Metadiscourse Patterns in Human-Written vs. Generative AI-Authored Research Abstracts: A Comparative Corpus-Based Analysis

This study examines how metadiscourse markers differ in research article abstracts written by human authors and those generated by AI (ChatGPT). We compiled a balanced corpus of human-written abstracts from peer-reviewed journals and AI-generated abstracts using the same titles and publication contexts. Using Hyland’s interpersonal metadiscourse framework, we annotated interactive (text-organizing) and interactional (reader-focused) devices and compared their frequencies across corpora. Quantitative analysis (frequency counts normalized per 1,000 words; chi-square tests) revealed systematic differences: AI abstracts contained more structural (interactive) markers (e.g. transitions, frame signals) but significantly fewer stance and engagement markers (e.g. hedges, boosters) than human abstracts. Qualitative analysis of exemplar abstracts confirmed that AI-generated abstracts adopt a clearer, more impersonal tone, while human abstracts show richer personal voice and reader engagement. These findings align with recent studies showing ChatGPT’s abstracts are coherent but lack nuanced authorial presence. We discuss implications for academic writing, noting that AI can mimic formal structure yet may omit the rhetorical subtlety valued in scholarly communication. Limitations (e.g. single AI model, one genre) and future work are identified. Overall, this comparative corpus analysis highlights characteristic metadiscourse profiles of AI vs. human authorship in scientific abstracts and informs discourse research in the AI era.