Thematic analysis with ChatGPT | PART 2- Coding qualitative data with ChatGPT

Qualitative thematic analysis with ChatGPT doesn’t replace the core work of coding—it mainly accelerates the messy middle: reorganizing a large list of initial codes into workable groups. After generating initial codes with ChatGPT, the next step is to sort those codes into meaningful clusters (not yet final themes), then tighten the categories so they align with the study’s research questions—especially when the analysis includes both positive and negative influences on outcomes like job satisfaction.

The workflow starts with a practical audit of the coded material. Codes are color-coded so researchers can trace each code back to the interview it came from. Just as important, quotes must remain attached to codes; without the original quotations, later theme-building becomes hollow because there’s no evidence to support claims. The transcript emphasizes using search tools (e.g., Microsoft Word “Find”) to locate the quotes for any given code name, since ChatGPT reorganization will require repeatedly checking that the right evidence is still connected to the right label.

Before reorganizing, researchers also need to ensure the code list is complete and faithful to the data. If ChatGPT missed a relevant angle—such as a code that captures a limitation or challenge—researchers should add it manually. The guidance is blunt: too many codes are preferable to too few, because ChatGPT can only classify what it sees. This stage also involves scanning code summaries and deciding whether they fit the intended direction of the analysis (for example, the transcript’s focus includes both positive and negative factors affecting job satisfaction).

For grouping, the transcript recommends a hybrid approach: do some manual organization, but use ChatGPT to speed up first-pass sorting. A sample prompt assigns every code to one of three buckets: “challenges” (codes implying negative effects on job satisfaction), “positive factors” (codes implying positive effects), or “other” (codes that don’t clearly fit either). ChatGPT can be fast, but accuracy still needs human correction. The transcript shows this with misplacements—like a code that relates to “desire for something new” being incorrectly treated as a challenge—requiring researchers to move codes to the right group after checking the associated quotes.

Once codes sit in the right buckets, the next refinement is “clean-up” within each group: fix category errors, keep wording consistent, and only merge or rename codes carefully. Renaming creates a traceability problem—if a code label disappears, the researcher may lose the ability to find its quotes later. The workaround is to preserve the old wording in a comment or note before combining codes, so the evidence remains retrievable.

Finally, the transcript describes optional sub-grouping to support later theme development. For example, positive factors can be split into “personal” versus “external/workplace” influences, using additional ChatGPT prompts as brainstorming aids. The overall message is iterative and flexible: ChatGPT helps propose structure, but researchers must verify fit against the quotes, correct misclassifications, and maintain audit trails so themes can be defended with evidence.