Metadiscourse in ChatGPT-generated and human-written research articles in linguistics

Abstract

ChatGPT has served as a reference for researchers seeking to generate scholarly articles since its emergence in 2022. Its increasing use in academic writing necessitates exploring how ChatGPT generates discourse compared with human authors. This study investigates the use of interactive and interactional metadiscourse in 100 human-authored and 100 ChatGPT-generated English-language linguistics research articles. Using AntConc, metadiscourse markers were identified and analyzed, drawing on Hyland’s (2005) model, revealing both convergences and divergences between the two corpora. Transitions occur the most frequently among interactive markers, reflecting ChatGPT’s ability to emulate human discourse organization. Evidentials are commonly used in both datasets; however, the striking difference is that the AI corpus often contains inaccurate and fictitious citations, raising concerns about the trustworthiness of its content. Moreover, the frequent use of frame markers in ChatGPT texts, compared to a human corpus, indicates a strong reliance on formulaic structure, leading to a lack of sophisticated argumentation. The sparse use of code glosses and endophoric markers in ChatGPT also suggests limited references within the text. Interactionally, hedges, boosters, and engagement markers are predominant in both corpora as conventions in academic writing. However, the ChatGPT corpus lacks self-mentions and fewer attitude markers, which typically reflect the impersonal nature of AI-produced texts. These findings offer practical pedagogical implications and significant insights into quality control, academic integrity, and plagiarism detection for all stakeholders amid the rapid advancement of AI.