Can You Use ChatGPT for Thematic Analysis? A Peer-Reviewed, Step-by-Step Answer

The qualitative community is divided about generative AI. Some researchers refuse it on principle; others quietly paste transcripts into a chatbot and hope. Both positions are wrong — the first forfeits real gains in transparency and speed, the second forfeits rigour. The 2025 paper in the International Journal of Qualitative Methods charts the third path: a complete, documented, peer-reviewed process for AI-assisted thematic analysis.

Why earlier attempts fell short

Studies before this one shared three limitations: they did not follow all six steps of an established thematic analysis approach; they gave the AI no contextual information about the research, such as the aim and research questions; and they never informed the AI of the researcher’s methodological considerations. The result was AI output disconnected from the study’s intellectual architecture — and findings, such as Prescott and colleagues’, that human-generated codes were more reliable than AI-generated ones. The remedy is not less AI; it is better inputs.

Step 1: Familiarise the AI before it sees a transcript

The heart of the process is what you give ChatGPT before analysis begins: your research focus, problem and context; your theoretical underpinning and philosophical position; your research questions, objectives and aim; a methodology summary; participants’ profiles; how the data were collected; and the six steps of systematic thematic analysis themselves. Then ask the AI to confirm it understands and is ready. This mirrors what a good supervisor does with a new research assistant — and it is what makes every later output traceable to your study rather than to generic patterns.

Steps 2–4: Keywords, codes and themes with quality criteria

Keywords are selected on the 6Rs: realness, richness, repetition, rationale, repartee and regal — criteria that ensure the robustness and relevance of extracted data. Codes are developed on their own 6Rs: robust, reflective, resplendent, radical, relevant and righteous, so each code reflects the meaning of its grouped quotations and keywords in service of the research questions. Themes are organised on the 4Rs: reciprocal, recognizable, responsive and resourceful, with the researcher supplying the keywords, codes, research context, aims and theoretical underpinnings so the AI groups codes by their genuine inter-relationships. At every stage, you instruct — the AI proposes, you evaluate against the published criteria and dispose.

Steps 5–6: Concepts and the conceptual model

In conceptualisation, you instruct the AI to define new concepts from the identified keywords, codes and themes, emphasising theoretical conceptual clarity — definitions that meaningfully apply to existing theories. In the final step, the AI helps synthesise concepts into a framework of arrows and boxes, establishing the significance of the relationships, addressing research gaps, and extending existing theories. The output is the same destination as manual STA: a named conceptual model.

Does it actually work?

The paper tests the process as a case study: it re-analyses the data set of a published manual systematic thematic analysis (a study of scan-and-go shopping apps in England) with ChatGPT, using the same research context, methodology, and theoretical underpinnings — and compares outcomes step by step. This head-to-head design is precisely what makes the process citable in your own methods section.

What this means for your ethics and integrity statements

Because every prompt is documented, AI-assisted STA produces an audit trail most manual analyses never have. This helps to reduce human bias and improves accountability and transparency. Check your target journal’s AI disclosure policy, state exactly which steps the AI supported, and keep your prompt log — transparency converts a perceived risk into a methodological strength.