You understood that last sentence, but not because you had met it before. You almost certainly read it for the first time just now, and whoever wrote it did not pull it out of memory either. Everyone produces and understands sentences that have never existed before, all day long, and a two-year-old has already started doing the same thing.
The hard part is what he has to work with: only the speech around him. No textbook, no table of rules, and hardly anyone correcting him line by line when he gets something wrong. One side says that isn't enough, so something must come pre-installed. The other side says it is enough, and that we have underestimated how much information that speech carries while overestimating how abstract a young child's grammar is to begin with.
Three things this time: what the "not enough" side offers as evidence, what the "enough" side answers with, and what shape the argument has taken today.
To turn an English statement into a yes/no question, you move something: the is in "The man is tall" goes to the front, and you get "Is the man tall?"
Lengthen the sentence and the trouble starts. "The man who is beating the donkey is mean" has two instances of is. Which one moves?
One rule counts by position: front the first is you come to. That rule is far simpler — you need no notion of a clause, just left-to-right counting. It gives you "Is the man who beating the donkey is mean", a sentence any English speaker sees is wrong at a glance.
The other rule has to cut the sentence into chunks first: "the man who is beating the donkey" is one whole chunk, the part that says who; the is that moves is the one outside that chunk. That gives "Is the man who is beating the donkey mean", the right one. This property — the rule looks at structure, not at position — is called structure dependence.
How do we know they don't take it? In 1987 Crain and Nakayama sat thirty three-to-five-year-olds in front of a puppet and asked them to put questions to it: "Ask Jabba if the man who is beating the donkey is mean." The children made all sorts of errors — most often saying is twice — but not one of them produced the error the counting rule predicts.
That is the shape of the argument known as the poverty of the stimulus: children draw distinctions that go beyond what the material they heard can support, so where does the extra come from?
Two parts of that foundation have since loosened. First, the corpora: Pullum and Scholz went and counted in 2002 and found the crucial sentence type is not absent at all, running to something like one percent of the interrogatives in a typical corpus; Legate and Yang replied the same year that what matters is the speech children actually hear, where it is rare to the point of near-absence. Second, the experiment: Ambridge, Rowland and Pine redid it in 2008 with modals like can and should and with plural are in place of is — and children did produce the counting-rule error.
So the accurate statement is not "children never make this error" but "the error is rare, and it depends on which auxiliary you use". Strength of evidence: the original result has been replicated, but change one word and the conclusion moves. One more caveat: a Mandarin yes/no question moves nothing at all — you add the particle 吗 (ma, a sentence-final question marker) and you are done. The puzzle simply does not arise in that language, and the evidence above is evidence English happens to supply.
The other side comes in not through a syntactic puzzle but through a diary. Tomasello recorded his daughter's verbs day by day in 1992 and found that within a single week she could use one verb in several different frames while the one she had just picked up came in only one. What she had was not "how to build a sentence" but "how cut works", "how draw works" — one island per word, which is why this is called the verb island pattern. The whole position on this side is called usage-based: structure is an abstraction slowly ground out of particular utterances, not a frame that words get dropped into.
In 1999 Akhtar ran a clean test of it. She taught two-, three- and four-year-olds a made-up verb, demonstrating it in a deliberately odd word order — English normally runs agent, verb, patient, and she used orders like agent, patient, verb. Then she brought out new puppets and asked what was happening. About half the two-year-olds reproduced the odd order. The four-year-olds almost all corrected it to English order. Word order is not learned in one shot; it accumulates verb by verb.
But this side took a hit too. In 2006 Gertner, Fisher and Eisengart switched to a looking task: twenty-one-month-olds heard "The duck is gorping the bunny" — gorp being a word they had never met — while one screen showed a duck pushing a bunny and the other a bunny pulling a duck. They looked longer at the first. On word order alone, they had sorted out who was doing what to whom with a brand-new verb.
One result says abstraction comes late, the other that it comes early, and they are not measuring the same thing: one asks a child to produce a sentence, the other only to look a moment longer. Comprehension runs ahead, production trails, and the gap between them is now a research topic in its own right.
More interesting still, the picture changes when you change language. Dittmar and colleagues tested German children in 2008. German marks who-does-what not only by order but by case marking — articles and endings that flag the agent — and its word order is looser than English. Two-year-olds understood only when order and case pointed the same way; five-year-olds could go on order alone; only at seven did they, like adults, trust case when the two cues conflicted. What gets abstracted, and how fast, tracks which cue that particular language makes reliable. Strength of evidence: these are replicated laboratory results, and the disagreement is about interpretation, not about the data.
The real change in this argument is not that somebody won; it is that the question changed shape. It used to be "is syntactic knowledge innate", which no data can settle. Now it is: how strong a bias does a learner need in order to arrive at structure dependence from roughly the amount of material a child gets, and is that bias specific to language or the sort any sequence learner has? Both versions can be run on a computer.
Perfors, Tenenbaum and Regier did one side in 2011. They fed real child-directed speech to an ideal Bayesian learner and let it choose between grammars of the counting kind and grammars of the chunking kind. It chose the latter — not because a preference had been installed, but because on this data chunking pays better overall: a slightly more complicated set of rules that explains the corpus far more economically. In that setting, at least, "the material cannot settle it" is false. The discount to apply: the set of candidate grammars was supplied in advance by the researchers.
Yedetore, Linzen, Frank and McCoy did the other side in 2023, and did it closer to life. Instead of an ideal learner they used two kinds of network with no built-in preference for hierarchy — a recurrent one and a Transformer — trained on child-directed text comparable in quantity and content to what a child hears. The result splits in two: the models captured the surface statistics well, and then generalised by the counting rule as soon as they had to form questions.
Scaling up did not overturn that. Lan, Chemla and Katzir moved the test onto six models in 2024 — two recurrent networks and four Transformers, including the general-purpose large models public at the time — probing a finer phenomenon in English wh-movement. They set a fairly lenient bar for success, and the models still often failed it.
So the state of play is this: on the same material, an ideal learner picks out structure dependence and actual networks do not. Where the difference lies is an experiment you can run, not a position to hold. Neither side gets a free step here — the models get text only, with no intonation, no scene, no back-and-forth feedback, and a child has all three; reading the models' failure as evidence for human innateness and reading their success as "so nothing need be innate" are the same unearned step.
The usage-based side has been making itself quantitative too. Children produce sentences nobody says to them, like "Don't giggle me", and then quietly withdraw them over a few years. The withdrawal runs on two things: preemption, hearing other people repeatedly use a different form in that same slot, and entrenchment, one form being heard so often that it squeezes its competitors out. Both are functions of frequency, so you can compute a prediction from a corpus and then go test it.
One reading is that it has been falsified. Another is that it retreats from "never occurs" to "occurs far less than chance", and the second still needs explaining. Different next steps separate the two: the first needs only a few more attested errors, the second needs a baseline — if children went purely by position, how many bad questions should there be? What you compute that baseline from is itself hard.
The worrying possibility is that it is neither — the two groups had different tasks, different materials, different labs. What would have to be equated for the cross-linguistic comparison to mean anything? And if a difference survived the equating, would you attribute it to how reliable each cue is, or to where each cue sits (case on the word ending, order at the front of the sentence)?
If a preference for hierarchy is really one thing, it shouldn't surface only in questions. Think about where else you have to cut chunks before you can operate: what a pronoun refers to, how far a negation reaches, the requirement that two coordinated items be the same kind of thing. A proposal that holds only for the phenomenon it was tested on is usually measuring that phenomenon, not the proposal.