Reflexivity in human-written and ChatGPT-generated English research article abstracts: A comparison of metadiscourse

Abstract Reflexivity, a unique feature of human language, is a key indicator evaluating the performance of ChatGPT in text generation. Comparing reflexivity in human-written and ChatGPT-generated texts could reveal how well ChatGPT could capture the fundamental features of human language. Using a self-built corpus and adopting a bottom-up approach and statistical methods, this study compares the reflexive language, metadiscourse, in human-written and ChatGPT-generated English research article abstracts. Results show that in both types of abstracts, metadiscourse fulfills three broad and eight specific discourse functions: Referring to text participants (Referring to writer, Referring to text), Describing text actions (Introducing, Arguing, Finding, Presenting), Describing text circumstances (Phoric marking, Code glossing). However, metadiscourse markers are much more prevalent in ChatGPT-generated abstracts. In addition, human-written abstracts employ metadiscourse markers mainly for writer-oriented introducing, while ChatGPT-generated abstracts for text-oriented introducing. Possible reasons for the similarities and differences are related to ChatGPT’s working mechanism, the training dataset, and writing rules learnt by ChatGPT. This research contributes to the development of large language models and artificial intelligence output detectors, writing instruction and practice, and metadiscourse research.