Who Keeps Interpretive Authority in AI Thematic Analysis?
Qualitati Research Team · 2026-07-15 · 7 min read
When an AI tool runs your thematic analysis, who actually decides what the data means? A 2026 study by Nyaaba and colleagues found the answer is clear: the AI structures the process, but interpretive authority stays with the human researcher. Their purpose-built tool, ITA-GPT, acted as a "procedural scaffold" — organizing the workflow and improving transparency — while the researchers exercised judgment by modifying, deleting, rejecting, inserting, and commenting on every AI suggestion.
What did the study test?
The study examined how experienced researchers actually work with an AI thematic-analysis tool, rather than whether the AI produces good themes on its own. According to Nyaaba, SungEun, Apam, Acheampong, Dwamena, and Zhai (2026), three experienced qualitative researchers used ITA-GPT — an Inductive Thematic Analysis GPT built around reflexive thematic analysis and verbatim coding principles — to analyze interview transcripts from teacher-education research in the Ghanaian context.
The work was guided by a framework the authors call HACITA (Human–Artificial Intelligence Collaborative Inductive Thematic Analysis). Crucially, the study focused on the analytic process, not the substantive findings about teachers. In other words, it asked a methods question: how is thematic analysis enacted when a human and an AI share the workbench?
How does an AI thematic-analysis tool actually support the work?
ITA-GPT supported four recognizable stages of reflexive thematic analysis, and enforced three quality guardrails throughout. Here is the breakdown reported in the study:
- Familiarization — structured initial reading and engagement with the transcripts.
- Verbatim in vivo coding — codes drawn directly from participants' own words.
- Gerund-based descriptive coding — action-oriented codes (e.g., "negotiating," "resisting") that keep analysis close to what people are doing.
- Theme development — grouping codes into candidate themes.
Alongside these steps, the tool enforced trace-to-text integrity (every code had to map back to the transcript), coverage checks (was the whole dataset accounted for?), and auditability (a reviewable record of how analysis progressed). These are exactly the transparency features reviewers and ethics boards increasingly ask for.
Who had interpretive authority — the AI or the human?
Interpretive authority stayed firmly with the human researchers. The study captured this through a rich evidence trail: interaction logs, AI-generated tables, and every researcher revision, deletion, insertion, comment, and reflexive memo. Analyzing those traces, the authors found that researchers did not passively accept AI output. They repeatedly exercised judgment through five recurrent actions:
- Modification — reshaping AI-proposed codes and themes.
- Deletion — removing suggestions that missed the point.
- Rejection — declining AI framings outright.
- Insertion — adding interpretations the AI never surfaced.
- Commenting — annotating and reasoning about the analysis.
The conclusion is worth stating plainly: ITA-GPT functioned as a scaffold that structured the workflow and enhanced transparency, but the meaning-making — the interpretive core of qualitative analysis — remained a human act. As the authors frame it, inductive thematic analysis was enacted "through responsible human AI collaboration."
What this means for researchers
The practical takeaway is that AI is most defensible in thematic analysis when it is designed as a process scaffold, not an autonomous coder. Three implications follow:
- Keep a decision trail. The study's credibility rested on logging modifications, deletions, and rejections. If you use AI to code, preserve the same audit record — it is your evidence of analytic rigor.
- Enforce trace-to-text. Every AI-generated code should point to a specific piece of data. Tools that let codes float free of the transcript invite exactly the drift qualitative reviewers distrust.
- Own the themes. Treat AI output as a first draft to interrogate, not a verdict to accept. The researchers in this study rejected and rewrote freely — and that is the point.
This mirrors how modern qualitative platforms are built. ThemeLens, Qualitati's AI thematic-analysis tool, is designed around the same principle: the AI proposes codes and themes with quotes traced back to the source, and the researcher edits, merges, and rejects them. If you also collect your own data, an AI Interviewer can feed clean, well-probed transcripts into that human-led analysis loop.
Limitations to keep in mind
This is a small, focused process study — three researchers working with education-interview transcripts from one national context. It was not designed to measure coding accuracy, inter-rater reliability, or generalizable effect sizes, and the authors are explicit that it centers analytic process over substantive findings. Read it as evidence about how human–AI thematic analysis unfolds and where authority sits, not as a benchmark of how "good" AI coding is in absolute terms.
Frequently asked questions
Can AI do thematic analysis on its own?
Not defensibly. In this 2026 study the AI scaffolded the workflow and improved transparency, but human researchers retained interpretive authority — actively modifying, deleting, and rejecting AI suggestions. The meaning-making stayed human.
What is ITA-GPT?
ITA-GPT is a purpose-built Inductive Thematic Analysis GPT used in the study, aligned with reflexive thematic analysis and verbatim coding. It supported familiarization, in vivo and gerund-based coding, and theme development while enforcing trace-to-text integrity, coverage checks, and auditability.
What is the HACITA framework?
HACITA stands for Human–Artificial Intelligence Collaborative Inductive Thematic Analysis. It frames thematic analysis as a shared process in which the AI structures the workflow and the human researcher exercises interpretive judgment.
How do I make AI-assisted coding rigorous?
Keep an audit trail of your edits, require every code to trace back to the transcript, and treat AI output as a draft you interrogate rather than accept. These were the practices that made the study's analysis credible.
Source: Nyaaba, M., SungEun, M., Apam, M. A., Acheampong, K. O., Dwamena, E., & Zhai, X. (2026). Human–Artificial Intelligence Collaborative Inductive Thematic Analysis. arXiv preprint. arxiv.org/abs/2601.11850
Last updated: July 15, 2026. This article is an independent editorial summary of third-party research and is not affiliated with or endorsed by the study's authors.