A critical period for second language acquisition: Evidence from 2/3 million English speakers

669,498 native and non-native English speakers

Worldwide internet-quiz respondents, native and non-native English speakers, roughly ages 7-70; recruited virally (the quiz was shared more than 300,000 times on Facebook)

CROSS-SECTIONAL, not longitudinal: age of first exposure, years of experience and current age are separated statistically by a computational model rather than by follow-up, which is the paper's central methodological move and its central vulnerability. The outcome is a RESEARCHER-DESIGNED written grammar quiz - 95 "critical items" retained from 132 grammaticality-judgment and sentence-picture-matching items, where a critical item is one on which at least 70% of native English-speaking adults aged 18-70 agreed. No standardized proficiency instrument, no speaking, no listening, no writing. Self-selection into a viral "guess your native language" quiz is severe and unmeasured, and the authors acknowledge they cannot rule out that the sample interacts with the variables defining critical-period effects. Nothing here is randomised: age of arrival is chosen by families.

The best-fitting model (R2 = 0.89) involves a rate change beginning at 17.4 years - a decline in the RATE of grammar learning in late adolescence, NOT a cliff in early childhood. Learning is slow in absolute terms: most of it takes place over the first 10-20 years of exposure, and native speakers do not reach asymptote until around 30 years of age. The widely-quoted practical implication follows from combining those two facts rather than from a maturational cliff: because learning takes so long, "learners must start by 10-12 years of age to reach native-level proficiency". Seven-year-olds were already scoring very well. Immersion and non-immersion learners follow comparable learning curves once years of experience are controlled; non-immersion learners simply need more years. The paper is routinely misreported as evidence for a young-childhood critical period; its own estimate puts the rate change at the END of adolescence.

High as an observational design, though the confound here is selection more than heredity: people who arrive in an English-speaking country young differ from late arrivals in schooling, input quantity, identity, motivation and - relevantly for this archive - in heritable language aptitude, none of which is measured. Additionally, who takes a viral internet grammar quiz, and who finishes it, is itself selected on language interest and ability.

Directly re-analysed and disputed by van der Slik, Schepens, Bongaerts & van Hout (2022), who fit the same data separately by learner type and report that a CONTINUOUS decline fits monolinguals, bilinguals and early-immersion learners better, with a discontinuity appearing only for non-immersion and later-immersion learners at 18.6 and 19.0 years - which they interpret as a schooling effect. Hartshorne (2024) rebuts. Independent large-sample work on real state proficiency exams (Schepens et al. 2022) finds monotonic decline after an arrival age of about 25 rather than a late-adolescent breakpoint. So: the age-of-acquisition GRADIENT replicates everywhere; the SHARP BREAKPOINT does not.