Effects of AI-generated feedback on l2 writing: a PRISMA 2020 systematic review of performance, revision, and engagement outcomes
TL;DR
The evidence indicates that GenAI feedback most consistently supports local accuracy, lexical development, revision activity, and learner engagement, whereas effects on syntactic complexity, higher-order composing, and long-term retention are mixed or conditional.
Abstract
Generative artificial intelligence (GenAI) tools can provide immediate feedback during second-language (L2) writing, but evidence on their effects across writing outcomes remains fragmented. This PRISMA 2020 systematic review synthesized peer-reviewed empirical studies published from 2023 to 2025 that examined GenAI-generated feedback for English writing in EFL/ESL contexts. Searches of Scopus, PubMed, and ScienceDirect identified 1,106 records; after 12 duplicates were removed, 1,094 records were screened, 71 full texts were assessed, and 20 studies met the eligibility criteria. Eligible studies were experimental, quasi-experimental, mixed-methods, qualitative, or design-based investigations reporting writing performance, revision, feedback engagement/literacy, or AI-feedback accuracy. Because study designs, writing tasks, outcome measures, and reported statistics were heterogeneous, statistical meta-analysis and pooled effect-size estimation were not conducted; findings were synthesized thematically and by effect direction. The evidence indicates that GenAI feedback most consistently supports local accuracy, lexical development, revision activity, and learner engagement, whereas effects on syntactic complexity, higher-order composing, and long-term retention are mixed or conditional. Comparative evidence generally supports a complementary model in which AI provides rapid, high-volume local feedback and teachers address global structure, disciplinary expectations, contextual appropriateness, and affective support. Key limitations of the evidence base include short intervention periods, small samples, heterogeneous measurement, limited longitudinal evidence, and uneven geographic representation. Future research should preregister screening procedures, use independent double screening with agreement statistics, report extractable effect sizes and delayed post-tests, and examine equity, transfer, and hybrid AI-human feedback models.