Someone has lived abroad for fifteen years. Large vocabulary, no trouble in meetings, reads technical material faster than plenty of locals. But "the" and "a" still land in the wrong places, and the accent is there the moment they open their mouth. The person in the same household who arrived at six sounds like a local child a year later.
The contrast is so common that everyone has an explanation ready: a door closes somewhere around puberty. The idea has a name — the critical period hypothesis, proposed by Eric Lenneberg in 1967. The trouble is that "the door closed" runs at least three things together: whether the decline is real, whether there is an elbow in the curve, and which part of language is being shut off.
Three steps: what the age curve actually looks like, why the isolated-child cases cannot prove what they are usually enlisted to prove, and a comparison that pulls "age" apart from "did you have a language early on" — which is the hardest ground on this topic.
In 1989 Johnson and Newport recruited 46 native speakers of Chinese or Korean who had arrived in the United States at ages ranging from 3 to 39, all of whom had lived there at least five years. The task was plain: listen to 276 sentences and say which ones are wrong. The bad ones looked like this — "Yesterday the hunter shoots a deer" (the verb should be the past tense, shot).
The result: arrivals before seven scored no differently from native speakers; after that, scores dropped steadily with age of arrival; past puberty the curve roughly flattened, but the spread between individuals suddenly widened — some very low, some rather high.
The curve was later scaled up. Hartshorne, Tenenbaum and Pinker (2018) used a grammar quiz that circulated online to collect 669,498 responses, then modelled apart three things that are normally tangled together: how old you are now, how old you were when you first met English, and how many years you have been at it. Their answer put the onset of decline in grammar learning near 17.4 years — well after puberty.
That number has not held up cleanly. Slik and colleagues (2022) reanalysed the same data: for monolinguals, bilinguals and early-immersion learners a model of steady, continuous decline fits better, and only the non-immersion and late-immersion groups need an elbow at all (landing at 18.6 and 19.0 years). Vanhove (2013) had flagged a more basic problem earlier: the hypothesis predicts a break in the curve, while most papers test whether age and score are correlated. Those are two different questions.
So, as of today: the decline is solid; whether there is an elbow, and where, is not settled.
And one thing that often gets dropped: this is not one door but several, closing at different times. Accent narrows earliest — as the piece on the grid the brain lays over sound described, infants start discarding contrasts that do not belong to their home language before their first birthday. Grammar goes much later. Vocabulary barely closes at all; adults memorise words more efficiently than children do. So "too late" needs a follow-up question: too late for which part?
November 1970, Los Angeles. A girl of thirteen years and seven months was brought to a social services office. She had spent almost her entire life shut in one room, much of it strapped to a potty chair, with nobody speaking to her. The pseudonym she is known by in the research is Genie.
The linguist Susan Curtiss followed her for years. Words came in numbers; short strings like "Father hit Genie" appeared. But syntax never got off the ground: no auxiliaries, no wh-questions, unstable word order. It looks like living proof of the slogan — words any time, grammar expired.
It proves nothing of the kind. She suffered prolonged malnutrition, abuse and total social isolation at the same time, and any one of the three suffices on its own to explain the language outcome. Nor does anyone know what she was like before confinement — her father said at the time that she was "slow", a claim that can neither be checked nor ruled out. The sample is one person, with no control group, and there never will be one.
Another case pulls the conclusion the other way. A girl found in 1938, called Isabelle in the record, was six and a half; she had been shut in a darkened room with a deaf mother who did not speak, communicating only in gestures. When found, she behaved like an infant of a few months. After intensive work she reached age-appropriate language in roughly two years. She and Genie differ in two things: seven years of age, and the constant presence of one communication partner in early life — even if that gesture system was not a full language.
Two cases, one person each, pointing in opposite directions. Together they license exactly one conclusion: we do not know.
There is another layer. The people caring for Genie were also the people building research on her data; funding was withdrawn in 1974, the study broke off, and she moved through a series of foster arrangements. The case now serves as both evidence and warning: this kind of "natural experiment" cannot be clean, and making it clean would mean keeping a child under observation. Evidence that can actually answer the question has to come from somewhere else.
Somewhere else is here. There is a group of people whose first encounter with a full language happened in late childhood: born deaf, in families that did not sign, meeting American Sign Language only at school. As the piece on sign language being a full language laid out, signed languages give up nothing grammatically to spoken ones — so this is not exposure to a substitute. It is a first language arriving late.
Mayberry (1993) compared 36 such adults. All of them began acquiring American Sign Language around ages nine to thirteen, and all had used it for decades. Twenty-seven were born deaf, so the language was a late first language; the other nine were born hearing and lost their hearing in late childhood — they already had a spoken language, so signing was a second one for them.
The task was recalling structurally complex signed sentences. The second-language group scored 82%; the group with no early language scored 43%. Matched starting age, matched decades of use, nearly a factor of two apart. And within the first-language group, the later the start, the lower the score.
That twists the causal story around. Age itself is not the driver; whether you had a language early on is what sets how far you get with any language later. A first language does not just hand you that language — it installs the machinery for learning language, and once installed you can call on it at any age.
Mayberry and Lock (2003) ran the same contrast in English and got the same direction.
The evidence strength, stated plainly: the groups are small (nine in the second-language group) and the design is quasi-experimental — nobody can randomly assign children to "had early language" and "did not". But it supplies precisely what Genie's case lacks: age is matched, so early language experience is the only difference left.
Do not over-read it the other way either. This is not a claim that learning a second language late is free — the Johnson and Newport curve is still there. It is a claim that the cost is an order of magnitude smaller than the cost of a missing first language. And some late starters do land inside the native range; the proportion is small but it is not zero. A lowered ceiling, not a locked door.
"Critical" says after this it cannot be done; "sensitive" says after this it is harder. The second version is nearly unfalsifiable — any successful late learner counts as "hard but managed". To give it content again you have to name a quantity: how much harder, in which subsystem, and how much input closes the gap. Without those three, all that changed is the tone.
It stops being a preference question about which mode of communication to choose and becomes a question with a time cost: before any option takes effect, the child is spending window every year. This is exactly where signing communities and parts of medical advice have argued for a long time, often citing the same data. Worth noting: the evidence here supports getting a full language early, and says nothing about which one.
Depends what the window is meant to explain. If it explains "nobody can", one counterexample settles it. If it explains a downward shift of the whole distribution, those people are just the right tail and no counterexample at all. Papers often slide between the two readings without marking the switch — and they need completely different evidence: the first only needs a person, the second needs the whole curve.
It cannot prove a critical period exists, but it works as a design brief: precisely because it lacks an age-matched control, it forced the kind of comparison later made within deaf populations. That use requires the researcher to admit first that the case in hand proves nothing.