Listen to an English-speaking three-year-old talk about yesterday and you may hear "I goed to the park." The past tense of go is went; goed is something he assembled himself, taking go and adding the past-tense -ed.
The odd part is not that he got it wrong. The odd part is that six months earlier he was saying went, and saying it reliably.
This error has nowhere to come from. Nobody ever said goed to him, so he cannot have picked it up. He has not forgotten went either, because he had been using it for months. The only explanation left is that somewhere in those six months he acquired a rule of his own making — and that the rule is strong enough to override a word he had already memorised.
Three things this time: how many of these errors there actually are (far fewer than almost anyone assumes), why they only ever run in one direction, and how, in another language, the child goes wrong somewhere else entirely.
The shape has a name: overregularization — applying a pattern that holds for most words to a word that takes an exception. Most English past tenses add -ed (walk, walked), but a few dozen of the commonest verbs do not: go becomes went, break becomes broke, eat becomes ate. Those few dozen are exactly the ones a child says first, and at first he says them correctly — he memorised them one by one.
Then he starts producing goed, breaked, eated, even wented (the already-correct went with -ed added on top). A few years later these quietly disappear. Drawn as a curve it is a U: high, then a dip, then back up.
The question that matters is how deep that U goes. In 1992 Marcus and several collaborators counted. They took the spontaneous speech of 83 children from a public archive of child recordings and pulled out every place an irregular past tense was called for — 11,521 of them. What share were overregularized? A median of 2.5%.
A rate of 2.5% means one slip in forty. That surprised a lot of people, since "children go through a phase of saying goed all the time" is close to folk knowledge. But it actually strengthens the argument: if there were a genuine defect in the child's grammar waiting to be undone, the errors should not be this thin and this persistent — what Marcus and colleagues recorded is a low, steady output running from just past two into the school years, not a storm.
On evidence strength: this is large-sample corpus work that has been recomputed repeatedly, but the 2.5% figure is sensitive to how densely a child is recorded. Maratsos, using much denser sampling in 2000 (the same child taped every week), got a markedly higher rate — the finer the sampling, the more you catch. What is stable is the shape and the direction; the number is not.
If the child were simply guessing, the errors should scatter both ways: sometimes an irregular treated as regular (go becoming goed), sometimes a regular treated as irregular — wing, say, coming out as wang on the model of sing and sang.
The second kind barely happens. Xu and Pinker went looking for these "weird forms" in 1995 and found them close to noise: roughly 0.2% in English, about 0.01% in Spanish, under 1% for German participles. Against 2.5%, the two directions differ by more than an order of magnitude.
One caution. Among the forms they counted as weird is brang (for bring), which is simply how some varieties of English say it, and is treated as nonstandard only from the position of the standard variety. Counting it as an error is a social judgement, not a fact about the language — and the researchers acknowledge the line is arguable.
The asymmetry itself holds, and it points at a structure: the rule is a default route, the exceptions are a list you have to look up item by item. If the lookup succeeds you use the exception; if it fails you fall through to the default. The default is always there, so it spills outward; an exception only appears when it is retrieved, so it never floods inward.
Is the default just whichever pattern is commonest? German says no. German nouns take several plural endings, and -s covers only a small share of the dictionary. Yet for loanwords, surnames, acronyms and any freshly coined word, German speakers reach for -s almost without exception: the family called Mann are die Manns. Marcus, Brinkmann, Clahsen, Wiese and Pinker turned this into an argument in 1995: a default is not established by counting, it is established by what gets used when nothing else applies. Evidence strength: still contested — recent nonce-word experiments have not reliably reproduced so clean an -s default, and single-route analogy models absorb part of the data.
There is a finer signature too. Prasada and Pinker gave adults invented verbs in 1993 and asked for past tenses. Willingness to supply an irregular form tracked how much the verb sounded like existing irregulars; willingness to add -ed tracked similarity to existing regular verbs not at all. That is what a rule is: it does not consult the neighbours.
A child learning Mandarin has no past tense to get wrong. Mandarin verbs do not change form: 去 (qù, "go") stays 去 (qù). So what does that child's goed look like?
It lands on classifiers. Mandarin puts a classifier between a number and a noun, and which one you use depends on the shape or class of the thing: 一本书 (yì běn shū, "one volume-classifier book"), 一张纸 (yì zhāng zhǐ, "one flat-sheet-classifier paper"), 一条鱼 (yì tiáo yú, "one long-thin-classifier fish"), 一只猫 (yì zhī māo, "one animal-classifier cat"). Children say 一个书 (yí gè shū) and 一个纸 (yí gè zhǐ) — the general classifier 个 (gè) applied everywhere.
Erbaugh recorded four Taiwanese children at home in 1986, aged one year ten months to three years ten months, 64 hours in all. The order she saw: the earliest classifier uses are item-bound, one classifier appearing only with one particular object, more like part of that object's name than a rule you could carry elsewhere; the abstraction grows in later. And along the way, 个 (gè) becomes the all-purpose part.
A precision worth keeping, or this turns into an exaggeration: 个 (gè) is already the broadest classifier in adult Mandarin and does substitute for a fair range of specific ones — but not for the long-and-thin class 条 (tiáo), the flat class 张 (zhāng), or the tiny-and-round class 粒 (lì). The child is not inventing anything; he is widening a path that was already wide. Same shape as goed: a default spilling into territory that belongs to an exception.
Change language again and the phenomenon nearly vanishes. Turkish marks who did what to whom with suffixes: one dedicated suffix per grammatical meaning, unfused, not dropped, in a fixed position. How transparent is that? There are records of children using case marking productively at fifteen months, still at the single-word stage, and using it quite steadily around two, with errors rare enough to barely constitute a research topic.
So how much overregularization appears is not a measurement of the child. It measures how transparent the rules of that language are, how many exceptions there are, and whether the exceptions cluster in the most frequent words.
One conclusion to block on the way out: this cannot be read as Turkish being "easier to learn" or Mandarin "simpler". Transparent morphology is paid for with longer words and more suffixes; Mandarin puts its distinctions into classifiers and word order, and that is where its children spend the time. The difficulties sit in different places, and totals do not add up for comparison that way — as "Words Are Assembled" put it, the sentence "Chinese has no morphology" has to be taken apart before it is accurate at all.
First sort out whose argument it serves. "Rare" is there to counter the idea that the child's grammar holds a defect that has to be undone later, and what that needs is "the errors do not come in bursts", not any particular number. The shape carries the argument; the number cannot. So what should be measured next — within one week of one child, are the errors spread evenly or clumped?
In form they are nothing alike. To put them together, the criterion has to sit in the process: both take a default abstracted from many examples and lay it over a slot that belongs to an exception. But that requires first showing the child really has abstracted a default rather than falling back on the word he has heard most. Those two are hardest to separate for exactly this item — 个 (gè) is both the default and the most frequent one.
Both readings work: his memory may be good enough that the rule is not needed, or the rule may not have formed. They predict different things — the first child will still add -ed to a verb he has never met, the second will not. Judging someone's learning by error rate alone often gets the direction backwards.
One is "does not match what adults say", the other is "ungrammatical". Goed is the first kind. Brang is not even that — it is simply how some varieties of English say it, and it counts as an error only from the standard variety's position. So when child error rates get tallied, how much of the judgement is really a tally of social norms?