A research team interviews 30 early-years teachers about workload. One member pastes the transcripts into a generative AI assistant and asks it to “identify the main themes.” Five minutes later there are eight neatly labelled themes: “administrative overload”, “parental expectations”, “lack of management support”… They match everyone’s impressions. But when the team lead rereads the data, she notices that the most important thread in the interviews, a sense of having one’s professionalism doubted, which many teachers talked around rather than about, is missing. It was never stated plainly, so the model never “saw” it.
Qualitative analysis is not summarising. It is a researcher interpreting data from a theoretical position, with responsibility to participants. An AI assistant can help at some stages, but it cannot take on the role of interpreter.
What you can delegate and what you must keep
| Stage | Reasonable role for AI | What stays with you |
|---|---|---|
| Transcription | Speech-to-text first drafts | Listening back and correcting, especially names, dialect and meaningful hesitations |
| Familiarisation | Almost none | Reading everything, writing reflexive memos |
| Initial coding | Suggesting codes for a passage for you to compare | Coding it yourself and deciding what to keep |
| Organising codes | Grouping codes by keyword, spotting duplicates | Deciding how codes relate in meaning |
| Developing themes | Acting as a critic: “which passages contradict this theme?” | Interpretation, naming, linking to theory |
| Finding illustrative quotes | Locating passages that contain a specific idea | Rereading the full context before using them |
| Writing findings | Not advisable | All of it |
Why AI misses what matters
- It favours what is said often and plainly. Good qualitative analysis often surfaces what is said rarely, obliquely, or by a single participant in a way that illuminates the whole dataset.
- It lacks your theoretical lens. The same excerpt will be coded differently by someone using professional identity theory and someone using a power framework, and both are valid.
- It cannot hear. A bitter laugh, a long pause or a change of tone disappears once only text remains.
- It can invent quotes. Asked for “illustrative examples,” an assistant sometimes paraphrases or blends statements into words nobody said. Every quote must be checked against the transcript.
Ethics and confidentiality come first
Interview data often contains personal stories, names and workplaces. Before feeding it into any AI tool:
- Reread the consent form and ethics approval: do they allow processing by external services? If they are silent, ask your ethics committee.
- Anonymise transcripts first, replacing names, institutions and places with codes.
- Prefer institution-licensed tools that commit to not training on your data, or models running locally.
- For vulnerable groups or sensitive topics, consider not using AI on raw data at all.
Automated transcription: the biggest saving, if done properly
For many teams the largest benefit of AI is transcription: an hour of audio can take five or six hours to type by hand. Automated drafts still need a full listen-through, and that listening is itself valuable familiarisation. Watch for typical errors: misheard technical terms, regional expressions, overlapping speakers merged into one turn. Mark pauses, laughter and hesitation in the transcript with conventions your team agrees on.
A workflow with AI as the cross-checker
The safest and most useful approach is to treat AI as a second coder you do not trust by default:
- Code the first four or five transcripts yourself and build a codebook with a definition and example for each code.
- Give the assistant the codebook and an anonymised transcript you have already coded, asking it to apply the codes strictly by definition and quote the passage justifying each one.
- Compare the two versions. For each disagreement, ask: is my code definition vague? Did I miss something? Or did the AI misread the context?
- Revise the codebook where needed and record the decision in your analysis log.
- Repeat with a few more transcripts. Never let the AI code the rest and use its output directly.
Step 3 is where the real value lies: disagreement forces you to sharpen definitions, just as working with a human co-coder does.
Should you calculate human–AI agreement?
Some teams compute an agreement statistic such as Cohen’s kappa between human and AI coding. It can help spot poorly defined codes, but do not present it as evidence of analytical quality. In many qualitative traditions, especially reflexive thematic analysis, inter-coder agreement is not a primary quality criterion; consistency, transparency and interpretive depth matter more.
Reporting it in your methods
Qualitative reviewers increasingly ask about AI directly. A good methods paragraph states:
- The tool, version and when it was used.
- Which stages involved AI and for what (for example, “cross-checking initial coding on 5 of 30 transcripts”).
- How data were anonymised and the ethical basis for processing by an external service.
- That all final coding and interpretive decisions were made by the research team and every quote was verified against the original transcript.
If you are early in your analysis, try the cross-checking workflow on one transcript you have already coded. Thirty minutes of comparison will show you how clear your codebook really is, and how useful AI actually is with your data.
Câu hỏi thường gặp
Can I use AI to analyse interview data?
You can use it to support some stages, such as draft transcription, suggesting codes to compare or locating illustrative passages, if ethics approval allows and data are anonymised. Interpretation and final coding decisions must remain with the researcher.
Can AI replace manual coding in thematic analysis?
No. AI favours what is said often and explicitly, lacks your theoretical lens and can miss what participants say indirectly. It is most useful as a second coder for cross-checking.
Is it ethical to upload interview transcripts to an AI assistant?
Not if your consent forms and ethics approval do not permit processing by external services. Check first, anonymise the data and prefer institution-licensed tools.
Can AI make up quotes from qualitative data?
Yes. When asked for examples, assistants sometimes paraphrase or blend statements into words nobody said. Verify every quote against the original transcript.
How do I report AI use in qualitative research?
State the tool, version and stages used, how data were anonymised, the ethical basis, and that all interpretive decisions and quotes were checked by the research team.
