Maria Kuteeva, Marta Andersson
Recent studies have started to examine differences in the use of stance expressions between human and GenAI writing, e.g. comparing student essays to those generated by LLMs. In professional research writing, stance expressions not only reflect authors’ attitudes and assessments but also function as manifestations of researcher’s positionality. However, to our knowledge, the impact of GenAI on the linguistic expression of researcher positionality, and stance in particular, has not received attention in the literature. To address this gap, our study zooms in on how ChatGPT-4 adopts and adapts stance with regard to knowledge claims in applied linguistics and pragmatics. We have systematically tested three kinds of prompting techniques in order to assess the model’s flexibility in generating stance expressions: a) zero-shot prompting with minimal input; b) few-shot prompting technique; c) role enactment prompting. Our analysis of ChatGPT-4 output in response to our prompts suggests that, overall, the model displays a rather rigid understanding of stance. The outputs are more acceptable for epistemic rather than attitudinal stance. Our results suggest that interactional prompting is the most effective strategy, as the model is guided to treat texts holistically and consider a range of nuanced elements, producing an output that better mimics genuine engagement with the text.