AI-moderated interviews collect new answers from real participants, with software asking the questions; Human Digital Twins generate modelled answers from existing information about people. Choose between them by deciding whether you need a fresh account of someone's experience or an early reaction to something you're making. Disclosure: I built AYA, so weigh this accordingly.
I'd start with the evidence the decision needs, because the same chat window can hide two very different research methods. If you need to know what happened when a customer tried to cancel yesterday, speak to that customer. If you're checking whether a draft cancellation message explains the next step, a twin can help you find wording to examine before fieldwork.
The most important person in an AI interview is whoever supplies the answer.
AI-moderated interviews vs Human Digital Twins: who answers?
In an AI-moderated interview with a real participant, the person supplies the answers and software steers the conversation. A researcher prepares the discussion guide, sets the audience criteria and reviews the output. Depending on the system, the moderator can ask follow-up questions when an answer needs more detail.
Research published at LaTeCH-CLfL in 2025 tested this distinction directly: university students answered questions from either human or AI interviewers using the same questionnaire. The researchers assessed the interviewers' adherence to guidance, alongside response quality and engagement. The respondent remained a person in both conditions.
A twin supplies a generated answer. For interview-based Human Digital Twins, the foundation is information collected in an earlier interview. You show the twin your new material and ask how it reads it. The output is a modelled reaction grounded in that source material.
Read both sides of a supplier's description: who moderates, and who responds. An AI moderator can also question a modelled respondent. A transcript alone leaves that distinction unresolved unless the report labels it.
| Question | AI moderator with real participants | Interview-based Human Digital Twins |
|---|---|---|
| Where does the new answer come from? | A participant taking part in this study | A model drawing on existing interview evidence |
| What must the team arrange? | Participant recruitment, a guide and fieldwork | An appropriate twin panel, stimulus and question |
| What can a follow-up explore? | A person's account of their experience | The twin's interpretation of the material |
| What belongs beside a quote? | Participant reference, study date and question | Twin label, source run and question |
| What should the team inspect? | Recruitment, moderator prompts and original responses | Source evidence, panel fit and generated responses |
Which method fits an early-stage concept test?
For an early-stage concept, match the method to the uncertainty you need to resolve. A draft full of vague promises gives you a wording task. An assumption about how customers currently solve a problem gives you a discovery task.
For wording, show the concept and ask: “What would you expect to receive?” Then examine where interpretations differ. The message-testing examples show how a small twin panel can expose details that need explaining.
For discovery, recruit people who recently faced the problem. Ask: “Tell me about the last time you tried to do this. What happened next?” Check whether the moderator follows the participant's account closely enough to uncover the steps, workarounds and consequences.
A traditional focus group adds something else: people responding to one another. Choose that format when the discussion between participants matters to the question. Our AI focus-group explainer explains why the label needs unpacking before you book a study.
A gym email shows where the methods part company
A twin reaction can identify a useful question for subsequent customer interviews. On 2 November 2026, six UK Human Digital Twins saw this hypothetical gym email, written by AYA, in a Concept Test:
From 1 January you can cancel your membership any time in the app, no notice period. We're also adding £2 a month to cover longer opening hours.
The question was: “Your gym sends you this email. What is your first reaction, and what would you do next?” Jordan, a 31-year-old twin from Redlands, Reading, answered:
I'll check whether the new hours mean earlier morning access, but otherwise I'll keep the membership as it is.
That gives the writer a concrete task: put the actual new hours beside the fee. It also suggests a separate fieldwork question for real members: “Tell me about the last time the opening hours affected a visit.” Their accounts could establish which hours caused difficulty and what they did instead.
Those customer interviews are a proposed next step. The completed run contains modelled reactions to the original email only. Its small panel is directional, built from recorded interviews, and provides no forecast of membership retention.
A brief you can copy before choosing a tool
A useful brief names the decision, the evidence required and what the team will do with it. Here is a hypothetical brief for the gym message:
Decision: What must our opening-hours announcement explain before we send it?
Material: The complete draft email, current timetable and confirmed new timetable.
Audience: For a twin check, select UK twins with relevant fitness interests where those details are available. For customer interviews, recruit current members who have recently tried to visit outside our opening hours.
Twin question: What changes for you after reading this email, and which detail would you look for next?
Customer interview opener: Tell me about the most recent time you wanted to visit while the gym was closed.
Follow-up rule: Ask for the sequence of events or clarify the participant's own words. Leave the expected benefit out of the question.
Deliverable: Original answers beside each proposed change to the email, with twin reactions and customer accounts labelled separately.
Before commissioning fieldwork, try the guide with a colleague giving a vague answer, an unexpected answer and a request to stop. Check the moderator's response in each case. Keep the resulting transcript so the researcher can inspect what was actually asked.
For a twin check, read the panel information first. If the relevant experience is missing, narrow the task to comprehension of the words or recruit real members for the experience question.
Limits
AI moderation still needs research judgement. Recruitment affects whose experiences you hear; question wording and follow-ups affect what they tell you. The cited study is a small, controlled comparison with students, not proof that any commercial moderator suits every audience. Source checked 11 November 2026.
A twin's new answer is not a new statement from the person behind its source interview. Twin reactions cannot establish current customer behaviour, willingness to pay or future sales. The gym example used six modelled responses to hypothetical copy; no live-member interviews or retention outcomes were collected.
This is a choice between research tasks, not a ranking of suppliers. Costs, accessibility and the amount of human oversight depend on the service and study design. Check those against your own brief.
How AYA fits this choice
AYA's Human Digital Twins start from recorded interviews with consented participants, with 120+ data points per twin. We use them for directional reactions to ads, concepts and messages. Read the methodology alongside your brief, then decide which questions need fresh answers from real participants.