The advent of ChatGPT, a novel AI-powered language model able to create grammatically accurate and coherent texts, has generated considerable concern among educationalists anxious about its potential to enable cheating among students and to undermine the development of critical thinking, problem-solving, and literacy skills. The similarities and differences between ChatGPT texts and human writing, however, remain underexplored. This study aims to bridge this gap by comparing the use of 3-word bundles in A-level argumentative essays written by British students with those generated by ChatGPT. Our findings show that ChatGPT essays contain a lower frequency of bundles but these have a higher type/token ratio, suggesting that its bundles are more rigid and formulaic. We also found noun and preposition-based bundles are more prevalent in ChatGPT texts, employed for abstract descriptions and to provide transitional and structuring cues. Student essays are characterized by more epistemic stances and authorial presence, crucial in persuasive argumentation. We attribute these distinct patterns in ChatGPT’s output to its processing of vast training data and underlying statistical algorithms. The study points to pedagogical implications for incorporating ChatGPT in writing instruction.

ChatGPT has created considerable anxiety among teachers concerned that students might turn to large language models (LLMs) to write their assignments. Many of these models are able to create grammatically accurate and coherent texts, thus potentially enabling cheating and undermining literacy and critical thinking skills. This study seeks to explore the extent LLMs can mimic human-produced texts by comparing essays by ChatGPT and student writers. By analyzing 145 essays from each group, we focus on the way writers relate to their readers with respect to the positions they advance in their texts by examining the frequency and types of engagement markers. The findings reveal that student essays are significantly richer in the quantity and variety of engagement features, producing a more interactive and persuasive discourse. The ChatGPT-generated essays exhibited fewer engagement markers, particularly questions and personal asides, indicating its limitations in building interactional arguments. We attribute the patterns in ChatGPT’s output to the language data used to train the model and its underlying statistical algorithms. The study suggests a number of pedagogical implications for incorporating ChatGPT in writing instruction.