A child’s syntax outruns the sentences that built it
Every child who grows up around auxiliary-fronted questions like “Is the cat that is sleeping on the mat hungry?” learns, by around age four, to attach the fronted “is” to the correct clause even in sentences the child is unlikely to have heard before, and does so without ever producing the alternative that a simple linear rule would predict: fronting the first auxiliary in the string regardless of which clause it belongs to. Noam Chomsky’s poverty-of-the-stimulus argument treats this as diagnostic. Children’s grammatical judgments are more determinate, more uniform across children with very different individual histories of exposure, and more tightly constrained than the finite, noisy sample of sentences any one child actually hears could by itself explain. Robert Berwick, Paul Pietroski, Beracah Yankama and Chomsky restated the argument in 2011 using this exact case: children never produce the linear, structure-independent rule for question formation, even though it is the simpler rule to state and even though the sentences that would expose its output as wrong are individually rare in child-directed speech [12]. Their framing is explicit about what the argument requires: a description of the knowledge actually attained, in its full richness, and then an account of whether that knowledge could plausibly have been extracted from the data actually available — not simply a demonstration that children hear less than an adult grammarian would like [12].
That is a substantive empirical claim, and it can be checked rather than merely asserted. If children are extracting a structure-dependent rule for auxiliary fronting from ordinary experience, the sentence types that would disambiguate it from the simpler linear rule have to appear somewhere in what they hear, however rarely; if such sentences are essentially absent, an experience-based account has a harder case to make. Berwick, Pietroski, Yankama and Chomsky’s own reading of transcribed child-directed speech is that the relevant construction type is close to nonexistent in usable corpora, which is precisely the poverty their argument is named for [12]. Critics of the poverty-of-the-stimulus tradition have long pushed back on exactly this point — whether the disambiguating evidence really is as scarce as claimed, whether a large enough corpus would turn up more of it than smaller samples suggest, and whether children need the evidence in the specific form nativists assume — and that dispute over what counts as usable evidence in a transcribed corpus has never fully settled either way. What is not in dispute is the shape of the question: a specific, countable gap between input and output, answerable in principle by better corpora and more careful modeling rather than by an appeal to intuition about what language “must” be like.
Pinker and Bloom priced grammar as a textbook adaptation
Steven Pinker and Paul Bloom’s 1990 target article in Behavioral and Brain Sciences took the poverty-of-the-stimulus tradition’s conclusion — that grammar is a specialized, largely innate computational system — and asked a question Chomsky had conspicuously declined to answer the same way: given that a language module exists, how did it get there? The paper opens by cataloguing the case for treating language as a biological specialization at all. Children become fluent speakers of complex grammatical sentences by around age three without formal instruction; some language disorders are genetically transmitted; aspects of language skill localize to characteristic brain regions; and the human vocal tract is shaped in ways that compromise breathing and swallowing in the service of speech [1]. None of this catalogue was original to Pinker and Bloom. Their own words for it are that this “list of facts… suggests that the ability to use a natural language belongs more to the study of human biology than human culture” [1]. Their contribution was to insist that this conclusion has an unavoidable second half. If grammar is a species-specific, task-specific biological adaptation, evolutionary theory supplies a specific, falsifiable account of how such traits arise: complex functional design, plus the absence of any equally plausible alternative process, and no other mechanism recognized in the modern synthesis is known to produce that particular combination [1].
The paper’s real target was not a reader who doubts grammar is special, but two people who agreed it was and still balked at natural selection as the explanation: Chomsky himself, and Stephen Jay Gould. Both had suggested, across the 1970s and 1980s, that language could have arisen as a byproduct of an increase in overall brain size or from unspecified “laws of growth and form,” rather than as a target of selection in its own right, a position Massimo Piattelli-Palmarini had recently restated in unusually explicit terms [1]. Pinker and Bloom’s answer treats this as a testable empirical claim rather than a philosophical objection. A byproduct or spandrel account predicts that a trait’s structure should be explicable by the developmental or physical process that generated it as a side effect; an adaptation account predicts intricate, quasi-engineered fit to a specific function that no known byproduct process tends to produce. Grammar, on their reading, shows the second signature — a serial channel doing propositional work that resembles no known cognitive byproduct — and they conclude flatly that “there is every reason to believe that a specialization for grammar evolved by a conventional neo-Darwinian process” [1].
It is worth being precise about what kind of evidence this is, because the paper is often summarized as though it settled something empirically that it in fact settled argumentatively. Pinker and Bloom did not present a new fossil, a new gene, or a new experiment; they presented a design argument, of exactly the kind evolutionary biologists use to infer adaptation from structure when a lineage’s history cannot be observed directly, and their case is only as strong as the premise that grammar’s complexity has no better non-selective explanation available. That premise has held up as the majority position among linguists sympathetic to nativism in the thirty-five years since, but it was never universally accepted even inside that camp — and the clearest evidence of that is that the very co-author Pinker and Bloom were arguing against would, twelve years later, propose a version of the nativist program dramatically narrower than the one Pinker and Bloom had spent their paper defending.
Chomsky narrowed his own claim to recursion alone
Marc Hauser, Chomsky and W. Tecumseh Fitch’s 2002 Science review draws a distinction between the faculty of language in the broad sense, FLB, and the faculty of language in the narrow sense, FLN. FLB includes a sensory-motor system, a conceptual-intentional system, and the computational mechanisms for recursion, and the paper is content to describe FLB in ordinary adaptationist terms; FLN is defined as whatever within that broader system is both uniquely human and uniquely dedicated to language [2]. Their central hypothesis is stated as plainly as a hypothesis in a hedge-averse field ever gets stated: “we hypothesize that FLN only includes recursion and is the only uniquely human component of the faculty of language” [2]. Everything else that makes human language look distinctive — the sensory-motor machinery of speech, the conceptual system that supplies meanings, even most of what an introductory grammar course calls syntax — is, on this view, either shared with other species or not specific to language. What that leaves is one recursive engine as the sole component that needs its own evolutionary account.
This is a genuinely narrower claim than Pinker and Bloom’s, and it matters that it originates inside the nativist camp rather than from a usage-based critic outside it. Where Pinker and Bloom treated the entire apparatus of grammar as the adaptation requiring explanation, Hauser, Chomsky and Fitch explicitly floated the possibility that FLN “may have evolved for reasons other than language,” with comparative research looking for recursive computation in domains such as number, navigation and social relations rather than in communication at all [2]. That is close to the byproduct account Pinker and Bloom had spent their paper arguing against, now advanced in a qualified form by one of that paper’s own addressees. The 2002 paper frames this as a research program rather than a closed conclusion — its stated aim is to promote a stronger connection between linguistics and evolutionary biology by clarifying what a theory of language evolution would need to explain, not to announce that the broader adaptationist case was wrong [2] — but the effect on the field was immediate. A hypothesis this specific, this narrow, and this directly attributable to Chomsky by name was always going to draw a response from adaptationists who had built their case on a much broader target, and it did not take long to arrive.
Pinker and Jackendoff said recursion-only throws away too much
Pinker’s co-author this time was Ray Jackendoff rather than Bloom. Their 2005 paper in Cognition frames the disagreement precisely: they “examine the question of which aspects of language are uniquely human and uniquely linguistic in light of recent suggestions by Hauser, Chomsky, and Fitch that the only such aspect is syntactic recursion, the rest of language being either specific to humans but not to language… or not specific to humans” [3]. Their verdict is blunt: they “find the hypothesis problematic,” because it “ignores the many aspects of grammar that are not recursive, such as phonology, morphology, case, agreement, and many properties of words” [3]. A language with no recursion at all would still need an inventory of speech sounds organized into contrastive categories, a system of word-formation rules, case and agreement marking that varies across languages in ways no theory of recursion predicts, and tens of thousands of learned, arbitrary form-meaning pairings — none of which reduces to, or automatically follows from, an operation that lets a rule apply to its own output.
Pinker and Jackendoff add that the recursion-only hypothesis sits awkwardly with independent evidence from outside syntax proper. It is, they write, “inconsistent with the anatomy and neural control of the human vocal tract,” and it is further weakened by findings that speech perception does not reduce to general primate auditory abilities, that word learning is not simply a form of fact-learning, and that at least one gene involved in speech and language was evolutionarily selected in the human lineage without being at all specific to recursion [3] — a pointed early reference to FOXP2, which becomes a story of its own later in this piece, and whose real evidentiary status turned out to be considerably less settled than either side of this 2005 exchange assumed. Pinker and Jackendoff go further than cataloguing omissions; they offer a diagnosis of where the recursion-only hypothesis came from, tracing it to Chomsky’s Minimalist Program, a syntactic framework that deliberately de-emphasizes exactly the aspects of grammar — rich morphology, language-particular idiosyncrasy, the sheer bulk of an arbitrary lexicon — that the 2002 paper leaves outside FLN [3]. Their claim is not that the Minimalist Program is wrong as a theory of syntax, but that a framework built to streamline syntactic description is a poor basis for a claim about what evolution needed to explain, since it may simply be describing away the parts of language that do not fit its architecture rather than showing those parts were cheap to evolve.
This is, in the end, a dispute between two camps that agree language is substantially innate and that recursion matters, disagreeing instead about how much of grammar an evolutionary account is actually on the hook for. It is worth stating plainly that Pinker and Jackendoff do not, in this paper, present new experimental evidence against FLN as such; their argument is conceptual and evidentiary-synthesis, weighing existing findings from anatomy, speech perception and genetics against one specific published hypothesis. That is exactly the register in which most of this particular disagreement has been conducted on both sides, and it is why the exchange reads less like a decisive refutation than like two research programs stating, in public and by name, how far apart their premises actually are.
Everett said Pirahã breaks the recursion universal
If the Hauser-Chomsky-Fitch hypothesis is correct, it makes an unusually clean, falsifiable prediction: every human language should show recursion, because recursion is supposed to be the one thing FLN cannot do without. Daniel Everett’s 2005 paper in Current Anthropology reports fieldwork on Pirahã, a language with a few hundred speakers in the Brazilian Amazon that Everett had studied since the late 1970s, and states the challenge to that prediction as directly as a linguist can state anything: Pirahã, he writes, “is the only language known without embedding (putting one phrase inside another of the same type or lower level, e.g., noun phrases in noun phrases, sentences in sentences, etc.)” [4]. His paper catalogues a long list of features he describes as absent alongside embedding: numerals and a concept of counting, color terms, quantifiers such as “all,” “each” and “every,” a perfect tense, and, he adds, most fiction, creation myths, and any individual or collective memory reaching back more than about two generations [4].
Everett’s explanation for the whole cluster is a single proposed cultural constraint he calls the Immediacy of Experience Principle: Pirahã culture restricts assertions to claims that fall within the immediate experience of the speaker or of someone the speaker knew directly, and he argues that this one cultural value, not a defect or a simpler grammar, accounts for why so many features that look unrelated from a grammarian’s chair disappear together in a single language [4]. He is explicit about what he thinks is at stake for linguistic theory more broadly, writing that “if the form or absence of things such as recursion, sound structure, word structure, quantification, numerals, number, and so on is tightly constrained by a specific culture… then the case for an autonomous, biologically determined module of language is seriously weakened” [4]. He also makes a point of his own authority as a source, reporting that he had lived among and studied the Pirahã since 1977, visiting nearly every year since, and describing himself and his wife Keren as the only two non-Pirahã people who speak the language fluently [4]. That combination — a genuinely startling empirical claim, offered by essentially the only outside linguist with sustained access to the relevant data — is what made the paper impossible for the field to ignore, and, as the next section shows, impossible to adjudicate quickly either.
Nevins, Pesetsky and Rodrigues said the data doesn’t show what Everett claims
Andrew Nevins, David Pesetsky and Cilene Rodrigues published a lengthy reassessment in Language in 2009, and its conclusion is a direct denial of Everett’s central empirical claim rather than a dispute about interpretation alone: “we find no evidence, for example, that Pirahã lacks embedded clauses, and in fact find strong syntactic and semantic evidence in favor of their existence in Pirahã” [5]. Their method was to return to the grammatical description Everett himself had published in 1986, before he proposed the Immediacy of Experience Principle, and argue that sentence types he had once analyzed as containing embedded clauses were being reanalyzed in 2005 as something else, without new data forcing the reanalysis. They extend the same style of argument to other items on Everett’s list, writing “likewise, we find no evidence that Pirahã lacks quantifiers, as claimed by Everett,” and they close by rejecting the theoretical structure of the whole argument, not merely its individual pieces: “we conclude that there is no evidence from Pirahã for the particular causal relation between culture and grammatical structure suggested by Everett” [5].
Everett’s reply, also published in Language in 2009, does not retreat. He restates his own working definition of the property under dispute — “recursion consists in rule (or operation) sets which can apply to their own output an unbounded number of times” — and argues that Nevins, Pesetsky and Rodrigues have mostly reanalyzed the same quarter-century-old material rather than bringing new fieldwork of their own, while his 2005 paper had corrected specific errors in his own earlier 1983 and 1986 analyses on the strength of additional time in the field [6]. He maintains without qualification that Pirahã “falsifies the single prediction made by Hauser, Chomsky, and Fitch (2002) that recursion is the essential property of human language” [6]. Neither side, in other words, produced a decisive new dataset between 2005 and 2009 that settled the question outright; both sides argued over the correct analysis of a body of already-published examples.
That points to the detail that makes this dispute unusual among linguistic controversies, and it is worth stating plainly rather than glossing over it. The Pirahã have perhaps a few hundred speakers living in a small number of villages along the Maici river and, as Everett’s own paper notes, have remained essentially monolingual despite more than two hundred years of intermittent contact with Portuguese speakers [4]. Everett states in his 2009 response that all the data behind that paper were gathered by himself between 1977 and 2007, or by a predecessor field linguist named Steve Sheldon between 1967 and 1976, and that he had spent twenty-one months in the field prior to his 1983 analysis and a further fifty-three months in the field since [6]. That is not a criticism of either side’s competence or honesty; it is a structural fact about the evidence base. For more than three decades, the empirical record on one of the most consequential test cases for a claim about all human language has passed through essentially two people, and no outside team has published an independent, large-scale Pirahã corpus that could adjudicate the embedding question without relying on Everett’s own transcriptions and translations. The sober verdict, and the one this piece takes, is that the theoretical dispute between Everett and his critics has not been resolved because the empirical dispute underneath it cannot yet be resolved: whether Pirahã embeds remains open not for lack of argument on either side but for lack of an independently collected dataset large enough to settle it.
Statistical learning gives infants machinery nobody needed to posit as innate grammar
While the recursion debate ran on argument and reanalysis, a separate research program was accumulating experimental results that never claimed to resolve the poverty-of-the-stimulus question but did show that infants arrive with more general-purpose statistical machinery than a strict nativist account had reason to expect. Jenny Saffran, Richard Aslin and Elissa Newport’s 1996 Science paper remains the field’s most cited demonstration. Eight-month-old infants heard two minutes of a continuous, monotone synthetic speech stream built from four made-up three-syllable “words,” with no pauses, stress cues, or other surface markers at word boundaries; the only information available to distinguish a word from a non-word was that syllables inside one of the artificial words always followed each other with a transitional probability of 1.0, while syllables that happened to fall across a word boundary followed each other with a transitional probability closer to 0.33 [7]. After just this two-minute exposure, infants tested with the head-turn preference procedure listened differently to the high-probability “words” than to equally frequent three-syllable sequences that merely straddled two words, which showed they had extracted the statistical regularity itself rather than memorizing particular sound patterns by rote [7]. The transitional probability doing the work in a design like this can be written directly:
A syllable pair inside one of the artificial words has a transitional probability at or near one; a pair straddling two words has a transitional probability closer to a third. Nothing in the design requires a grammar module, a parameter to be set, or an innate syntactic category — only a mechanism sensitive to the conditional frequency of what follows what, operating over a stream the infant has never heard before in its life.
Michael Tomasello’s usage-based research program extends the same general-purpose picture from segmentation into early syntax. His 2000 review in Trends in Cognitive Sciences reports that naturalistic and experimental work converges on a single generalization about toddlers’ early multi-word speech: it is organized around “concrete, item-based linguistic schemas” tied to particular verbs and particular constructions, not around the abstract adult categories — noun phrase, verb phrase, subject — that a nativist grammar would predict children deploy from the outset [8]. On this account, children imitatively learn specific expressions from the speech they hear, then use general cognitive and social-cognitive skills, rather than a dedicated syntax module, to categorize, schematize and gradually recombine those learned expressions into more abstract constructions, arriving at adult-like generalization only “gradually and in piecemeal fashion” [8]. Tomasello’s own framing is pointed about the implication for the nativist program: item-based development, he argues, “presents significant problems for nativist accounts of children’s language development that use adult-like linguistic categories, structures and formal grammars as analytical tools” [8], because those accounts predict earlier and more uniform abstraction than the item-based data actually show.
What this research program explains well is exactly what it was built to explain: how a language-general statistical sensitivity could carry an infant from an undifferentiated stream of sound to a first working vocabulary, and how item-based imitation could carry a toddler from single memorized constructions to piecemeal productivity, without positing any machinery specific to grammar as such. What it has a harder time explaining, and what Berwick, Pietroski, Yankama and Chomsky press hardest in their 2011 restatement of the poverty-of-the-stimulus case, is the specific, structure-dependent character of adult syntactic knowledge — the fact that children never even attempt the simpler, linear alternative for auxiliary fronting, not merely that they eventually converge on the structure-dependent rule after enough exposure to it [12]. A statistical learner that tracks what tends to follow what, or a usage-based learner that generalizes cautiously from one item to the next, has to be shown to derive that specific asymmetry, not merely to be broadly compatible with it after the fact. Both research programs have active work attempting exactly this, and neither side treats the underlying question as closed.
The “language gene” was a genetic mirage built on the wrong sample
No result in this dispute traveled further outside linguistics, or got more badly garbled on the way, than FOXP2. Cecilia Lai, Simon Fisher, Jane Hurst, Faraneh Vargha-Khadem and Anthony Monaco’s 2001 Nature paper studied the KE family, a three-generation British family in which roughly half the members, across several generations, have a severe, autosomal-dominantly inherited disorder combining orofacial dyspraxia with linguistic and grammatical impairment [9]. The team mapped the disorder to a point mutation in a single gene on chromosome 7, a forkhead-domain transcription factor they named FOXP2, and confirmed that the same gene was independently disrupted, by a chromosomal translocation, in an unrelated child with a similar disorder [9]. This part of the story has held up well and been extended by other affected families and other FOXP2 mutations described since: a broken copy of this one gene reliably produces a specific, severe speech and language disorder, which is exactly the kind of finding a modular, partly innate view of the language faculty would want to have.
What did not hold up nearly as well was the story attached a year later about why FOXP2 looks the way it does across species. Wolfgang Enard and colleagues, in a 2002 Nature paper, sequenced the FOXP2 coding region in humans and several other primates and reported a pattern of amino-acid changes and reduced nucleotide diversity around the human gene that they read as the signature of a selective sweep — positive selection driving a beneficial variant rapidly through the population within roughly the last fifty thousand to one hundred thousand years, a window the paper linked to the rapid flourishing of anatomically and behaviorally modern humans in Africa [10]. The finding was reported, almost immediately and far outside the original paper’s own careful hedges, as the discovery of “the language gene”: a single locus whose recent sweep supposedly explained why humans, and not other apes, have grammar at all. The sample behind that sweep signature was modest by the standards later genomic work would apply to it — the original analysis covered only a fraction of the gene in roughly twenty individuals, drawn predominantly from populations of Eurasian descent rather than from a geographically representative sample of humanity.
That sampling choice turned out to matter enormously. Elizabeth Atkinson, Amanda Audesse, Julia Palacios, Dean Bobo, Ashley Webb, Sohini Ramachandran and Brenna Henn’s 2018 paper in Cell reanalyzed the entire FOXP2 locus using much larger, globally diverse panels of modern human genomes together with archaic hominin genomes, and reported plainly that they “do not find evidence for recent positive selection at FOXP2” [11]. Their diagnosis of what went wrong in 2002 is a lesson in sampling composition rather than in genetics as such: a summary statistic built to detect a sweep is sensitive to the ancestry mix of the sample it is computed over, and a sample drawn mostly from populations descended from the out-of-Africa migration will show reduced genetic diversity relative to a broader, Africa-inclusive panel for reasons that have nothing to do with selection at that specific locus [11]. When Atkinson and colleagues reran comparable tests treating African and non-African populations appropriately, rather than pooling a small, ancestry-skewed sample together, the sweep signal that had anchored more than a decade of “language gene” reporting did not survive [11].
FOXP2 remains one of the clearest known examples of a gene whose disruption produces a specific, severe and heritable speech and language deficit — the Lai et al. finding is not in question and has never been seriously challenged. What is now the field’s own textbook cautionary tale is the further inferential leap from that disruption phenotype to a story about recent adaptive evolution in the human lineage. That leap depended on population-genetic evidence that a better, more representative sample simply could not reproduce, and any general-audience account still describing FOXP2 as “the gene for language, selected fifty thousand years ago” is repeating a finding that its own field walked back once it had better data to work with.
What would actually decide the instinct question
None of the four disputes traced here — adaptation versus byproduct, recursion-only versus broad grammar, embedding versus no embedding in Pirahã, statistical learning versus an innate module — has been closed by a single decisive result, and it is worth being explicit about why, because the reason differs from dispute to dispute rather than reducing to one general excuse for inconclusiveness. The adaptation question is, in Pinker and Bloom’s own framing, a design argument rather than a directly observable fact about a history nobody can rerun; it will move only if someone produces either a non-selective process demonstrably capable of generating grammar’s specific complexity, or clearer evidence that the complexity itself has been overstated [1]. The recursion-only question is adjudicable in principle by comparative cognitive science, and Hauser, Chomsky and Fitch said as much when they proposed looking for recursive computation in non-linguistic domains such as number and spatial navigation [2]; that program has produced suggestive results in other species without yet forcing a consensus revision of FLN in either direction. The Pirahã question is, distinctively, an evidence-access problem before it is a theoretical one: it will not move until an independent research team gathers a large Pirahã corpus without relying on the existing published record, and both sides of the 2009 exchange would, on their own stated terms, have to accept whatever that corpus showed. The statistical-learning question is the most tractable of the four, because it is being fought with the same kind of evidence on both sides — corpus counts, computational models, and infant experiments — rather than across incommensurable kinds of argument, which is exactly why Berwick, Pietroski, Yankama and Chomsky felt the need to restate the poverty-of-the-stimulus case against newer statistical models point by point rather than dismiss them outright [12].
What should make a careful reader hold the whole question open, rather than adopting whichever camp’s rhetoric currently sounds more confident, is that every camp here has been corrected at least once by its own side. Chomsky’s 2002 paper narrowed a much broader nativist claim under the weight of its own authors’ reasoning, and did so in public, over Pinker and Jackendoff’s explicit objection three years later. Everett’s most quoted sentence is not a flat assertion but a conditional — if the facts are as he describes them, the case for a biologically autonomous language module is weakened — and his sharpest critics dispute the antecedent, the facts, rather than merely the conclusion he draws from them. And the single most widely repeated genetic fact in this entire subject, that a gene swept toward fixation because humans needed it for language, turned out on closer inspection to be an artifact of which twenty people happened to be sequenced first. The instinct question is not settled in either direction on the evidence assembled here, and the discipline this article has tried to practice throughout is the same one Atkinson and colleagues modeled when they went back and checked a fifteen-year-old result rather than continuing to cite it: before treating any one of these four claims as the field’s considered verdict, ask what evidence would have to exist to overturn it, and whether anyone has actually gone looking yet.
