How well can GenAI (GPT-4) provide written corrective feedback on English-language learners’ writing?

Providing written corrective feedback (WCF) on English-language learners’ (ELLs) writing is an efficacious, but time-consuming practice in English for Academic Purposes (EAP). Advances in generative artificial intelligence (GenAI), such as large language models (LLMs) may afford EAP language educators time-saving tools for providing WCF. In this study, we investigate the capability of GPT-4 Turbo, a LLM, in providing WCF on ELLs’ writing, and compare the feedback provided by GPT-4 with feedback provided by two experienced English-language educators. Data came from teacher and GPT-4 generated feedback on fifty ELL writing samples, collected from two classes at an intensive English programme for non-matriculated college ELL students. We explored the types and numbers of error marking codes utilized, the number of error codes used correctly and incorrectly, and the number of errors missed by the teachers and GPT-4. We found GPT-4 to be less reliable and accurate than the two trained language educators. Both teachers and GPT-4 used error marking codes correctly and incorrectly, as well as missed errors in student writing all together. We conclude, however, that if LLMs continue to improve at the current rate, with better prompt engineering, GenAI may soon allow language educators to share the time-consuming burden of providing WCF on students’ writing. This article was published open access under a CC BY-NC-ND licence: https://creativecommons.org/licenses/by-nc-nd/4.0/ .