Every qualitative researcher has had the thought by now: could AI just do the coding? The honest answer about AI in qualitative data analysis is neither the hype nor the horror story, it is a job description. Used as a fast, tireless assistant, AI genuinely shortens the distance between forty transcripts and a first view of the patterns. Used as the analyst, it will hand you confident findings that nobody can trace back to the data.
This article is a sober map of that line: where AI helps qualitative data analysis, where it quietly fails, and how to build a workflow that takes the speed without giving up the rigour.
What AI does well in qualitative data analysis
The strengths are real, and they cluster around the mechanical middle of analysis:
- Transcription and tidy-up. Automated transcription is now good enough that correcting it beats typing it, and turning messy speech into readable text is exactly the kind of patience-heavy work machines should do.
- Summarising sessions. A faithful summary of each interview helps you orient, brief stakeholders, and decide where to focus a close read. It is a map, not the territory, but maps are useful.
- First-pass clustering. Given a pile of excerpts, AI is decent at grouping similar statements and proposing candidate labels. As a starting structure to react against, this can save hours of shuffling.
- Retrieval. "Find every place anyone discusses pricing, including indirect mentions" is a question AI answers across forty transcripts in seconds, and it beats keyword search because it catches paraphrase.
- Playing devil's advocate. Handing over your draft themes and asking for counter-evidence and alternative readings is a genuinely good use: it is a cheap, ego-free disconfirmation pass.
Notice what these have in common. They are all first drafts and lookups, tasks where an error is cheap because a human checks the output. The model also brings one underrated virtue: it is as attentive on transcript thirty-eight as on transcript one, which no tired researcher can claim.
Where it quietly fails
The failures are less visible than the strengths, which is what makes them dangerous.
Hallucinated themes. A language model has absorbed a vast amount of research writing, and when your data is thin it will helpfully average the literature: themes that sound entirely plausible for a study like yours but are not actually in your transcripts. The output reads as insight and is really genre.
Quote trouble. Ask for supporting quotes and you may get light paraphrases presented as verbatim, composites stitched from two participants, or attributions to the wrong person. Every one of these is disqualifying in a report.
Lost nuance. So much of qualitative meaning lives outside the literal words: the hesitation before an answer, the sarcasm, the thing conspicuously not said, the difference between a first-time user and a power user saying the same sentence. Models flatten this, and they flatten it silently.
No provenance. Rigour in qualitative work means an unbroken trail from claim to code to excerpt to participant, the same audit-trail standard we describe in our guide to thematic analysis. Raw AI output gives you conclusions without a trail. A theme you cannot trace is not a finding, it is a vibe.
Instability. The same transcripts on a different day, or with a slightly different prompt, can yield a different code list. That is survivable for brainstorming and fatal for anything you call systematic.
A human-in-the-loop workflow that holds up
The workable pattern is not human versus machine, it is a strict division of labour where meaning stays human-owned.
- Read your data first. Familiarisation cannot be delegated; Braun and Clarke made it phase one for a reason. If your first contact with the data is a machine summary, the machine has already framed what you will notice.
- Let AI draft second, not first. Once you have impressions of your own, AI-generated candidate codes and clusters are something to argue with rather than something that anchors you.
- Treat output as a junior colleague's first pass. Check every suggested code against the excerpts beneath it, exactly as you would when building a codebook by hand; our guide to coding qualitative data sets out what a defensible code looks like. Expect to merge, rename, and reject.
- Enforce a verbatim rule. No quote enters a report unless you have found it, word for word, in the source. No exceptions, however plausible it sounds.
- Keep the audit trail yourself. Themes link to codes, codes link to excerpts, excerpts link to participants. If a tool cannot support that trail, it is a brainstorming aid, not an analysis tool.
- Spot-check by sampling. For each AI-assigned label, pull a random handful of tagged excerpts and verify them. Error rates vary wildly by code, and you want to know where before you trust anything.
One more obligation that predates all of this: consent. Participants agreed to research, not necessarily to their words being processed by whatever tool you like. Check what your consent covers, anonymise before processing, and know where the data goes.
The analyst is still you
None of this argues for abstinence. Refusing the speed on offer is as unserious as outsourcing your judgement to it. The tools will keep improving; what will not change is the obligation underneath qualitative work, that a human who has genuinely engaged with the data can stand behind every claim and show the trail when asked. Let AI do the tireless parts, keep the meaning-making, and when someone asks "how do you know?", make sure the answer is never "the model said so".
People of Research is a community and newsletter for UX researchers, behavioural scientists, and qualitative researchers working out what AI is actually for. Join us and get the good stuff in your inbox.