A review of 101 higher-education English-language studies maps five roles for AI and recurring evidence gaps

Wang and colleagues systematically review how artificial intelligence has been used in English-language teaching and learning in higher education. Their July 2026 Cogent Education article covers empirical ESL and EFL research published from 2020 to January 17, 2025. The team searched Web of Science, Scopus, ScienceDirect and ERIC under a PRISMA process. The scope is deliberately narrower than a general review of educational technology: participants had to be higher-education English-language learners or teachers, and the work had to report empirical evidence.

The searches returned 1,536 records. After 105 duplicates were removed, 1,431 titles and abstracts were screened. The reviewers sought 296 full texts, could not retrieve two, assessed 294 and excluded 193, leaving 101 studies. Non-English publications, non-empirical papers and studies centered on native-language learning or unclear learner profiles were excluded. Three authors screened independently after a ten-paper calibration exercise, with a senior researcher resolving disagreements. All included studies were appraised using the 2018 Mixed Methods Appraisal Tool.

The synthesis combined descriptive and thematic analysis with latent Dirichlet allocation. It organized applications into five functional families and five recurring educational roles: teaching assistant, learning companion, assessment and feedback provider, translation and comprehension support, and personalized learning or motivation support. Publication volume rose sharply across the review period, especially as generative AI became easier to access. Yet activity was regionally concentrated, language skills were studied unevenly, and mixed-method, student-focused designs were especially common.

Across the included literature, researchers often reported improved engagement, language practice, feedback access, writing support or learner confidence. The review also identifies repeated concerns: overreliance, inaccurate or variable feedback, limited cultural adaptability, inequitable access, privacy and ethical uncertainty. These patterns should be read as a map of reported findings, not a pooled effect estimate. A large number of favorable conclusions across small or short studies cannot establish that one AI approach reliably improves every learner's proficiency.

The review itself is bounded by English-language publication, four databases and a higher-education ESL/EFL focus. Its evidence base ends in January 2025, before later products and policies. Many original studies use localized convenience samples, brief interventions, self-report or quasi-experimental designs, and the literature is weighted toward writing. The article does not provide a meta-analytic causal effect, and publication patterns may favor novel or positive findings. Long-term retention, unaided performance and culturally responsive behavior therefore remain important gaps.

For Hong Kong universities and teacher education, the five-role framework can structure a disciplined pilot portfolio. Institutions can test feedback, conversation, translation and personalization separately across Cantonese, English and Putonghua, with teacher review and source checking. Research should include receptive and oral skills as well as writing, compare assisted work with delayed unaided performance, and report access across student groups. Teacher workload and the reliability of corrective feedback should also be measured rather than assumed. The review supports targeted experimentation, not blanket adoption: each role needs a defined learning purpose, local language evidence and an outcome that remains meaningful after the tool is removed.