LLM-assisted inductive thematic analysis of research interviews

Workflow & human oversight

Human-orchestrated qualitative analysis across coding and theme development, partially following the six thematic-analysis phases. The researcher supplies data, prompts and interpretation and compares outputs with earlier analyses; no autonomous tool-using agent is demonstrated.

Workflow level is not an evidence score; higher is not better evidence.

Problem

Can a language model assist interpretation-intensive qualitative data analysis without treating generated themes as equivalent to human understanding?

Approach

The author applies GPT-3.5-Turbo to interviews with videogame players and university data-science lecturers and compares resulting themes with prior researchers’ analyses.

Outcomes (as reported by the originating team)

The author reports that the model inferred most main themes identified in the earlier studies. This is an exploratory comparison on two datasets, not proof of equivalent interpretive understanding or a general productivity benefit.

Limitations

The study is explicitly exploratory and does not establish formal procedures. LLM pattern generation is not human interpretation; generalization, model variability and reflexivity require care.

Reproducibility

The paper identifies two open-access interview datasets and describes the experiment. Exact outputs may vary with model availability, version and prompts; this catalogue has not independently reproduced the analysis.

Adoption requirements

Appropriate consent and data-governance review before using an LLM, qualitative-method expertise, recorded prompts and model settings, and human checking against the original interviews.