To summarize user interviews with AI without flattening what participants meant, work in two passes. First create a traceable summary for each session. Only then compare sessions in a theme matrix. Every finding should retain the research question, speaker label, timestamp, nearby context, evidence status, counterexample, and a researcher’s interpretation. The model can organize supplied material; it cannot decide what a participant intended or whether a pattern represents the wider population.

Start with the general guide to using AI safely, and use the separate research interview guide workflow before fieldwork. This article begins after interviews have been completed and a usable record exists.

Key Takeaways

Confirm consent and allowed uses before uploading any recording or transcript.

De-identify the smallest useful transcript excerpt and freeze its version.

Summarize one session at a time before looking for cross-session patterns.

Keep questions, speakers, timestamps, context windows, quotes, and omissions attached.

Separate what was observed from the researcher’s interpretation.

Search deliberately for contradictory evidence and negative cases.

Let a researcher approve findings, quotes, and publication language.

Why does interview context disappear?

Interview summaries often fail because they compress too early. A sentence such as “users dislike setup” may combine an interviewer’s leading question, one participant’s conditional complaint, and another person’s unrelated example. Once those distinctions disappear, a polished synthesis can no longer be checked against the session.

Context includes more than the sentence around a quote. It includes the question being answered, the participant’s role in the scenario, whether they described direct experience or a hypothetical, the sequence of events, and any correction later in the conversation. It also includes what the transcript could not capture, such as inaudible audio or a missing segment.

The answer is not to paste every transcript into one enormous prompt. Use a controlled evidence structure, preserve links back to the source, and delay theme naming until individual sessions have been reviewed.

Step 1: Freeze the research question and consent boundary

Write a short analysis brief before using a model. It should identify the research question, included sessions, excluded material, intended audience, decisions the work may inform, and the person authorized to approve the synthesis. If the question changes, create a new brief rather than silently reusing the old analysis.

GOV.UK guidance says informed research consent should explain the purpose, what data will be collected, how it will be used, who may process it, and how participants can withdraw.［1］ Check the signed or recorded agreement, the privacy notice, and organizational policy. Consent to join an interview is not automatically consent to send the recording to any AI provider.

Record the permission decision for each input:

Field	Example using synthetic labels
Study ID	`STUDY-ALPHA`
Research question	How do first-time users recover from a failed setup?
Session ID	`SESSION-04`
Allowed material	De-identified transcript excerpts
Prohibited material	Raw audio, names, account details
Approved processor	Organization-approved workspace only
Retention rule	Delete working copy after review window
Analysis owner	Named researcher

If any permission is uncertain, stop that session’s processing. Do not treat a model’s privacy settings, paid tier, or general assurance as evidence that your specific use is permitted.

Step 2: Minimize and de-identify the input

Make a working copy, preserve the original separately, and remove information that is not needed to answer the research question. Replace names with stable labels. Remove email addresses, account identifiers, precise locations, secret URLs, and unrelated personal stories. Generalize details only when doing so does not change the meaning being studied.

GOV.UK’s note-taking guidance recommends protecting participant privacy, anonymizing notes, and keeping detailed observations that can be used in analysis.［2］ De-identification is a review task, not a search-and-replace trick. A rare job title plus a city and an incident date may identify someone even after their name is removed.

Use the AI privacy risk checklist when selecting a workspace and deciding what to omit. Keep the re-identification key outside the model. If the research question can be answered from excerpts, do not upload the complete transcript.

Step 3: Create an interview context packet

Build one packet for each interview. Do not mix participants at this stage. A strong packet has enough structure to reconstruct why an observation was made.

Field	Required content
Session ID and transcript version	Stable identifiers and capture date
Research question	The approved question this analysis answers
Participant label	Non-identifying label and only relevant segment
Interviewer question	Full question, including material framing
Speaker and timestamp	Who spoke and where to find the passage
Context window	Nearby turns before and after the focal statement
Transcript quality	Missing, overlapping, translated, or uncertain text
Direct evidence	Exact short quote or faithful paraphrase
Observation	What the passage explicitly shows
Researcher interpretation	A separate, reviewable inference
Counterexample	Evidence that limits or contradicts the interpretation
Open question	What cannot be concluded from the session

The context window should be large enough to preserve the condition around a statement, but no larger than necessary. When the participant revises an earlier answer, include both passages. When a timestamp is missing, label it missing; never ask the model to invent one.

Step 4: summarize user interviews with AI one session at a time

Ask the model to use only the packet, keep the field names, and return not evidenced when the transcript is silent. Request a summary of goals, described behavior, obstacles, workarounds, consequences, and unanswered questions. Require every statement to cite a session ID and timestamp.

A useful per-session output has three layers:

Observed evidence: what the participant explicitly said or described.
Interpretation: what the researcher thinks that evidence may mean.
Limit: ambiguity, contradiction, missing context, or a competing explanation.

Do not let the model replace ordinary language with diagnostic or psychological labels. “The participant paused and asked for help” is an observation. “The participant was anxious and lacks confidence” is an interpretation that the transcript may not support. The same rule applies to inferred intent, identity, expertise, and purchasing power.

Review the output line by line against the transcript. Check speaker attribution, timestamps, negation, quantities, conditional language, and whether a hypothetical was rewritten as an actual event. If many corrections are needed, repair the packet or transcript before continuing instead of polishing an unreliable summary.

Step 5: Audit omissions before synthesis

An accurate sentence can still produce a misleading summary if important material was omitted. Compare the summary against the interview guide and mark each relevant question as covered, not asked, unanswered, or excluded. List significant events and quotes that did not fit the draft.

GOV.UK’s session-analysis guidance emphasizes reviewing observations soon after research, grouping evidence, and involving the team so findings are not based on one person’s memory.［3］ A short debrief can reveal misunderstood terminology, missing situational context, and observations that were not captured in the transcript.

Run an explicit negative-case review:

Which evidence contradicts the apparent story?
Did the participant describe a workaround that succeeded?
Was the problem conditional on a device, plan, role, or earlier step?
Did the interviewer suggest the answer?
Does a later statement correct the quoted passage?
What important question remains unanswered?

Keep these answers inside the session record. Do not hide them in a private note that will be lost when the synthesis is shared.

Step 6: Build a cross-interview evidence matrix

Only reviewed session summaries should enter cross-session analysis. Give every candidate theme a neutral identifier before naming it. Use a matrix rather than a paragraph so support and limits remain visible.

Theme field	What to record
Theme ID	Stable identifier such as `THEME-03`
Candidate statement	Narrow claim, not a slogan
Supporting sessions	Session IDs and exact evidence references
Contradicting sessions	Negative cases and alternative paths
Conditions	Segment, task, environment, or sequence
Coverage	Included sessions, not a population estimate
Confidence note	Evidence quality and remaining uncertainty
Researcher decision	Accept, revise, merge, split, or reject

Ask AI to cluster semantically similar observations, but require it to show every proposed membership. Researchers should be able to move an observation, split a broad cluster, or reject a theme without rewriting the evidence. Do not use the model’s confidence number as a substitute for sample design or research judgment.

For open-ended survey data, use the dedicated survey theme analysis workflow. Survey responses and moderated interviews have different context, sampling, and evidence structures.

Step 7: Write findings with claims and limits

Turn each approved theme into a finding only after the matrix is stable. A reviewable finding contains the claim, the situation in which it occurred, supporting evidence, contradictory evidence, affected sessions, unknowns, and an implication framed as a question or hypothesis.

Avoid counts that imply statistical prevalence when the study was not designed for it. “Four of six included sessions mentioned recovery email delays” describes the analyzed material. “Most users have this problem” makes a population claim that the study may not support.

If the next deliverable is a customer journey map, pass along evidence IDs and uncertainty rather than copying theme labels as established journey stages. If the task is only to compress a policy or report, use the document summarization workflow instead.

Step 8: Approve, version, and retain the audit trail

Have the research owner review every finding and every publishable quote. A second reviewer should spot-check high-impact conclusions, unexpected themes, and translated passages. Record decisions to merge, split, reject, or soften findings.

Version the analysis brief, transcript working copy, session summaries, theme matrix, and final report together. Preserve a change note explaining what changed and why. If a transcript correction arrives, identify the affected evidence IDs and re-run only the dependent summaries and themes.

NIST’s Generative AI Profile describes risks such as confabulation and data privacy problems that require documented risk management rather than trust in fluent output.［4］ Treat the model response as a transformation to inspect, not as proof that research has been analyzed correctly.

A compact prompt contract

Give the model a narrow contract: use only the supplied session packet; preserve session IDs, speakers, timestamps, negation, and conditions; separate observations from interpretations; quote only provided text; show counterexamples; return not evidenced for missing facts; and do not infer identity, emotion, intent, representativeness, or population prevalence. Ask for a table first. Narrative findings come after human review.

Final review checklist

Consent and processor permissions match the actual input.
The working copy contains the minimum necessary, de-identified material.
Every summary points to a session ID, speaker, timestamp, and context window.
Observations, interpretations, and unknowns are visibly separate.
Quotes match the transcript and preserve conditions and corrections.
Omissions and counterexamples were reviewed before theme creation.
Cross-session themes list both supporting and contradicting sessions.
Findings do not claim population prevalence without an appropriate design.
A researcher approved the report and the change history is retained.

FAQ

Can I upload raw interview recordings to an AI tool?

Only when the participant agreement, organizational policy, processor terms, security controls, and research owner all permit that exact use. A de-identified transcript excerpt is often safer and sufficient.

Should I summarize all interviews in one prompt?

No. Review one session at a time first. Combining transcripts too early makes it easier to mix speakers, lose conditions, and hide contradictory cases.

How much surrounding context should I keep?

Keep the question and enough adjacent turns to preserve conditions, corrections, and references. Record why the chosen context window is sufficient.

Can AI decide the themes automatically?

It can propose clusters from supplied evidence. Researchers must inspect membership, contradictions, scope, and naming before accepting a theme.

Are AI-generated quotes acceptable?

No invented or reconstructed quote is acceptable. Use exact transcript text, verify it against the source, and follow the participant’s quotation permissions.

How do I handle an inaccurate transcript?

Mark uncertain passages, consult the recording only if permitted, correct the transcript version, and re-run dependent analysis. Never silently smooth over missing words.

Can I report how many users had a problem?

You can state how many included sessions contained specified evidence. Do not generalize that count to all users unless the research design supports that inference.

What should happen to counterexamples?

Keep them beside the theme they limit. They may reveal conditions, alternative paths, or a theme that needs to be split.

Who approves the final interview summary?

The named research owner, with appropriate privacy or subject-matter review, approves the findings and any external quotations. The model has no approval authority.

Sources

GOV.UK Service Manual — Getting users’ consent for research — https://www.gov.uk/service-manual/user-research/getting-users-consent-for-research
GOV.UK Service Manual — Taking notes and recording user research sessions — https://www.gov.uk/service-manual/user-research/taking-notes-and-recording-user-research-sessions
GOV.UK Service Manual — Analyse a research session — https://www.gov.uk/service-manual/user-research/analyse-a-research-session
NIST — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Sources checked 4 September 2026.

How to Summarize User Interviews with AI and Keep Context