Generative AI in English language learning: what does the research say?

Research shows that generative AI tools such as ChatGPT can help language learning, but the benefit depends more on how they are used than on the tool itself. According to a meta-analysis published in 2025, the effect grows clearly when learners are taught how to use the tool, plan before the task and have their work assessed by a teacher. A systematic review published the same year finds that the most frequent problem is learners relying on the tool too much.

A meta-analysis: 25 studies, 2,431 learners

Li, Shadiev and Chiu (2025) combined 25 experimental studies in which generative AI was used in task-based language teaching; 2,431 learners took part in total. Learners who used AI did significantly better than those who did not; the average effect size (Hedges' g) was 0.75, which the authors interpret as a large effect. ChatGPT was the tool used most often; most tasks were about writing (64%), some about speaking (24%), and the effect was similar for speaking and writing.

The more important finding was three conditions that made the effect bigger or smaller. The effect sizes from the meta-analysis are below; by a common rule of thumb, values around 0.8 and above read as large, values around 0.2 as small.

The authors link this result to the tools used: a general-purpose tool not designed for language learning may require learners to know how it works, to plan their task, and to have a teacher involved in assessment. The learners' level of education and language proficiency did not change the effect significantly.

A systematic review: 49 classroom studies

Lee, Choe, Zou and Jeon (2025) examined 49 studies with real language learners published between January 2023 and December 2024. Most were done at university and on learning English; most relied on learners' self-reported data and measured perceptions such as attitude, self-efficacy and motivation. Studies that measured language development looked mostly at writing. So the firmest answer to “does AI improve English?” is, for now, about university students' writing; there is little evidence for children or for long-term use.

The review lists five roles for AI: feedback provider, tutor, thinking prompt, interaction facilitator and conversation partner. The problems are of two kinds. On the technical side, content quality and the difficulty of asking the right question; the review reports that the tools can produce inaccurate information and fabricated references that learners often fail to notice. On the educational side, the most frequent problem was over-reliance on the tool (22 studies), especially when learners focused on the tool's output rather than on the learning process. Pedagogical limits such as lower motivation and engagement and fewer chances to practise (21 studies), and concerns about academic integrity (15 studies), followed.

The limits of these findings

Both reviews stress that the field is very new. The meta-analysis covers only English-language, peer-reviewed journal articles from three databases; most studies were done at university, and the authors write that they found it hard to identify studies with rigorous designs. The review calls for long-term studies and studies with school-age learners. So read the results not as “AI works” but as “it can work when used well”.

What does this mean in class and at home?

The same principle applies to learners working alone: write your own paragraph first, then check the tool's suggestions one by one and know why you change something. For short paragraph practice, English writing practice gives small tasks that fit this cycle.

Sources: Li, Y., Shadiev, R. and Chiu, T. K. F. (2025), Applying Generative Artificial Intelligence to Task-based Language Teaching and Learning: A Systematic Review and Meta-analysis, TechTrends, https://doi.org/10.1007/s11528-025-01140-7; Lee, S., Choe, H., Zou, D. and Jeon, J. (2025), Generative AI (GenAI) in the language classroom: A systematic review, Interactive Learning Environments, https://doi.org/10.1080/10494820.2025.2498537. Both are open access; the full texts were read on 9 October 2026.

Frequently asked questions