Rhetorical divergence of AI: Comparing interactive and interactional metadiscourse in GPT-generated and human-authored research abstracts across disciplines

Abstract

Academic writing is inherently social and persuasive, relying on metadiscourse to mediate interaction between writers and readers. The emergence of Large Language Models (LLMs) such as ChatGPT challenges this dynamic, raising questions about their ability to reproduce human rhetorical complexity. Although previous studies have noted AI's limitations in interactional features, a comprehensive cross-disciplinary analysis of both interactive and interactional metadiscourse in empirical research article (RA) abstracts remains underexplored. Given that abstracts play a crucial gatekeeping role by shaping first impressions and influencing readership, understanding how AI deploys organizational and persuasive language in this genre is essential.

Addressing this gap, the present study offers a systematic comparative analysis of metadiscourse use in ChatGPT-generated and human-authored RA abstracts across Engineering, Medicine, and Applied Linguistics. Using a novel methodology that integrates corpus linguistics with computational semantic modeling (BERT embeddings and t-SNE visualization), the analysis moves beyond frequency counts to examine meaning in context. Findings indicate that AI-generated abstracts place greater emphasis on explicit textual organization through interactive metadiscourse, sometimes exceeding human usage of features such as code glosses and evidentials. However, AI demonstrates marked weaknesses in interactional metadiscourse, including limited stance modulation and reduced personal engagement, particularly through the underuse of self-mentions and inconsistent use of hedges across disciplines. Moreover, AI shows limited sensitivity to discipline-specific rhetorical norms. These results suggest that current AI models fall short of replicating the interpersonal and disciplinary sophistication of expert academic writing, offering implications for AI detection and academic writing pedagogy.