Prompts, Pearls, Imperfections: Comparing ChatGPT and a Human Researcher in Qualitative Data Analysis

Abstract

The impact of ChatGPT and other large language model–based applications on scientific work is being debated across contexts and disciplines. However, despite ChatGPT’s inherent focus on language generation and processing, insights regarding its potential for supporting qualitative research and analysis remain limited. In this article, we advocate for an open discourse on chances and pitfalls of AI-supported qualitative analysis by exploring ChatGPT’s performance when analyzing an interview transcript based on various prompts and comparing results to those derived by an experienced human researcher. Themes identified by the human researcher and ChatGPT across analytic prompts overlapped to a considerable degree, with ChatGPT leaning toward descriptive themes but also identifying more nuanced dynamics (e.g., ‘trust and responsibility’ and ‘acceptance and resistance’). ChatGPT was able to propose a codebook and key quotes from the transcript which had considerable face validity but would require careful review. When prompted to embed findings into broader theoretical discourses, ChatGPT could convincingly argue how identified themes linked to the provided theories, even in cases of (seemingly) unfitting models. In general, despite challenges, ChatGPT performed better than we had expected, especially on identifying themes which generally overlapped with those of an experienced researcher, and when embedding these themes into specific theoretical debates. Based on our results, we discuss several ideas on how ChatGPT could contribute to but also challenge established best-practice approaches for rigorous and nuanced qualitative research and teaching.