<p>Recent studies have pointed to detectible differences between academic writing produced by AI algorithms versus human authors. With the aim to shed further light on this emerging line of enquiry, we analysed the rhetorical structure and metadiscoursal features of 320 research article abstracts, half of which were generated by ChatGPT while the other half were written by human authors. Drawing on a genre-based analysis and as reported by Hyland (Disciplinary discourses: Social interactions in academic writing, Pearson Education Limited, 2000) move structure model, we confirmed and extended the findings of earlier studies. We found that the ChatGPT did not exhibit comparable to human writing variety in the combinations of moves, provided less background information and methodological detail than did the human writing, as well as tended to overemphasize findings and conclusions. Further, the ChatGPT was found to use fewer textual metadiscourse markers, such as transitions, compared to human writers; it also overused sentential frame markers, attitude markers, and verbs used to hedge and boost. We discuss our findings in light of current literature in the field, as well as their implications for chatbot designers, AI detection developers, writing instructors, and students writing for academic purposes.</p>