A corpus-based comparison of interactional patterns in EFL student and LLM-generated essays

Insights into authenticity and audience awareness

Abstract

The emergence of ChatGPT has attracted increasing scholarly attention to the affordances and challenges of integrating large language models (LLMs) in EFL learners’ writing skills development. Although some studies have investigated AI-generated text detection and characteristics, research on LLMs in EFL education remains limited. This paper seeks to address this gap and advance learners’ critical understanding of AI applications. It reports on a small-scale study comparing genuine EFL undergraduate student essays, suspected LLM-generated essays, and essays generated using ChatGPT 4.0 and DeepSeek-V3. Drawing on Hyland’s (2005a, 2005b) framework of interactional metadiscourse, the study examined patterns of stance and engagement in the texts. Combining quantitative analysis with qualitative interpretation, the study identified consistent use of engagement as the defining characteristic of genuine student writing, whereas LLMs showed limited capacity to engage readers. Furthermore, although LLMs approximated human stance-taking, they nevertheless favored a detached style, contrasting with the more personal tone of student writing. We believe that these findings can inform both EFL instructors and students by offering insights that may contribute to a deeper understanding of the rhetorical limitations of LLMs in academic writing contexts.