Work

Stance in argument: Comparing GPT-4o essays and second language learner essays

Year: 2026
Type: article
Abstract: Previous work has identified the differences in the overall frequency of stance marker use in human- vs. artificial intelligence (AI)-produced argumentative essays without clarifying whether such differences reflect distinct argumentative functions or different realizations of the same function through different stance-making choices. This corpus-based study, therefore, compared argument structure and stance marker uses in relation to argumentative functions. The corpus contained 125 timed learner essays and 125 GPT-4o-generated essays in response to the same prompt. Sentences were coded into six Toulmin argumentative elements and four stance categories. Between-group analysis showed that GPT-4o predominantly presented two-sided complete arguments, while learner essays were predominantly one-sided. Learners used more self-mentions, while GPT-4o employed more attitude markers, hedges and boosters. When presenting Claims, Data, and Rebuttals, learners foregrounded themselves as evaluators, whereas GPT-4o did not. The same pattern of self-mention and attitude marker use was observed in essays generated by other AI models in a cross-model check, but hedge and booster use showed cross-model variations. Such results suggest that the between-group differences reflect both argumentative structure and within-function realization variations, providing a nuanced understanding of authorial positioning in human- and machine-generated texts. Implications of such results for writing assessment and instruction are subsequently discussed.
Language: English
Open Access status: diamond
OpenAlex ID: W7204835822
FWCI:
Cites:
Cited by:
Related to: 9