpoint is a song; there is nothing playing the role that prosody plays in Gervain’s story, namely as a lever into morphosyntax. And the methods differ sharply — bird researchers mostly measure heart rate in ovo, a cruder window than the NIRS and EEG imaging Gervain uses on sleeping neonates.
From prosody our exchange moved to a harder question: is there early evidence for something specifically linguistic, as opposed to just generally auditory, being innate? Here the evidence becomes more pointed. Gervain’s 2008 PNAS paper showed that newborns’ brains, including Broca’s area — the region central to language in adults — respond distinctively to syllable sequences with an abstract repetition structure (ABB: mu-ba-ba) but not to random sequences, and do so from the very first trials, before any learning could have occurred. This is not acoustic discrimination; the infants are detecting an abstract positional identity relation across varying tokens. A 2022 multi-authored study pushed further, showing that newborns’ left-hemisphere language regions respond more strongly when structured sequences are exchanged communicatively between two speakers — suggesting sensitivity not just to structure but to the informational use of structured sound.
Whether any of this is specifically Chomskian — i.e., evidence for innate, autonomous syntactic principles like structure dependence, hierarchical phrase structure, or binding constraints — is a more demanding question, and the honest answer is: not yet, and possibly not yet demonstrable at this age level, because truly syntactic constraints only manifest when the infant is processing semantically contentful multi-word utterances, which comes later. The phoneme categorical perception story (the famous R/L case: Japanese babies can tell these sounds apart categorically at six months but lose the ability by twelve, because Japanese merges them) is the cleanest early evidence for a prepared, language-specific learning mechanism — but it sits at the phonological rather than the syntactic level.
What the full picture suggests is something like this: humans arrive with a nervous system that is pre-tuned for language in ways that go well beyond general auditory sensitivity — left-hemisphere lateralization at birth, sensitivity to abstract sequential structure, rapid prosodic bootstrapping into grammatical word order by seven months — but the specifically Chomskian claim, that there are innate, purely syntactic, principles operating autonomously of semantics and pragmatics, still lacks direct developmental evidence in infancy. The prepared biases still look more like a richly language-fitted general learning system than like a pre-loaded universal grammar. Whether that distinction matters for the autonomy of syntax remains, fittingly, an open question.
