To prepare a research interview guide with AI, translate approved research objectives into a small set of topics, open starters, and neutral probes. Add the real introduction, consent, recording, withdrawal, timing, accessibility, and moderator instructions required by your research governance. Pilot the guide with a human moderator; AI is a drafting assistant, not an ethics reviewer or interviewer.

GOV.UK recommends planning each research round around clear questions, participants, methods, responsibilities, and analysis.［1］ Its in-depth interview guidance favors open questions and follow-up prompts that explore experience in the participant’s own words.［2］ Use the general guide to using AI for workflow basics and research-with-AI safeguards before processing study material.

Key Takeaways

Start from approved research objectives, not a list of desired answers.

Use topics, open starters, and neutral probes rather than a rigid script.

Keep one idea per question and remove leading assumptions.

Use the organization’s real consent, recording, and withdrawal language.

Do not give AI identifiable participant data or sensitive disclosures.

Pilot timing and comprehension, then obtain accountable approval.

Step 1: Define the research decision and objectives

Write what the team needs to learn and what decision the learning will inform. “Understand how new users recover from a failed identity check so the team can prioritize the next design investigation” is more useful than “find out whether users like the new process.” The first invites experience; the second invites approval-seeking.

For each objective, list what is already known, the evidence gap, the participants needed, the interview method, and what is out of scope. GOV.UK’s planning guidance recommends agreeing research questions and involving the team in planning and observation.［1］ Keep stakeholder assumptions separate from research questions so the guide does not smuggle a preferred conclusion into the interview.

Do not provide AI with participant names, contact information, health details, account records, or recruiting notes. Draft with generic participant characteristics that are necessary for method design. Apply the AI privacy-risk guidance before sharing any source material.

Step 2: Build a topic-to-objective matrix

Turn objectives into a manageable sequence of topics. Begin with context and recent experience, move into the focal task, explore variation and recovery, then close with anything missed. A 45-minute interview usually cannot support twenty equally important topics.

Topic ID	Research objective	What to explore	Evidence needed	Sensitivity	Minutes	Analysis tag
T-01	Understand recovery	Most recent failed attempt	Sequence and workaround	Medium	8	recovery-path

Ask AI to identify duplicate topics, objectives with no question, and questions with no objective. Require it to preserve IDs. The matrix prevents an interesting tangent from consuming the session while a critical research question goes untouched.

Step 3: prepare a research interview guide with AI using open starters and neutral probes

For each topic, draft one open starter and several optional probes. GOV.UK advises asking open questions, following what participants say, and avoiding leading them toward your assumptions.［2］ A starter such as “Tell me about the last time you tried to complete this task” is stronger than “Was the new flow easier?”

Neutral probes include:

“What happened next?”
“How did you decide what to do?”
“Can you tell me more about that?”
“What, if anything, did you expect at that point?”
“Could you show me how you did it, if you are comfortable?”

Instruct AI to flag leading, loaded, hypothetical, double-barreled, binary, and jargon-heavy questions. Ask for an explanation rather than an automatic replacement. Use better prompt-writing practices to constrain the model to the approved objectives and vocabulary.

Run a question-level traceability check. Every essential starter should point to one research objective, and every objective should be covered by at least one starter. A probe may support more than one objective, but do not overload it with several ideas. Mark whether a question asks about a recent event, an observed task, a general belief, or a hypothetical scenario; these produce different evidence and should not be analyzed as interchangeable.

Prefer recall of a specific recent experience before abstract opinion. “Tell me about the last time” gives the participant a concrete anchor, while “What do you usually do?” often produces an idealized account. Hypothetical questions can be useful for concept exploration, but label them and avoid reporting the answers as observed behavior.

Step 4: Add introduction, consent, recording, and withdrawal scripts

Insert the organization’s approved language, not wording invented by the model. The introduction should identify the moderator and study purpose, explain the session structure, clarify that the product—not the participant—is being evaluated, and invite questions.

GOV.UK says informed consent should explain why research is being done, how data will be used and stored, who will see it, whether sessions are recorded, and how participants can withdraw.［3］ Consent is an ongoing process, not merely a checkbox at the beginning. Separate agreement to participate from permission to record, and record the actual withdrawal route and any practical limits.

Do not ask AI to decide whether consent is valid, whether a disclosure must be escalated, or whether your study requires ethics, legal, safeguarding, or data-protection review. Route those questions to the real accountable owner.

Step 5: Review sensitivity, safeguarding, and accessibility

Mark potentially sensitive topics and explain why each is necessary. Put less sensitive context before more sensitive questions, give participants a clear option to pause or skip, and prepare a stop or escalation procedure. Do not ask for sensitive detail merely because it might be interesting.

Check language complexity, interpreting needs, captioning, screen-reader compatibility, communication preferences, breaks, session length, device or venue access, and whether showing a prototype could exclude someone. Ask participants about accommodations through the approved recruiting process rather than inferring disability or access needs.

Review cultural assumptions and power dynamics. A manager observing an employee, a caregiver speaking for someone, or an incentive controlled by a service provider may change what a participant feels able to say. AI cannot resolve these ethics questions; a qualified research owner must decide the protocol.

Step 6: Add timing and moderator notes

Allocate time to welcome and consent, warm-up, core topics, participant questions, and close. Identify which questions are essential and which probes can be skipped. Include short moderator notes for transitions, silence, off-topic answers, distress, technical failure, and time pressure.

Avoid writing a word-for-word interrogation. A guide supports listening; it should not force the moderator to ask every probe in order. Notes can say “follow the participant’s language,” “do not correct their mental model,” and “ask for a recent example before a general opinion.”

Create a session log template with guide version, moderator, observer, consent and recording status, start and end time, topics covered, skipped questions, incidents, and debrief owner. Never prefill it with expected findings. The AI fact-checking workflow is relevant when the guide includes factual context, but participant accounts remain research data to interpret, not claims to correct mid-interview.

Step 7: Pilot the guide and revise it

Run at least one pilot with someone who is not part of the target sample, or with a suitable internal proxy when appropriate. Test whether the introduction is understandable, consent is complete, questions sound neutral when spoken, transitions are natural, sensitive topics are handled safely, and the session fits the allotted time.

After the pilot, debrief the moderator and observer. Record question ID, issue, evidence from the pilot, proposed change, owner, and approval. If participants answer the same interpretation of a supposedly open question, check whether wording or context is leading them. If a probe repeatedly produces confusion, simplify or remove it.

Pilot the operational parts as well as the questions. Confirm that the meeting invitation matches the consent explanation, recording indicators are visible, observers know when they may speak, note-taking tools work with the retention policy, and the moderator can stop recording without disrupting the session. Rehearse a participant declining recording, skipping a sensitive topic, asking to withdraw, or becoming distressed.

Time the guide from the participant's perspective. Moderator setup and observer debrief do not belong inside promised interview time. Leave space for pauses and clarification rather than treating silence as a delay to fill. If the pilot only fits when every answer is short, reduce the topic list.

Do not use synthetic AI personas as proof that a guide works. They may help the writer spot awkward wording, but they cannot reproduce lived experience, consent, fatigue, trust, cultural context, or unexpected participant behavior.

Step 8: Prepare analysis and obtain final approval

Before fieldwork, map each topic ID to a preliminary analysis tag and research objective. This is a filing structure, not a list of expected findings. Allow new codes to emerge and preserve disconfirming cases. Define where notes and recordings will be stored, who may access them, retention rules, and who leads the post-session debrief.

NIST’s Generative AI Profile highlights confabulation, privacy, information-integrity, and human-oversight risks.［4］ Preserve the AI input boundaries, draft output, human edits, pilot evidence, and final approver. Use a human-approval workflow so only the reviewed version reaches moderators.

Freeze the final guide with version, date, study ID, owner, approved participant group, required scripts, and change history. If a material question or consent term changes during fieldwork, document the reason and decide whether earlier sessions remain comparable.

FAQ

Can AI generate interview questions from research goals?

It can draft topics, open starters, and probes from approved objectives. A researcher must remove assumptions, confirm coverage, and approve the final wording.

What makes an interview question leading?

It suggests a preferred answer or embeds an unverified premise, such as “Why was the faster process better?” Ask about the participant’s experience without prescribing it.

How many questions fit in a 45-minute interview?

There is no fixed number. Budget time by topic, identify essential starters, and treat probes as optional so the moderator can listen and follow relevant detail.

Can AI write the consent script?

It may format already approved requirements, but it should not invent legal or ethical language. Use the organization’s authorized script and accountable review.

Should the guide include sensitive questions?

Only when necessary for the objectives and covered by the research protocol. Explain the reason, offer skip and stop options, and prepare safeguarding steps.

Can synthetic personas replace a pilot interview?

No. They can expose awkward wording but cannot validate comprehension, timing, consent, accessibility, trust, or real participant responses.

Should the moderator follow every question in order?

No. Keep topic coverage consistent, but allow neutral follow-up based on the participant’s account and skip optional probes when time is limited.

What should be frozen before fieldwork?

The approved objectives, guide version, participant scope, scripts, timing, moderator notes, storage rules, analysis preparation, and named approver.

Sources

GOV.UK Service Manual, Plan a round of user research — https://www.gov.uk/service-manual/user-research/plan-round-of-user-research
GOV.UK Service Manual, Using in-depth interviews — https://www.gov.uk/service-manual/user-research/using-in-depth-interviews
GOV.UK Service Manual, Getting users’ consent for research — https://www.gov.uk/service-manual/user-research/getting-users-consent-for-research
NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf

Sources checked 30 August 2026.

How to Prepare a Research Interview Guide with AI