Blog / AI & search /

Using AI to organise interview notes without inventing quotations

You ask an AI tool to tidy up your interview notes. It hands back a lovely summary, complete with a perfect quote. One small problem: nobody actually said it.

Organising interview notes with AI can blur the line between a summary and a quotation. A tool might merge two comments, polish someone’s grammar, or produce a sentence no participant ever spoke. If the notes will inform an article, a product decision or a public claim, keep the source intact and be clear about what the summary can and can’t support.

AI is genuinely useful for grouping repeated themes or turning a long set of notes into a working index. It shouldn’t create evidence. A good workflow keeps the original notes to hand, anonymises sensitive details, and has a person check every quotation and conclusion before anything is shared.

Prepare notes before using a tool

Review consent and confidentiality first. Did participants agree to have their responses processed by an AI service? Does the organisation permit this tool for interview data? Remove names, contact details and identifying context when they are not required. Some details can identify a person even after their name is removed, so use judgement rather than assuming an anonymised file is automatically safe.

Keep the original recording or transcript in the approved location, with access limited to the people who need it. Make a separate working copy for analysis. Don’t paste private notes into an unapproved tool, or upload material when you don’t know how it will be handled.

Explain the purpose and limits in the prompt. Ask for themes based only on the supplied notes, ask it to mark uncertainty, and tell it not to generate quotations. A better prompt reduces the risk. It doesn’t guarantee the model will follow every instruction.

Keep a trace from summary to source

Ask the tool to reference a participant ID, paragraph or note location for each theme. Then open the original text and confirm that the cited material supports the summary. If the system cannot provide a trace, treat the output as a rough index and verify the underlying notes manually before using the claim.

Keep paraphrases and direct quotes apart. A paraphrase can condense someone’s point, but it mustn’t change the meaning or suggest the person used the generated wording. Direct quotations should match the source exactly, apart from clearly marked cuts or edits your interview policy allows. This is a line I hold firmly: if I can’t find the words in the transcript, they don’t go in quotation marks.

Don’t turn a recurring theme into a number unless you counted responses with a clear method. “Several participants mentioned…” may still need a qualifier if the sample is small or unrepresentative. The notes show what those people said, not a universal truth about every customer.

Review the interpretation, not just the text

Ask whether the summary preserves disagreement and exceptions. If one participant described a different experience, the model may omit it while producing a smoother theme. Keep minority observations when they change the design or content decision. A rare issue can still be important if its impact is high.

Check the question that produced each answer. Leading questions shape responses, and AI summarising can’t undo that. Read the original questions and think about how they colour the themes. If notes are incomplete, record the gap. Don’t ask the model to fill it.

Have the researcher or interviewer review the output. They can identify context, tone or follow-up details that a transcript summary misses. For sensitive research, involve the person responsible for ethics, privacy or customer data before using the findings externally.

Use findings responsibly in content

Before publishing an interview insight, confirm you have permission to use it, and make sure a combination of details doesn’t reveal who said it. If the person didn’t agree to be quoted, describe the lesson in general terms and don’t present it as a testimonial. Never invent a quote because it sums up the theme nicely. I’d rather have a slightly flatter paragraph than a fabricated voice.

Keep the analysis record separate from the public article. The article should present only verified information and explain the evidence at a level readers can evaluate. AI can help organise the material, but the researcher remains responsible for the interpretation.

An interview participant may have agreed to a particular use of their information, not to every possible analysis. Review the consent language and the research owner’s guidance before processing notes. If the notes include sensitive experiences, names or details that could identify someone, anonymise the working copy and remove context that is not needed for the analysis.

Do not upload a recording to a transcription or AI service without confirming that the service and its data practices are approved. A transcript may contain more identifying information than the notes. If the team has not agreed on a safe tool, use a manual process or seek approval rather than improvising.

Keep a separation between internal research synthesis and public marketing. A participant’s comment is not automatically a testimonial. Ask permission before quoting them, preserve the exact wording and context, and do not edit a quote until it becomes an endorsement the person did not make.

Represent disagreement and minority views

AI summaries often prefer a coherent theme. Check the notes for dissenting views, uncertainty and differences between participants. A small number of comments can still reveal a usability or safety issue. Do not erase them merely because they do not fit the dominant pattern.

Separate what participants said from the researcher’s interpretation. “Two interviewees described difficulty finding the setting” is different from “users cannot find the setting”. The first reports the sample; the second implies a broader conclusion. Include the scope and method when you present findings to colleagues.

If a summary says “most people” or gives a percentage, verify that the sample and counting method support it. Do not turn qualitative notes into pseudo-quantitative evidence. Use a count only when the data was coded consistently and the audience understands what it represents.

Keep the audit trail light

Store the prompt, output and reviewer notes only as needed under the research data policy. Record which themes were verified and which remain tentative. Restrict access to the original notes and avoid copying sensitive excerpts into a public content draft.

Before using a finding in an article, check the original context and the participant’s permission. If no quote is approved, describe the general lesson without suggesting a specific customer said it.

Here’s a habit worth building: next time AI groups your notes, pick three themes at random and trace each one back to the exact words in the transcript. Can you trace every important theme back to the source without exposing the participant?