Is there a critical period for learning a second language, and when should a child start?

Verdict

"Younger is always better" survives because three different questions share one name. They have three different answers, and once they are separated the practical advice changes completely.

1. How fast do you learn at first? Older learners are faster. This is not controversial and it is not new — it has been the finding since 1978.
2. How good do you eventually get? Earlier starters end up better. The gradient is real, large, and found in every dataset anyone has looked at, from 46 Korean immigrants to 2.3 million census respondents. What is not found reliably is the discontinuity — the cliff — that the phrase "critical period" asserts.
3. Should a school start its language programme in Year 1 or Year 3? On the evidence, it barely matters, and where it has been measured properly the earlier start has as often lost as won.

The verdict is mixed because the honest summary contains a real effect and a real null sitting next to each other. Age of first exposure genuinely predicts ultimate attainment. An earlier classroom start, at the doses schools actually use, genuinely does not.

What the evidence shows

The most misreported paper in language education is Hartshorne, Tenenbaum & Pinker. It is routinely cited as proof that children must start young because of a childhood critical period. Its own estimate puts the change in learning rate at 17.4 years — the end of adolescence, not the start of it. The "start by 10-12" recommendation is a derived claim with a completely different logic: learning a language to native level takes ten to twenty years, native speakers are still improving at 25, so anyone who wants to finish has to begin in late childhood. That is an argument about duration, not about a window slamming shut. A ten-year-old and a fourteen-year-old are on the same side of every break this literature has ever estimated.

The discontinuity is the contested part, and the gradient is not. Every dataset agrees that people who start younger end up better. Almost none of them, when the breakpoint is estimated rather than assumed, produces the sharp elbow that the words "critical period" promise. Hakuta found none in 2.3 million census records. Vanhove showed the field's standard analyses could not have detected one either way. Van der Slik's reanalysis of Hartshorne's own data finds a continuous decline for every learner type except non-immersion learners — and puts their break at 18.6, which looks a great deal like the age at which formal schooling stops.

The evidence a school actually needs points the other way from the folklore. Three large European studies have compared children who began classroom foreign-language instruction at different ages. Barcelona: later starters learn faster at equal hours. Germany: the early cohort's lead had inverted by Year 7. Switzerland: no general advantage across five years, with a benefit visible only in children whose families were already bilingual and supportive. Only Porsch's Year-9 sample finds a durable early-start advantage, and it is the least well-identified of the four. Two of the three cleanest school-based comparisons find that starting earlier at ordinary primary-school intensity buys nothing.

Hereditarian-lens assessment

Risk: medium, and the reason is that this topic contains two literatures with very different exposure to the premise.

The naturalistic age-of-acquisition literature is high-risk. Age of arrival in a new country is a family decision. Families that emigrate when their children are small differ from families that emigrate later in education, resources, motivation, and — this archive's standing point — in heritable language aptitude, which is passed to the children along with the decision. Every one of these studies is observational; not one randomises anything. The direction of that bias is toward finding age effects, which is worth noting: the confounds and the hypothesis push the same way, and even so the discontinuity does not appear.

The classroom-start literature is medium-risk, and that is the half the practical verdict rests on. Jaekel's and Porsch's cohorts are defined by state policy and birth year; Muñoz's by a curriculum change. Parents did not choose the starting age, so passive gene–environment correlation cannot generate the contrast. What remains is cohort effects, teaching quality, and school-track allocation — real threats, but not the premise's threat.

One result belongs squarely under the lens. Pfenninger & Singleton's finding that the early start helped only "biliterate simultaneous bilinguals receiving substantial parental support" is a correlation between a family-selected, heritable-ability-loaded category and a response to treatment. It is exactly the shape of finding this archive refuses to read as an aptitude–treatment interaction, and it should not be used to argue that early language instruction works for "the right kind of family."

And the far-transfer claim fails as usual: Gunnerud's meta-analysis finds no reliable cognitive advantage for bilingual children. Nothing in this topic bears on g.

Boundaries & what critics say

• Nothing here is randomised. Nothing. There is no experiment in this entire literature, because you cannot assign a child a birth date or an age of immigration. The strongest identification available is a state policy that changed the starting grade for a whole cohort, and this archive grades that C. Anyone quoting a critical-period effect size is quoting an observational estimate.
• The internet-quiz measure is a written grammar test. Hartshorne's 95 retained items are grammaticality judgements and sentence–picture matching, taken online, by people who chose to take a viral quiz about their native language. No speaking, no listening, no writing, no standardized instrument. The best-instrumented age study in the literature — Schepens 2022, on real Dutch state examinations across four skills — finds a monotonic decline beginning around arrival age 25, which looks much more like ordinary cognitive ageing than like a language-specific window.
• The Hartshorne / van der Slik dispute is unresolved and both sides are modelling the same self-selected sample. Neither has produced new data. That is why this topic does not treat the 17.4 breakpoint as established, and equally does not treat it as refuted.
• Even early starters mostly fail close scrutiny. The half of Abrahamsson & Hyltenstam that gets quoted is that no late learner reached the native range. The half that does not get quoted is that only a few of the early learners did either, in a sample deliberately selected for sounding native. If "native-like" is the goal, the evidence says almost nobody gets there by any route.
• Age of onset and years of exposure are entangled in almost every design. A person who arrived at 5 has had more years of the language than one who arrived at 25, at every observation age. Hartshorne's contribution is a model that tries to separate the two; the criticism of that model is precisely about whether the separation worked.
• Critical-period advocates would say the classroom studies are irrelevant — that a couple of lessons a week is not the input a maturational mechanism needs, so their nulls say nothing about the window. That is a fair objection, and it is also a concession that matters more than the dispute: it concedes that the critical period, whatever its shape, does not license an earlier school start.

Practical guidance

• For real immersion, start early. The age-of-acquisition gradient is one of the most reproducible findings in the field even though it is not a cliff. If a genuinely bilingual environment is available, earlier is better, and there is no age at which the benefit switches on or off.
• For a school foreign-language programme, buy hours, not an earlier start date. Two of the three best European comparisons find that adding years at ordinary primary intensity produces no durable advantage, and one finds the early starters behind two years later. The Barcelona result — later starters learn faster at equal hours — is the mechanism: an older child extracts more per hour.
• Do not sell an early-start programme on the critical period. The anchor paper's own break is at 17, its reanalysis says there is no break, the census data show no break, and none of them studied a school. A programme justified this way is vulnerable to the first parent who reads the paper.
• Do not promise native-likeness. Not from school instruction at any age, and — on the only study that scrutinised the claim properly — barely even from early naturalistic bilingualism.
• Do not add a cognitive-benefit argument. "Bilingualism makes children smarter" is a far-transfer claim, the children's meta-analysis finds nothing, and attaching it to a language programme makes the programme falsifiable on a claim it never needed.
• If you must pick an age on thin evidence, pick the one that maximises total hours before the child leaves you — which usually means starting later and going harder, not starting earlier and going slower.

Open questions

• No study in this topic randomises anything, and the obvious design — randomise schools to an earlier or later start at matched total hours — has never been run, despite Europe having reorganised primary curricula around the answer.
• Jaekel and Porsch disagree on overlapping German populations at different horizons. Nobody has reconciled them, and the reconciliation is the single most valuable next study for a school.
• Rate, attainment and the classroom question have never been measured in the same sample. Every claim in this topic comes from a different literature with a different instrument.
• Almost every ultimate-attainment estimate uses a written grammaticality-judgement task. The one study using real four-skill proficiency exams finds a different shape entirely. Whether the breakpoint literature is measuring language or measuring grammar quizzes is unresolved.
• Nothing measures what a school cares about most: whether a child who started at 6 versus 10 can hold a conversation at 18. Every school-based outcome in this topic is a receptive test.