Clause complexing in research-article abstracts: Comparing human- and AI-generated text

Abstract

The ability of chatbots to produce plausible, human-like responses raises questions about the extent of their similarity with original texts. Using a modified version of Halliday’s clause-complexing framework, this study compared 50 abstracts of scientific research articles from Nature with generated versions produced by Bard, ChatGPT, and Poe Assistant. None of the chatbots matched the original abstracts in all categories. The only chatbot that came closest was ChatGPT, but differences in the use of finite adverbial clauses and –ing elaborating clauses were detected. Incorporating distinct grammatical features in the algorithms of AI-detection tools is crucially needed to enhance the reliability of their results. A genre-based approach to detecting AI-generated content is recommended.