Can AI conduct real research interviews? A 2026 study by Wuttke and colleagues ran AI-led conversational interviews with 571 respondents on migration attitudes and found that AI interviews surfaced reasoning that standardized surveys miss — and that participants rated the AI interview at or above the survey. AI interviewing works, but the delivery mode matters a lot.
\n\nThe researchers tested whether a large language model can run open-ended, semi-structured interviews at survey scale. In their paper AI Conversational Interviewing: Scaling Up Semi-Structured and In-depth Interviews (Wuttke, Lang, Klamm, Würschinger & Kreuter, 2026), respondents from the Prolific and Payback panels were interviewed about migration policy by an AI moderator, then completed a matching standardized survey so the two data-collection methods could be compared head to head.
\nCrucially, the study was designed around a long-standing tension in research: the trade-off between depth and scale. In-depth human interviews are rich but slow and expensive; surveys scale but flatten nuance. The question was whether an AI interviewer could deliver some of the depth of a qualitative interview while keeping the reach of a survey.
\n\nAccording to Wuttke et al. (2026), the text-based interviewer ran on OpenAI's GPT-4o (model gpt-4o-2024-11-20, temperature 0.8), and the voice-based interviewer used OpenAI's GPT Realtime model (gpt-realtime-2025-08-28, temperature 0.7). Notably, the team reported that switching the text interviewer to GPT-5 caused a \"marked deterioration in interviewer performance\" — a useful reminder that a newer model is not automatically a better interviewer.
Respondents were randomly assigned across three conditions:
\nOf 1,039 respondents initially assigned (with no significant imbalance across conditions, χ²(2)=3.67, p=0.159), the final analytic sample was N=571 with complete interviews linked to survey records.
\n\nVoice interviews produced substantially more content. According to Wuttke et al. (2026), the numbers break down as follows:
\n| Metric | Voice interview | Text interview |
|---|---|---|
| Mean words per respondent | 608.8 | 299.8 |
| Mean session length | 8.3 min | 11.5 min |
| Speaking/typing rate | 52.5 words/min | 21.1 words/min |
In short, voice interviews generated roughly twice as many words as text interviews and did so faster, because speaking is quicker than typing. But there is a catch: attrition was concentrated in the voice-only condition, and when respondents were free to choose, they \"overwhelmingly selected the chat modality.\" The authors attribute the voice drop-off to \"addressable imperfections\" in their implementation rather than a fundamental limit of voice interviewing — a signal that voice UX, not the concept, is the bottleneck.
\n\nPositively. Among respondents who completed the interview, evaluations of the AI interview were at or above those of the standardized survey across all modes (Wuttke et al., 2026). That matters because a common objection to AI interviewing is that it will feel cold or frustrating; here, the completers preferred it to the familiar survey format. The open question the study leaves is completion, not satisfaction: getting people to finish — especially in voice — is where the design work remains.
\n\nThe headline methodological finding: conversational transcripts exposed \"considerations and reasoning that a comprehensive standardized battery does not capture.\" The AI interviews surfaced markedly different mental models of migration among respondents who had similar attitude scores on the survey. Two people can click the same point on a 7-point scale for completely different reasons — and only the open-ended conversation makes that visible. This is the core argument for AI interviewing: it recovers the \"why\" behind the number without abandoning scale.
\n\nFor academic, UX, and market researchers, three practical takeaways stand out:
\nIf you want to run this kind of study, QualiTaTi's AI Interviewer conducts voice or text interviews at scale and transcribes them automatically, and ThemeLens can then code the resulting transcripts for themes — turning the \"why\" the study describes into structured analysis.
\n\nNot wholesale, but it can scale interviewing dramatically. The 2026 study shows AI-led interviews capture reasoning surveys miss and are rated favorably by participants, making them a strong fit for large-N opinion and experience research where recruiting human interviewers would be impractical.
\nThey serve different goals. Voice produced about twice the word count (608.8 vs. 299.8 words) and faster responses, but had higher attrition; text is what most respondents chose when free to decide. For maximum richness, use voice with strong UX; for completion and comfort, offer chat.
\nThe text interviewer used OpenAI's GPT-4o and the voice interviewer used OpenAI's GPT Realtime model. The authors found GPT-4o outperformed GPT-5 for the interviewing task, showing that model choice should be validated, not assumed.
\nYes — the study found respondents with identical survey attitude scores held substantively different underlying mental models that only the conversational data revealed. AI interviews recover the reasoning behind the numbers.
\n\nPrimary source: Wuttke, A., Lang, M. M., Klamm, C., Würschinger, Q., & Kreuter, F. (2026). AI Conversational Interviewing: Scaling Up Semi-Structured and In-depth Interviews. arXiv:2606.20064.
\nLast updated: July 31, 2026. This article is an independent editorial summary of third-party research and is not affiliated with or endorsed by the study's authors.
\n","related":[{"slug":"digital-twins-survey-respondents-2026-study","title":"Can Digital Twins Replace Survey Respondents? A 2026 Study","description":"A 2026 Columbia Business School study found LLM digital twins correlate with real survey answers at only r = 0.20 across 164 outcomes: funhouse mirrors, not substitutes.","category":"Guides","date":"2026-07-29","readTime":"8 min read","author":"Qualitati Research Team"},{"slug":"is-ai-qualitative-coding-biased-2025","title":"Is AI Qualitative Coding Biased? What a 2025 Study Found","description":"A 2025 Sociological Methods & Research study found ChatGPT and Llama 2 coding errors correlated with respondent traits; a small supervised model was more accurate and less biased.","category":"Guides","date":"2026-07-28","readTime":"8 min read","author":"Qualitati Research Team"},{"slug":"best-llm-thematic-analysis-reliability-study","title":"Which LLM Is Best at Thematic Analysis? A 2026 Study","description":"A December 2025 study ran Gemini 2.5 Pro, GPT-4o and Claude 3.5 Sonnet on an interview transcript: all passed kappa 0.80 for inter-rater reliability, with Gemini highest.","category":"Guides","date":"2026-07-27","readTime":"8 min read","author":"Qualitati Research Team"}]}