Teaching children a foreign language

The program (what a school should actually do)

Foreign language is unusual in this archive. In most subjects the popular practice is under-evidenced and the research is merely disappointing. Here the popular practice rests on a specific empirical claim — the critical period — that is genuinely contested inside the field, that the field's own anchor paper places at age 17, and that has never once been tested on a school.

So the programme starts by deciding what you are actually buying.

Decide first: are you buying fluency, or are you buying a school subject? These have different answers and conflating them wastes years.

• If you can offer genuine immersion — a bilingual strand, content taught in the target language, years of it — start as early as you can and commit to the whole school career. Every effect in the immersion literature is measured after five to nine years.
• If you can offer two or three lessons a week, the starting age is close to irrelevant and the hours are what matter. (evidence: critical period)

Ages 4–10 — do not buy an earlier start; buy intensity.

• The early-start experiment has effectively been run across Europe, and it did not work. German children who began English in Year 1 were ahead in Year 5 and behind by Year 7. Six hundred Swiss students starting English at 8 versus 13 showed no general advantage across five years — the early start helped only children who were already biliterate bilinguals with strong parental support. The Barcelona Age Factor project found later starters learn faster at equal hours. One large German study points the other way at Year 9, which is why this is mixed and not no-effect.
• What the critical-period evidence actually licenses is a statement about immersion, not about lessons. Ultimate attainment declines with age of first exposure — that gradient is in every dataset. The cliff is not: 2.3 million census records show no discontinuity, and the reanalysis of the famous 669,498-person study finds a continuous decline for every learner type except non-immersion learners, whose break sits at 18.6 — suspiciously close to the end of schooling.

Ages 5–18 — if you run immersion, run it whole-school and for years.

• The reassuring finding is the null. Under a genuine admissions lottery, immersion students did no worse in mathematics and science and +0.13 to +0.22 SD better in English reading at grades 5 and 8. The standard objection — that time taught in another language is stolen from the curriculum — does not survive randomisation. (evidence: immersion)
• Know what has never been measured. That lottery tested English reading, English maths and English science. It did not measure partner-language proficiency at all. The best available answer to "how well do immersion students speak it?" is a grade-D narrative review: high functional fluency, receptive skills near native, production persistently non-native.
• If you serve English learners, expect the metric to look bad before it looks good. Two-language students reclassify as English-proficient more slowly in elementary school and finish ahead by the end of high school.

Ages 10–18 — teach the forms explicitly, and keep them producing language.

• Teach grammar explicitly. This is the opposite of what this archive concludes about grammar in English, and the difference is structural: L1 grammar teaching is a transfer claim that fails, L2 form instruction is an acquisition claim that succeeds. Explicit beats implicit across four meta-analyses, and in the one that tested it, explicit instruction improved spontaneous production, not merely controlled knowledge. (evidence: comprehensible input vs instruction)
• Correct errors with prompts, not recasts. Push the learner to self-correct rather than reformulating for them. In the only meta-analysis restricted to real classrooms, feedback effects were significant, durable, largest on free production, larger for prompts, and larger for younger learners.
• Run a large extensive-reading programme — budget d ≈ 0.46, the estimate that has a control group, not the 0.71 that does not. Curate the library and attach light accountability: the 2025 meta-analysis found effects larger when text choice was limited and larger when accountability was added, which is not what "free voluntary reading" means.

What to refuse to spend money or time on

• Moving a low-dose primary language programme earlier, justified by the critical period. Two of the three best European comparisons find nothing or worse; the anchor paper's own breakpoint is at 17; and none of the critical-period literature studied a school.
• Promises of native-like attainment. Not from school instruction at any starting age. In the one study that scrutinised near-native bilinguals properly, no late learner and only a few early learners performed inside the native range — and that sample was selected for sounding native.
• The cognitive-benefit argument for early bilingualism. It is a far-transfer claim, the children's meta-analysis finds nothing, and attaching it makes a language programme falsifiable on a claim it never needed.
• European CLIL evidence as a reason to adopt anything. Its advantages are attributable to selection and extra hours; there is no CLIL lottery anywhere.
• Any L2 effect size above ~0.8 quoted without naming the test. A d of 1.06 measured on the structure that was just taught, the same week, on undergraduates, is a fact about the test.
• Krashen's input hypothesis as an operational plan. Reading a lot works. "Provide comprehensible input and acquisition follows" has been unanswered as a criticism since 1984 because it specifies no condition under which it would be wrong.
• Task-based programme evaluations quoting d ≈ 0.93. That number pools implementation evaluations by implementers, and reports "positive stakeholder perceptions" as a finding. This archive already knows what practitioner satisfaction is worth: over 90% in a randomised trial that returned −0.02.

The measurement problem in this subject

Every subject here has one. Foreign language has two, and they point in opposite directions.

The instructed-SLA literature measures what it just taught. The pooled effect of form-focused instruction is g = 1.06 — larger than anything credible elsewhere in this archive — and it moves significantly with the mode of the outcome measure, the setting, and the intensity. The tell is in the corrective-feedback meta-analysis: laboratory effects exceed classroom effects, and shorter treatments produce larger effects than longer ones. A real instructional effect grows with dose. An effect that shrinks as you teach longer is measuring freshness, not acquisition.

The age literature measures a grammar quiz. The 669,498-participant study everyone cites used 95 retained items of grammaticality judgement and sentence–picture matching, taken online by people who chose to take a viral quiz about their own native language. The one large study using real four-skill state proficiency exams — 56,024 adult immigrants sitting Dutch speaking, writing, listening and reading tests — finds a monotonic decline beginning around arrival age 25, which looks far more like ordinary cognitive ageing than a language-specific window.

And then the immersion literature measures the wrong language entirely: the only lottery in the subject reports English reading, English maths and English science, and never asked how much Spanish or Mandarin anyone learned.

The hereditarian bottom line for a founder

The premise does uneven work here, and it is worth knowing where.

1. The age-of-acquisition literature is the archive's premise in its purest form. Nothing is randomised — you cannot assign a birth date or an age of immigration. Families who emigrate when their children are small differ in education, resources, motivation and in heritable language aptitude, and they pass all four along together. Note the direction: those confounds push toward finding age effects, and the sharp critical period still does not appear.
2. The classroom-start evidence escapes the premise and still finds nothing. German and Spanish starting ages were set by state policy and birth cohort, not by parents. Passive gene–environment correlation cannot generate that contrast — and the early start still bought no durable advantage in two of three large samples. That is the most important fact in this subject, and it is the one nobody quotes.
3. The immersion lottery is genuinely clean, and its effect is a quarter of the observational estimate. Self-selected immersion studies give +0.33 to +0.45 on English reading. The lottery gives +0.13 to +0.22. That gap is the selection, measured.
4. Every "it works for the right kind of child" finding here is a heritable category wearing an aptitude–treatment interaction's clothes. The Swiss study's early-start benefit appeared only in "biliterate simultaneous bilinguals with substantial parental support"; the L2 instruction meta-analysis finds onset proficiency moderates the effect. Neither licenses allocating instruction by family type.

Nothing in this subject bears on g, and the one place a g-adjacent claim is made — the bilingual cognitive advantage — is a far-transfer claim that the children's meta-analysis does not support.

Confidence

Open questions a founder should watch

• Nothing in the age literature is randomised, and the obvious trial has never been run: schools assigned to an earlier or later start at matched total hours. Europe reorganised primary curricula around the answer without it.
• No lottery has ever measured second-language proficiency. The single most valuable study in this subject would be the Portland design plus a partner-language test.
• Two large German studies disagree about early starts — Jaekel finds the advantage reversing by Year 7, Porsch finds it lasting to Year 9. Nobody has reconciled them.
• The head-to-head everyone argues about does not exist: input-rich versus form-focused L2 instruction, equal hours, a school year, school-age children, an independent proficiency test.
• Almost nothing in this subject reaches children under about 10 on any credible design. The instructed-SLA corpus is undergraduates; the classroom-feedback synthesis is 15 studies; the early-start comparisons measure receptive tests, not whether anyone can hold a conversation.
• Durability past a school year is unmeasured everywhere. "Delayed post-test" in this literature means weeks.