The field grew, the programme did not: differential growth and substrate portability in artificial intelligence and education research, 2016–2026

Abstract Reviews of a field in transition routinely read a falling share of records as evidence of decline. Where annual output rises more than fivefold that inference fails: a research form can lose four-fifths of its share while publishing more papers each year than before. This study reconstructs the artificial-intelligence-in-education literature across a decade — 1,218 records from Scopus, Crossref and Google Scholar, screened under a PRISMA 2020-aligned protocol and split at ChatGPT’s release into pre-generative (2016–2022, n = 333) and generative (2023–June 2026, n = 885) periods — converting every prevalence figure into an annualised count evaluated against pre-specified criteria with exact rate-ratio intervals. A growth-adjusted index G expresses each form’s growth as a fraction of the field’s own 5.32-fold increase in annual output; G is algebraically the prevalence ratio, so the share ratio is already growth-adjusted and the share difference is what misleads. The result is differential growth rather than displacement. Educational data mining contracts (rate ratio 0.31, 95% CI 0.15–0.58), as does learning analytics (0.67, 0.46–0.95), though that contraction is design-dependent and reverses under a retrieval exposing both periods to both vocabularies; intelligent tutoring is statistically flat (0.96, 0.64–1.42); three further forms grow at a fifth to three-fifths of the field’s rate, one keeps pace, and four outgrow it. Manual coding of a stratified sample (n = 200, κ = 0.80) supplies the code-level series a term-based instrument cannot recover: constructs presupposing instrumented longitudinal data return G = 0.18 and 0.52, those that do not return 1.12, 1.15 and 2.67. Substrate portability, not functional definition, predicts survival.