Beyond Performance: A Mixed-Methods Evidence Synthesis of Generative AI–Mediated EFL Education

Background: Generative artificial intelligence (GenAI) now mediates much of English-as-a-foreign-language (EFL) instruction, yet fluent AI-supported output is routinely reported as if it demonstrated language learning. Whether reported gains reflect independently available, durable competence remains an open validity question. Objectives: To synthesise quantitative and qualitative evidence on GenAI-mediated EFL education (2023–2026) and to test how far reported outcomes support inferences of independently demonstrated, transferable, and durable learning rather than assisted performance. Methods: A convergent segregated mixed-methods systematic evidence synthesis of a protocol-governed, integrity-verified corpus of 52 reports (50 included studies; ≈4,060 participants). Every outcome was classified on a five-level performance-to-learning evidence ladder and by assessment condition. Quantitative evidence was synthesised without meta-analysis (SWiM) after none of five prespecified pooling sets met release gates; qualitative evidence underwent thematic synthesis; strands were integrated through a joint display and calibrated meta-inferences. Results: Evidence concentrated at perceptual (16 studies) and process (13) levels; only five studies measured immediate independent performance or delayed retention, and none tested transfer. Where unassisted-format contrasts were estimable, effects favoured GenAI-supported instruction (Hedges g 0.62–2.09), alongside one null and one mixed-direction study. Six themes located benefit in orchestrated scaffolding, verification routines, and feedback literacy, while documenting dependence, integrity tensions, and miscalibrated trust. Conclusions: GenAI reliably improves supported and immediate task performance, but the field has not yet demonstrated that these gains endure as independent competence. Claims, assessment design, and reporting should be calibrated to the assessment condition under which outcomes were produced.