Learn · Qualitative methods
Qualitative research did not arrive as one method. It grew from arguments about three questions: what counts as knowledge, who may interpret experience, and how to make interpretation credible. Large language models reopen all three. Here is the story in fifteen short stops, and what changes at each.
The three questions to hold as you read:
An LLM can process language at scale. It has no lived relationship with participants, no stable position, and no accountability for an interpretation. Keep that in view.
Early ethnography documented cultures through prolonged observation, giving us field notes and the researcher as an instrument of inquiry. It also let Western researchers speak for people who had little control over how they were represented.
LLM parallel: a new distant interpreter. A model can summarise voices from cultures its training data covers unevenly, while hiding the structures that shaped the summary.
Life histories, case studies, and urban ethnography made meaning and process the unit of interest, not just measurable traits. The rich material was often dismissed as too local to generalise.
LLM danger: scale can fake generalisability. Processing 10,000 narratives does not make them representative, or the reading of them valid.
Experience cannot be separated from meaning, interpretation is situated, and language builds reality rather than just describing it. Quality here rests on coherence, reflexivity, and fit between claims and material.
LLM contradiction: a model is extremely good at producing coherent interpretations. Coherence is not evidence that it understood the participant's world.
Glaser and Strauss showed theory could be built from data through coding, constant comparison, and theoretical sampling. The label later got attached to studies that coded interviews without any of it.
LLM danger: a model can generate a whole theory before you have engaged with the data, reversing grounded theory. An early structure then anchors everything you read next.
Constructivist, critical, feminist, and participatory traditions rejected the idea of one neutral method. Reflexivity, standpoint, and power moved to the centre. Terms like saturation and triangulation started travelling between traditions as generic badges of rigour.
LLM problem: most general models default to a methodological blend, mixing terms that co-occur in publications even when the underlying ideas are incompatible.
Software brought searchable transcripts, code hierarchies, and linked memos. It stored and retrieved your analytic work. It did not interpret the data.
The shift now: an LLM also generates analytic content. The move is from computer-assisted organisation to machine-generated interpretation, and that is a different thing.
Who benefits, who controls the categories, whose language is privileged, who owns the knowledge. Participants moved toward co-researcher roles.
LLM risk and possibility: models can re-centralise authority by importing dominant-language categories and normalising minority expression. Under community control, local models could instead support translation and participant-led analysis.
Form matters as much as content. Pauses, sequence, audience, and context do social work that plain text loses.
LLM limitation: models are strongest when text is treated as semantic content. They are weak on hesitation, timing, irony, code-switching, and silence. A cleaned transcript is already one transformation; the model adds another.
Braun and Clarke gave an explicit, flexible framework, later split into reflexive and coding-reliability versions. Today the label covers very different procedures.
LLM problem: models naturally produce topic summaries, and topic summaries get mistaken for themes. A theme is an interpretive claim about the research question, not a cluster of frequent topics.
COREQ (2007) and SRQR (2014) asked for transparent reporting of team, sampling, analysis, and the link between claims and material. The risk is that a checklist becomes a retrospective compliance exercise.
LLM two ways: a model makes it easy to reconstruct a plausible methods section that never happened. The same technology can instead preserve a prospective record: prompt logs, versions, human edits, rejected suggestions, claim-to-transcript links.
Social life moved online, and so did method: netnography, social-media research, digital traces. Availability is not permission; a public post can still be sensitive and context-dependent.
LLM danger: a model can strip statements from platform culture, thread history, identity performance, and timing, turning situated discourse into loose fragments.
The pandemic made video interviews, online focus groups, and automatic transcription routine. It also spread digital exclusion, uncertain data location, and third-party handling of voice data. This built the infrastructure generative AI walked into.
Researchers began using models for interview guides, summaries, initial coding, theme generation, and drafting. Early studies found both enthusiasm and serious worry about privacy, bias, and the absence of norms. One analysis of interviews with Rohingya refugees found annotation errors that were not random across participant characteristics, meaning model use could produce systematically biased inferences.
Studies began comparing human and LLM analysis. A consistent pattern appeared. Models are relatively good at broad descriptive coding, locating recurring content, organising large text, and applying a fixed codebook. They stay weak at latent interpretation, cultural meaning, emotional nuance, reflexivity, the significance of silence, and telling a good theme from a merely plausible one.
Recent work evaluates structured human-AI pipelines rather than one-shot prompting, splitting transcription, coding, theme development, and human adjudication into explicit stages. A 2026 scoping review of 75 LLM qualitative studies found coding and theme identification dominant, one provider's models in most of them, and little guidance in existing standards for documenting AI use.
The question has changed. Not whether qualitative researchers will use LLMs, but what role the model is assigned, what weight its output carries, and how human accountability is preserved.
The field's central weaknesses predate AI. Labels used without the practice behind them. An invisible analytic path, where papers show final themes but not the discarded ones. Claims that a single quote cannot support. LLMs did not create these. They can amplify every one, because a fluent model makes a missing method look present.
Keep your method and your words aligned, from protocol to publication.
A reviewer asks which paradigm your study sits in, and you notice your methods section borrowed terms from three. That is not carelessness: the words travel between traditions and shed their assumptions on the way, so saturation, triangulation, and member checking end up as badges rather than commitments. The fix is to name your approach and keep your terminology consistent with it. MethodVahti checks that your stated method, your terms, and your written claims line up, and flags combinations that do not belong together, like reflexive thematic analysis paired with intercoder reliability. It flags the mismatch. It does not pick your epistemology. That stays yours.
See MethodVahtiSee also using LLMs on interview data ethically and AI qualitative coding, human-reviewed. More guides are on the Learn page.