Short answer: Yes, with limits. A 2026 study of 21 qualitative researchers found that rehearsing an interview guide with LLM-simulated participants led most of them to revise who they would recruit and what they would ask. The same study shows the risks: simulated answers are too cooperative to expose weak questions, and AI-generated themes can anchor real fieldwork.
Most research on synthetic participants asks whether their answers resemble real people's. A new paper asks a more practical question: what happens when qualitative researchers use simulated participants to rehearse an interview protocol before fieldwork? \"Inspectable LLM-Simulated Participants for Qualitative Interview Rehearsal\" is by Hangyue Zhang, Qiran Hu, Ziyi Zhang, Hyanghee Park and Yun Huang of the University of Illinois Urbana-Champaign. It was posted to arXiv (cs.HC) on 19 October 2026 as a preprint.
\n\nIt means running a draft interview guide against LLM personas before any human is interviewed, then using what comes back to revise the guide. Traditional piloting does the same job with real people, but access to suitable participants is often the bottleneck. This is especially true for hard-to-reach groups such as clinicians.
\nThe authors argue that earlier work judged simulated participants by how human-like their output looked. That says little about how researchers actually use and critique the material while designing a study. Rehearsim was built as a research probe to study exactly that.
\n\nIt runs in three stages, each kept inspectable and under the researcher's control:
\nThe tool never rewrites the protocol itself. Every revision is made by the researcher.
\n\nAccording to Zhang et al. (2026), 21 researchers took part in 1.5–2 hour sessions. They included Master's and PhD students, postdocs, faculty and UX researchers from HCI, social science, education, health, communication and design. Their experience ranged from one or two qualitative studies (6 participants) to six or more (6 participants). All had used generative AI in research, 16 (76.2%) to develop interview questions, but only 3 (14.3%) to simulate participants.
\nEach person brought an early-stage study, entered their initial protocol, rated it, used the tool while thinking aloud, revised the protocol, and rated it again. The team analyzed system logs, before-and-after protocols, surveys, think-aloud data and exit interviews.
\n\nIt changed them, though the study cannot show the changes made them better. Seventeen of 21 participants revised their protocol. Those who did modified 1.94 of three components on average and changed about 38 words. Research context, target population and interview questions were each revised by 11 participants.
\nThe coded revisions targeted breadth (12 participants), target population (11), research context (10), coverage (9), depth (8) and methodological rigor (8). One researcher studying why people use AI for some tasks but not others saw a generated \"AI avoider\" subgroup and added people with negative and neutral attitudes to her sample. Another, studying online health-information seeking, saw a theme about what happens after searching and added a whole stage to her guide.
\nFour participants made no changes. Two felt the simulation lacked context about the systems they were studying. One chose not to follow up on a surprising answer because it came from an AI. One considered his protocol already well developed.
\n\nParticipants rated their protocols higher afterwards on coverage, depth, breadth, rigor, context clarity and feasibility. But when the authors compared those ratings with what had actually been revised, agreement was weak: 51.8% of pairs matched (κ = 0.14). In 70 of 168 comparisons (41.7%), participants reported improvement in a dimension where no revision was coded. The reverse happened in only 11 (6.5%).
\nAt the person level, those who revised more dimensions did report broader improvement (Spearman's ρ = 0.58). The authors also note there was no comparison condition, so some of the gain may come from simply reflecting on and rewriting a protocol.
\n\nThe benefits and risks appeared together at every stage:
\n| Stage | Benefit researchers reported | Risk researchers reported |
|---|---|---|
| Simulated responses | Surfaced missing topics and new follow-up directions | Answers were articulate and complete, so weak questions looked fine and there was little to probe |
| Synthetic provenance | Low-stakes space to practice probing | Surprising details were dismissed with \"it's AI\" instead of explored |
| Generated subgroups | Revealed overlooked participant variation | Simulated diversity exceeded what could realistically be recruited |
| Generated themes | Fast overview of what the guide elicits | Themes could anchor expectations before any human data exist |
| Unfamiliar suggestions | Prompted reflection on blind spots | Hard to tell a genuine insight from a model error |
Nine of the 21 participants said the simulations made interviewing too easy. A researcher who works with Deaf and Hard of Hearing participants noted that the AI's language level left no room to practice the rephrasing and repair real interviews need. Others worried about later misremembering a simulated theme as one that recurred in real data.
\n\nSimulated rehearsal looks most useful for three things: widening who you plan to recruit, spotting missing topics, and training novice interviewers to probe. It is weakest at testing whether a question is clear, which is the job human piloting still does best.
\nA practical protocol, based on the paper's findings:
\nIf you run interviews with an AI Interviewer, a short test run on a few contrasting personas can catch gaps in the guide before launch. A Digital Twin Panel can play a similar rehearsal role, as long as its output is treated as preparation rather than evidence.
\n\nNo. The authors explicitly position simulated material as rehearsal, not evidence or a substitute for human participants, and their participants found simulated answers too smooth to test question clarity.
\nParticipants said many single functions could be reproduced in a general chatbot. What Rehearsim added was structure: literature-grounded profiles, several participants compared side by side, and themes traceable back to the turns that produced them.
\nParticipants feared so. Several said they might enter fieldwork expecting the generated themes, which is a particular concern in grounded theory, where researchers try not to presuppose answers.
\n\nLast updated: 2026-10-21.
\nThis is an independent editorial summary of third-party research by the Qualitati Research Team. Qualitati is not affiliated with the authors. Read the original paper for full methods and results.
\n","related":[{"slug":"synthetic-participants-pilot-study-design-2026","title":"Can Synthetic Participants Pilot Your Study Design? (2026)","description":"Can synthetic participants pilot your study design? A 2026 study found LLM reasoning traces caught more survey flaws than a paid human pilot or a chatbot review.","category":"Methodology","date":"2026-10-20","readTime":"7 min read","author":"Qualitati Research Team"},{"slug":"ai-follow-up-interviews-vs-fixed-surveys-experiment-2026","title":"AI Follow-Up Interviews in Surveys: A 2,243-Person Test (2026)","description":"Do AI follow-up interviews improve survey answers? A 2026 POQ experiment with 2,243 people found richer reasons, stable satisfaction, and some answer shifts.","category":"Methodology","date":"2026-10-16","readTime":"7 min read","author":"Qualitati Research Team"},{"slug":"batching-synthetic-respondents-inflates-correlations-2026","title":"Batching Synthetic Respondents Inflates Correlations (2026)","description":"Generating many synthetic respondents in one LLM request inflates correlations, a 2026 LSE study finds. What it means for synthetic survey data validity.","category":"Methodology","date":"2026-10-13","readTime":"7 min read","author":"Qualitati Research Team"}]}