What is an AI-moderated interview and how does it work?
How the method works step by step, what the published evidence says about quality, and when a human moderator is still the better choice.
An AI-moderated interview is a research interview in which an AI model asks the questions, follows a discussion guide and probes each answer, by text or voice, with a real participant. It lets a team run hundreds of short qualitative interviews within hours. In a 2025 LSE study, sociologists graded AI-led interviews close to, but below, face-to-face human experts.
How does an AI-moderated interview work, step by step?
A researcher writes the guide and the rules; the AI conducts each conversation; the researcher reviews and analyses the transcripts.
- Write the guide. Set the objective, the topics to cover, the questions that must be asked and the rules for probing, such as "ask for a concrete example whenever an answer is abstract".
- Recruit and inform real participants. People come from a panel, a customer list or an invitation link. Tell them an AI will interview them and where their answers will be processed. Researchers at LSE note that interviews run on commercial models require participants to agree to their data being processed by the model's provider (Geiecke and Jaravel, 2025).
- The AI runs the conversation. It asks one open question at a time, listens to the answer and decides on a follow-up: clarify, ask for an example, or move to the next topic. It can do this by text or voice, with many participants at once.
- Transcripts are coded. Answers are grouped into themes, often with AI help, and a researcher checks a sample of the coding by hand.
- The researcher interprets. People decide what the evidence means and what to do about it.
The design choices that matter most sit in step 1. Geiecke and Jaravel built their interviewer on established practice from sociology: non-directive questions, follow-ups that make sure each point is clear, and requests for concrete rather than abstract evidence.
How good are AI-moderated interviews compared with human moderators?
Close, on the published evidence, but the evidence is young and covers short interviews.
In an LSE evaluation, eight sociology PhD students graded interview transcripts on a 1 to 5 scale without being told who or what had conducted them (Geiecke and Jaravel, 2025):
| Interview type | Grade against a top face-to-face expert | Grade against a top online text-chat expert |
|---|---|---|
| Human expert, face to face | 3.6 | 3.6 |
| AI interviewer, voice | 3.4 | 3.9 |
| AI interviewer, text | 2.8 | 3.2 |
| Human expert, online text chat | 2.3 | 2.8 |
Against a face-to-face standard, human experts in the room scored highest; against a text-chat standard, AI voice interviews did. The authors sum it up as AI-led interviews "approaching the performance of human experts".
In the same paper, a sample of 466 US adults wrote 142% more words in AI-led interviews than in open text fields, and some respondents said they preferred the AI because it felt non-judgmental. A separate randomised study, in which university students were interviewed by either an AI or a human using the same questionnaire on political topics, found the AI produced data of comparable quality (Wuttke et al., 2024).
What is not yet known: how AI moderators do in long interviews, with senior business buyers or experts, on distressing topics, or in languages other than English. The LSE graders also disagreed with each other a fair amount (two grades of the same transcript correlated at 0.41), which is a reminder that interview quality is hard to score.
When does an AI-moderated interview fit, and when is a human moderator better?
Use AI moderation for many short interviews on a defined topic; use a person when the conversation itself has to be judged, steered or cared for. The table is practitioner judgement informed by the studies above, not a tested finding.
| AI-moderated interviews fit when | A human moderator is better when |
|---|---|
| You need dozens to thousands of interviews on one topic | You need a few long, exploratory interviews where the guide may change mid-study |
| You would otherwise use open text questions in a survey | The participants are senior buyers or experts who expect a knowledgeable peer |
| Participants are spread across time zones and schedules | The topic is distressing or sensitive and the participant may need care |
| A non-judgmental interviewer could make people more candid | You need to observe behaviour, a screen or a room, not only words |
| You want the same probing rules applied to every participant | The research depends on group dynamics, as in a focus group |
What should you check before running one?
Six checks catch most problems.
- Disclosure. Tell participants an AI is interviewing them, at the start. In the EU this has been a legal duty for providers since 2 August 2026, under Article 50(1) of the AI Act.
- One question at a time. The LSE authors found current models "struggle to always restrict their message" to a single question; read pilot transcripts for stacked questions.
- A pilot you read in full. Run a handful of interviews and read every transcript before launching.
- Fake and AI-written answers. The same study saw some panel respondents use AI to answer. Probing makes these easier to spot than in open text fields, but check.
- Coding checks. Hand-code a sample of transcripts and compare with any automated themes.
- Version records. Interviews change when the underlying model changes, so record the model, guide and prompt version with the data.
PulseLake runs AI-led interviews: conversational and avatar-based AI interviewers that follow a protocol, probe adaptively and interview real participants, with each conversation turned into structured evidence inside the study. Researchers keep judgment and approvals. Our FAQ describes the research modes in more detail.
Frequently asked questions
Is an AI-moderated interview the same as a chatbot survey?
No. A survey asks every person the same fixed questions. An interview follows a guide but chooses its follow-up questions from what each person says, so two transcripts on the same topic can go in different directions.
Do participants have to be told that an AI is interviewing them?
They should be, as a matter of research ethics. In the EU, Article 50(1) of the AI Act has required since 2 August 2026 that AI systems which interact with people are designed so the person is told, from the first interaction, that they are dealing with an AI.
How long should an AI-moderated interview be?
The published evidence covers short interviews. The LSE study notes that its online panel respondents expected to spend under thirty minutes, so its AI interviews were much shorter than typical sociology interviews. Longer AI interviews have been studied far less.
Can an AI moderate a focus group?
The studies cited here test one-to-one interviews only. A focus group depends on managing interaction between participants, and we found no comparable published evaluation of AI moderators doing that.
Sources
Sources for the facts on this page, last checked October 9, 2026.
- Geiecke and Jaravel, Conversations at Scale: Robust AI-led Interviews with a Simple Open-Source Platform (LSE, NBER conference version, 16 July 2025) checked October 9, 2026
- Wuttke et al., AI Conversational Interviewing: Transforming Surveys with LLMs as Adaptive Interviewers (arXiv 2410.01824; LaTeCH-CLfL 2025) checked October 9, 2026
- European Commission, Transparency obligations under Article 50 AI Act (FAQ, updated 24 July 2026) checked October 9, 2026
Related guides
- What questions should you ask in a concept test?A concept test asks the same questions about every concept: appeal, relevance, uniqueness, believability, purchase intent, price and why. A ready-to-use set.
- Qualitative coding at scale
- Where AI actually fits in the research workflow
- PulseLake FAQ: research modes, including AI-led interviews
Run research end to end. Keep the knowledge working.
One AI-native operating system for market research and insight professionals — from study design and evidence generation to agents, institutional knowledge, delivery and action.