Thematic Analysis

How do I use AI for thematic analysis?

Use it for the parts that reward consistency and scale: initial coding across a whole corpus, applying an approved code frame to every transcript identically, tracking prevalence by participant, and finding passages that contradict a theme. Do not use it to decide what a theme means, to resolve ambiguity, or to interpret cultural context. The critical control is sequencing: a human reviews the code frame before it is applied at scale, because a definitional error applied to thirty transcripts is thirty errors.

Can AI code qualitative data reliably?

It can code consistently, which is not the same thing. Two AI passes over the same material measure stability, not validity: they show the frame was applied the same way twice, not that it was applied correctly. Report it as stability, adjudicate the disagreements with a human, and fix the code definitions rather than the individual instances.

What is the difference between reflexive thematic analysis, framework analysis and grounded theory?

Reflexive thematic analysis treats themes as constructed by an analyst making visible decisions, and is the best default for most applied research. Framework analysis suits fixed objectives, cross-case comparison and clients who need a matrix. Grounded theory is for building theory and requires theoretical sampling and constant comparison, which most commercial timelines do not permit. Say which you used and why.

Should I code inductively or deductively?

Hybrid, usually. A deductive spine drawn from the research objectives ensures the analysis answers the questions the study was commissioned to answer. Inductive freedom below it lets the material say things nobody anticipated. Mark which codes came from where, so a reader can see how much of the frame was imposed.

How many participants make a theme?

There is no threshold, and any number offered as one is arbitrary. Report the count and the denominator and let the reader judge. Prevalence tells you how widely a view is held. It does not tell you how much it matters, and those two judgements belong in separate columns.

How do I know if a theme is important rather than just frequent?

Assess it against the decision the research informs, the consequence if it is true, whether it explains other themes, whether the people who hold it occupy a structurally significant position, whether anyone could act on it, and whether it is already known. A theme raised by five recent defectors often outranks one raised by everyone, because everyone is describing the category and the five are describing the failure.

What is saturation, and can I claim it?

Code saturation means no new codes are appearing. Meaning saturation, which matters more, means no new nuance is appearing within existing codes. You can only assess either after the fact, by tracking when new codes appeared. If new codes were still arriving in your final transcripts, you did not reach saturation, and a sample size on its own never demonstrates one.

How do I stop AI inventing or smoothing quotes?

Keep the participant ID attached to every extract from transcript to report, verify every client-facing quote word for word against source, and allow only the standard edits: filler removal, marked elision, bracketed clarification, flagged transcription corrections. Never merge two speakers, never present a paraphrase as verbatim, and check that a quote attributed to a segment came from someone in it. A quote that arrives without an ID cannot be checked, and an uncheckable quote is functionally a fabricated one.

Why does AI keep flattening contradictions in my qualitative data?

Because language models optimise for coherence, and ambivalence reads as noise to be cleaned up. This is the single most damaging failure in AI-assisted qualitative work, because the ambivalence is frequently the finding. Run contradiction as a separate deliberate pass, and write the tension into the theme definition rather than choosing a side.

When should I not use thematic analysis?

When the material is too thin to carry meaning (use open-ended response coding), when there are more than roughly 5,000 items and genuine familiarisation is impossible (use large-scale text analytics), when the unit of interest is one person (use transcript analysis), or when the real question is how many, which a qualitative sample cannot answer.

Do I still need a researcher if AI does the analysis?

Yes, at three specific points: reviewing the code frame before it is applied at scale, adjudicating ambiguous passages where two readings are both defensible, and interpreting cultural or linguistic context. These are positional limits rather than capability ones, and the output marks them explicitly rather than leaving them for the reader to find.