How to check whether AI interview findings have evidence

A citation gives you somewhere to start checking

A fluent AI analysis with participant quotations underneath can look reassuring. There are still two separate questions: do those words occur in the source, and do they support the claim above them?

Text matching can help answer the first. The second requires attention to the question asked, the participant's circumstances, and conflicting material. A real quotation can sit underneath an overreaching finding.

All people, numbers, and conversations below are invented teaching examples. They are not ScribeFlash customer data or evidence of a tool's accuracy.

Open the source and read the actual words

Suppose an AI finding says “New users abandon colleague invitations because the price is too high,” citing: “Adding people increases the cost, so I haven't invited them yet.”

Open the linked interview and locate the complete statement. Check for missing negation, answers stitched together from distant passages, or a researcher's summary presented as a quote. If the source actually says “I don't think it's expensive; we haven't finalized the list yet,” the explanation changes entirely.

Set aside a quotation you cannot locate. Asking AI to produce a more natural-sounding replacement cannot repair missing evidence. If you paraphrase, label it as a paraphrase and retain the source for checking.

For a transcript derived from a recording, replay important amounts, names, and negative statements where possible. With text-only material and no recording, state that you checked the available text, not the audio.

Read the question before it and the response after it

Consider this invented exchange:

Interviewer: If adding each person doubled the bill, would you hesitate?

P02: Of course. But I'm happy with the current price.

Extracting “Of course” could turn a response to a leading hypothetical into “The participant rejects the current price.” The surrounding exchange tells a different story.

Check who was speaking, too. A facilitator's prompt, a colleague's interruption, and a participant quoting their manager are not interchangeable statements of the participant's own needs. If speaker labels are unreliable, record the uncertainty instead of guessing from tone.

Short answers such as “fine,” “okay,” and “no problem” deserve a few extra lines of context. Was the person judging a design, agreeing to continue, or simply being polite?

Ask which part of the finding the words establish

The first invented quotation connects seat count with cost and describes a delayed invitation. It does not establish whether the price exceeds a budget, whether the participant abandoned the product, or whether approval is still pending.

A narrower working finding is “P01 considers added-seat costs before inviting colleagues.” Follow up on budget and what happened next. “High pricing causes new-user churn” needs further evidence.

Recommendations require the same care. Confusion about billing does not make a price reduction the only response. Pricing explanations, approval processes, and actual prices are different possible issues.

NN/g's distinction between data, findings, and insights is useful here: a quotation supplies material, while interpretation requires context. Do not let a tidy sentence conceal that interpretive step.

Five quotations may come from one person

If P01 mentions cost five times, cost may matter a great deal to P01. It does not mean five users consider the product expensive. Count distinct participants, then check that their statements concern the same question.

Silence is not indifference. P03 may never have been asked about price, while P04 used a prototype without seeing a billing screen. Including both in a denominator can make an unclear result look precise.

Actively look for another explanation. Suppose P05 says “The price is fine; procurement approval is what holds us up.” Keep that account beside the current finding. It may suggest separating participants by purchasing process rather than treating every delay as a price objection.

If you reviewed the material and found no counterexample, say that none was found in this study. That statement has a much smaller scope than “all users behave this way.”

A practical review sequence in ScribeFlash

1. Open the project's findings and select a claim that could affect a decision. Inspect its supporting evidence.
2. Follow the source to the interview and check the words, attribution, and surrounding exchange. Use source lines for imported text and recording positions for audio when available.
3. Review counter-evidence and gaps. Narrow an overbroad claim; correct or remove an unreliable reference.
4. Confirm the finding after reviewing its meaning. Check references again when generating the report. Later source edits or deletions may require another review.

For text sources, ScribeFlash checks quotations against the linked current source ranges and interview attribution. This can identify invalid citations. It does not establish that a quotation logically proves a finding. Passing a report citation check cannot replace reading the material.

If you have not imported your data, follow the existing-transcript guide. An AI system checking its own answer again is not independent verification.

Leave a short record of what changed and why

Using the same invented case, a sufficient note might read:

Original claim: High prices cause new-user churn.

Review: P01's words exist, but describe a delayed invitation, not abandonment. P05 reports an approval delay.

Decision: Narrow to “Some interviewees need to confirm cost or approval conditions before invitations.” Keep the two explanations separate. Willingness to pay needs follow-up.

Before delivery, check the important quotations you include in the report. Sampling can uncover problems in a large dataset, but record what you sampled. Checking ten quotations does not verify the entire report.

If only two findings survive review, deliver two. A specific question for the next study can also be a useful research outcome.

References

• NN/g: Accelerating Research with AI, reviewed January 2026, discusses human review of AI-assisted research.
• NN/g: Data vs. Findings vs. Insights distinguishes source material from interpreted findings.
• GOV.UK: Analyse a research session covers working with observations and research judgments.

The cover is an illustrative interview scene. For the handoff document, continue with writing a user research report.