Learn · Qualitative methods

What LLMs can do in qualitative research, and where they stop

An LLM can summarise, compare, and suggest language. It has no relationship with the people whose words it is reading, and it cannot carry responsibility for an interpretation. The question is which role the model should play, not whether it can produce themes.

By Heidi Andersén, MD, PhD. Updated 18 July 2026.6 minute read

Why a model cannot replace the researcher

A large language model generates text by learning statistical relationships between words and predicting the most probable next one. That gives you fluency, not understanding, and there is no built-in mechanism to check whether an answer is true. Because it returns the most likely patterns from its training data, its default output is the conventional reading, and even careful prompting produces outputs that vary between runs and can include fabricated content.

Qualitative analysis is interpretive and contextual. It rests on embodied judgment and reflexivity, on a researcher who was in the room. A tool built on pattern-matching cannot stand in for that. Practitioners who have tested these models say the same thing: they produce passable surface-level analyses quickly, but rarely at the level of an experienced researcher. At best they speed up routine tasks.

Where LLMs help

Used deliberately, a model supports the early, labour-intensive stages. In published comparisons, models like GPT-4 and Claude pull coherent themes from large sets of narratives and catch many of the same core ideas a human team finds — a 2026 scoping review counted 75 such studies, most using models for coding and theme identification. They organise complex text fast and draft preliminary themes, which saves time.

Different models have different habits: some condense themes into broad buckets, others over-segment. The approach that works treats the model as a junior analyst. You review its themes, merge the overlapping ones, and hold the line on conceptual clarity. Role-specific prompting and a clear analysis plan produce richer output, and sometimes the model surfaces a reading that challenges your assumptions and sends you back to the data.

The ethics, in one line

A model can produce plausible errors and give different answers to the same prompt. Its training data may be undisclosed, and a provider may retain or reuse what you submit. Qualitative data often contain sensitive narratives, so check consent, ethics approval, institutional agreements, and the provider's data policy before uploading them. A local model can reduce exposure, but it does not replace documentation or human judgment. The full ethics guide covers this in depth.

The five ways researchers over-trust AI

Five epistemic risks are worth knowing by name, because naming one is how you catch yourself doing it.

1. Category error. Mistaking the model's fluent language for genuine analytic insight.
2. Unreliable outputs. The same prompt producing inconsistent, irreproducible results.
3. Anthropomorphic fallacy. Attributing understanding or collaboration to a system that is repeating linguistic patterns.
4. Causal misattribution. Blaming your prompts for a failure that is really a structural limit of the model.
5. The oracle effect. Trusting an output as objective or neutral because it came from a machine.

The failure modes follow from the mechanism: a model has no reflexive capacity, over-weights frequent expressions, and can miss subtle or minority perspectives. Results shift with the prompt, the model, and the token limit. So check AI output against the raw data before it earns a place in the analysis.

A workable division of labour

The division below keeps two things straight: the role the model was given, and who stayed accountable.

Related guides

See also AI qualitative coding, human-reviewed, using LLMs on interview data ethically, and a short history of qualitative research. More guides are on the Learn page.