Does AI-Assisted Thematic Analysis Actually Help? (2025)
Qualitati Research Team · 2026-06-23 · 7 min read
AI-assisted thematic analysis can meaningfully speed up qualitative coding without taking the interpretation away from the researcher. In a 2025 study of 18 qualitative researchers, an LLM-supported toolkit called DeTAILS produced codes and themes that aligned closely with researchers' own refinements while lowering perceived workload and keeping humans in control of every analytic decision.
What did the study test?
The study evaluated DeTAILS (Deep Thematic Analysis with Iterative LLM Support), a toolkit that folds large-language-model assistance into a workflow modeled on Braun and Clarke's six-phase thematic analysis framework. According to Sharma, Cochrane, and Wallace (2025), the system helps researchers generate and refine codes, review clusters, and synthesize themes through interactive feedback loops "designed to preserve analytic agency" — the researcher accepts, edits, or rejects every AI suggestion rather than receiving a finished codebook.
The authors recruited 18 qualitative researchers via academic and professional networks, stratified evenly by experience: 6 novices (under 1 year), 6 proficient (1–4 years), and 6 experts (more than 4 years). Each analyzed Reddit data using the tool, and the team measured both how closely the AI-supported output matched the researchers' final decisions and how the experience felt.
How accurate is AI-assisted thematic analysis?
Alignment between the LLM-supported outputs and what researchers ultimately kept was high across every phase. The authors report F1 agreement scores (where 1.0 is perfect overlap) per analytic stage:
| Analytic phase | Agreement (F1) |
|---|
| Concept outline | 0.98 |
| Global coding | 0.97 |
| Initial coding | 0.90 |
| Reviewing codes | 0.90 |
| Related concepts | 0.86 |
| Generating themes | 1.00 |
According to Sharma, Cochrane, and Wallace (2025), this pattern suggests the model is most reliable at structuring and clustering work, while the more interpretive early coding still benefits from heavier human revision — which is exactly where the tool routed researcher attention.
Does it actually reduce the work?
Yes, on the study's measures. Participants reported an overall NASA-TLX workload of 26.3 out of 100 (SD = 12.4) — low for an analysis task — with mental demand at 39, effort at 37, and a median frustration score of just 10. Self-rated performance averaged 73 out of 100.
Perceived usefulness was also strong: an overall rating of 4.21 out of 5, with 86% of responses landing at 4 or 5. Researchers rated the tool 4.4 on both "accomplishes tasks more quickly" and "makes it easier to do my job." On the AttrakDiff scale (out of 7), participants described it as structured (6.3), creative (5.9), captivating (5.8), and straightforward (5.7).
How does the workflow map to thematic analysis?
DeTAILS supports the full arc of a Braun and Clarke–style analysis rather than a single shortcut:
- Background research — orienting to the data and research question.
- Loading data — ingesting the corpus (Reddit threads, in this study).
- Coding — generating initial codes and refining them with AI suggestions the researcher can accept or override.
- Reviewing codes — clustering and consolidating into a working codebook.
- Generating themes — synthesizing codes into candidate themes.
- Report — assembling the analytic narrative.
Initial coding consumed the most time and differed significantly from other phases (p < 0.001), confirming that the interpretive heavy lifting stayed with the humans even when the AI accelerated the surrounding mechanics.
What this means for researchers
The headline is not "AI replaces the coder." It is that a well-designed human-in-the-loop interface can compress the tedious parts of thematic analysis — clustering, restructuring, drafting theme candidates — while leaving the researcher in charge of meaning-making. Participants in the study credited transparency and control with building their trust in the AI's suggestions, a recurring theme in human-in-the-loop thematic analysis work.
If you want to try this approach on your own transcripts, ThemeLens applies AI-assisted thematic analysis with researcher review at each step, and AI Surveys can collect the open-ended responses you'd later code. As always, treat AI output as a draft to interrogate, not a verdict to accept.
Frequently asked questions
Does AI-assisted thematic analysis bias the results?
The risk exists, which is why the study emphasizes analytic agency: researchers reviewed and revised every code. High agreement scores reflect outputs that survived human editing, not raw model guesses.
How many researchers were in the study?
Eighteen, evenly split across novice, proficient, and expert experience levels, each analyzing Reddit data.
Which thematic analysis framework does DeTAILS follow?
Braun and Clarke's six-phase approach, from familiarization through to reporting.
Can AI generate the themes by itself?
It can propose them — theme-generation agreement was 1.00 in this study — but the value came from researchers refining and validating those proposals rather than outsourcing judgment.
Primary source: Sharma, A., Cochrane, K., & Wallace, J. R. (2025). DeTAILS: Deep Thematic Analysis with Iterative LLM Support. arXiv:2510.17575 [cs.HC].
Last updated: June 23, 2026. This is an independent editorial summary of third-party research; figures are drawn from the cited preprint and are reported here for informational purposes.