Thematic analysis with ChatGPT | PART 1- Coding qualitative data with ChatGPT

The central takeaway is that ChatGPT can speed up qualitative thematic analysis—especially the early “coding” stage—when researchers treat it as an assistive tool rather than an automatic analyst. The workflow described centers on feeding the model clear excerpts from qualitative data and asking it to generate candidate codes and code labels that can then be reviewed, refined, and organized into a coding framework. That human-in-the-loop step matters because thematic analysis depends on interpretive judgment, not just pattern matching.

The coding process starts with preparing the raw material: selecting manageable text segments (such as interview responses or open-ended survey answers) and ensuring the prompts request outputs that are usable for qualitative work. Instead of asking for a single “theme,” the approach emphasizes granular coding—producing short, descriptive codes tied to specific excerpts. This makes it easier to check whether the codes fit the data, whether multiple excerpts should share the same label, and whether the coding scheme is becoming too broad or too narrow.

A key practical insight is prompt design. The transcript highlights that effective prompts specify the task (e.g., generate codes), the unit of analysis (the excerpt), and the desired output format (e.g., a list of codes with brief definitions and/or example mappings to the text). When prompts are vague, the model tends to produce generic categories that don’t track closely to the participants’ wording. When prompts are structured, the output becomes more auditable—researchers can trace each code back to the underlying quote and decide whether to keep, merge, split, or discard it.

The workflow also addresses consistency. Because qualitative coding can drift across time, researchers can use ChatGPT to propose codes for multiple excerpts using the same instructions, then compare results across segments. That supports a more systematic approach to building a codebook. Still, the responsibility for final decisions remains with the researcher: codes must be checked against the full dataset, and themes should emerge through iterative refinement rather than one-pass automation.

Finally, the transcript frames ChatGPT as a way to reduce the mechanical burden of initial coding—turning long transcripts into organized, reviewable units—while preserving the interpretive core of thematic analysis. The promised “Part 1” focus on coding suggests a staged method: first generate and validate codes, then later aggregate them into themes. In that sequence, the model’s strength lies in drafting and structuring, while the researcher’s strength lies in theoretical alignment, reflexivity, and ensuring the analysis remains grounded in the data.