Ethical Thematic Analysis with ChatGPT: Step-by-step Tutorial

Ethical thematic analysis with ChatGPT hinges on one practical idea: keep tight control of the coding workflow so the model can’t “shortcut” to ready-made themes, and preserve traceability from every code back to the original participant quotes. The payoff is credibility—results can be treated as trustworthy and valid—because the process stays transparent and bias is actively managed rather than outsourced to an AI output.

The method is built around a three-stage thematic analysis workflow. Stage one creates a foundation of detailed, descriptive initial codes. These codes must be granular and grounded in what participants actually say, not abstract labels. The workflow also demands breadth: the coder should ask for lots of codes so every part of the transcript is captured, rather than letting the model return a small set of broad categories that resemble themes too early. Crucially, the prompt should require a full list of quotes associated with each code. Without quote-level traceability, later reporting becomes guesswork and transparency collapses.

Stage two refines and organizes those initial codes through “focus coding.” Here, the transcript-level outputs are transferred into Microsoft Word (or similar external tools like Excel) so the researcher can manage the audit trail. Each participant’s material is kept separate using distinct color formatting in Word, creating a visual map of where every code and quote came from. Focus coding is then done primarily manually: codes are grouped into higher-level clusters that represent patterns or trends, not final themes. This manual step helps prevent the model from collapsing the process into a theme list. When duplicates appear across participants—codes that mean essentially the same thing but use different wording—the researcher consolidates them by creating a new code description and attaching the original code names as comments, preserving the lineage back to the original quotes.

Stage three develops themes from the focus-coded structure. The recommended approach is manual again: themes and sub-themes are organized to communicate what the data supports, ensuring the reader understands the findings as well as the researcher does. ChatGPT can be used at this stage only if the researcher chooses, but the emphasis remains on human judgment for the final narrative structure.

To keep the workflow ethical, the tutorial repeatedly warns against two shortcuts: asking ChatGPT to analyze transcripts thematically in one step, and disclosing the study’s research questions or exact aims in the prompt. Revealing those details encourages the model to steer toward predetermined answers. Instead, the prompt should provide context (e.g., the transcript is from teacher interviews about remote teaching) while withholding the specific research question framing. The workflow also recommends uploading or pasting one transcript at a time (even with ChatGPT Plus) to reduce confusion and maintain consistent prompting.

Overall, the tutorial’s core message is that AI can assist coding, but ethical thematic analysis requires the researcher to control inputs, enforce quote-level traceability, and do the key organizing and theme-making steps themselves—using Microsoft Word as the transparency backbone and manual focus coding as the safeguard against bias and premature conclusions.