This report draws on a reflexive, team-based qualitative analysis: three levels of coding generated through close reading, group discussion and iteration. A generative-AI model joined that process as a second analyst — and the exercise sharpened our sense of where the help is real and where it is illusory.
What AI did in this study
The model produced an independent re-coding of all 947 excerpts, drew exemplar and nuance quotes for each high-frequency theme, computed cross-quadrant valence patterns at corpus scale, and offered counter-readings the human team had backgrounded. Throughout, the team's original coding remained the spine of the analysis. The AI's outputs were material to interrogate and refine a framework — not a replacement for it.
We used AI to interrogate, not replace, our reflexive thematic analysis. The model helped us work across the full set of excerpts, while the human team retained responsibility for the analytical framework, interpretation and claims presented in this report.
What AI did well
AI is well-suited to corpus-scale operations and data visualisations. Cross-tabulating participant data across categories of questions, surfacing thematic patterns from multiple scales of reference, computing tag-coverage asymmetries, drafting counter-readings from explicit theoretical positions — all of it happens rapidly, and across the whole corpus without sampling.
This matters for rigour. Qualitative rigour often depends on sampling adequacy; for comparative pattern-detection, AI removes the sampling trade-off. The scissor finding (AI's preferred term) was a good example: computing a clean 4×3 valence matrix over hundreds of quotes is exactly the kind of pass a small human team would usually have to approximate.
What AI cannot do
Several limits are structural, not provisional — they will not be fixed by a better model.
- No interview contact.The model heard no one speak. Tone, hesitation, laughter, recovery, the work a silence does in a conversation — all invisible. Reflexive analysis treats the researcher's embodied presence in the interview as part of the data.
- Not genuinely independent.An “independent” re-coding is a trained artefact of prior exposure to GenAI discourse. Convergence with the human framework may reflect a shared training distribution rather than validation.
- Limited reflexivity.The model can imitate reflexivity — this section is an example — but the self it reflects on has no continuity across sessions. It cannot notice that its own discomfort with a quote might mark a blind spot, because it carries no discomfort between conversations.
- No longitudinal relationship.The team has been with these participants for months. The interview relationship is part of what makes the data meaningful; the model arrives after that relationship has finished doing its work.
- No local knowledge.Sector-specific context — Australian higher education, NSW education politics, UAE infrastructure, the Chilean literary canon — is unevenly present in training. The model mis-calibrated some references; only local readers will catch which.
Where AI complements human work
The natural fit is the role of second analyst: running cross-cutting calculations, producing counter-readings that surface what the primary coding backgrounds, drafting initial codebooks for human refinement, and computing inter-coder-style convergence against a human-authored framework. Used this way, AI expands the rigour of the cross-cutting passes that small teams usually have to sample, and surfaces patterns a small analysis might miss.
Used as the primary analyst, it strips out exactly the part of qualitative inquiry that gives the work its claim to insight: the researcher's situated, reflexive relationship to participants and to their own discomfort.