Why Do People Distrust AI Interviews? A 2026 Study
Qualitati Research Team · 2026-06-26 · 7 min read
Short answer: People distrust AI interviews mostly because of a gap between what they are promised and what they experience. A 2026 CSCW study of asynchronous AI interview systems found that inflated "AI-powered" and "conversational" framing set expectations the systems could not meet, leaving participants feeling unheard, surveilled, and pushed toward defensive workarounds. For qualitative researchers, the lesson is that trust in an AI-moderated interview is earned through honest framing, genuine responsiveness, and participant control — not through marketing language.
What did the 2026 study examine?
The study tracked how people actually react to AI that conducts interviews without a human in the loop. In "Expecting Too Much, Getting Too Little: Exploring the Challenges and Design Opportunities of Asynchronous AI Interviewers" (CSCW 2026), researchers Md Nazmus Sakib, Naga Manogna Rayasam, and Sanorita Dey of the University of Maryland, Baltimore County, ran a two-phase investigation of asynchronous AI interviewers used in hiring.
According to Sakib, Rayasam, and Dey (2026), the work combined a large observational analysis with first-hand accounts:
- Phase 1 — community data: roughly 18,000 Reddit entries (597 posts, 12,727 comments, and 4,961 replies) across 11 subreddits where applicants discussed AI interview experiences.
- Phase 2 — interviews: 17 semi-structured interviews to probe the concerns surfaced online.
- Interface evaluation: a follow-up study with N = 180 participants (30 per condition across six groups), grounded in Self-Determination Theory, testing how small design choices change a participant's sense of agency.
The systems studied were hiring tools, not research instruments. But the trust dynamics it documents transfer directly to any AI-moderated interview — including the AI-led research interviews and conversational surveys now common in UX and market research.
Why do people distrust AI interviews?
The core finding is a mismatch between expectation and reality. According to the authors, organizations described their tools with terms like "AI-powered" and "conversational platform," which led applicants to expect something "human-like" and ChatGPT-fluent. What they met instead often "felt like static asynchronous video recording tools." Five recurring sources of distrust stood out:
- Performing to silence. With no reciprocal interaction, participants described "performing to silence," and generic responses read as "automated flattery" rather than genuine listening. As one participant put it, "there might've been an LLM behind the screen, but to me, it still felt like no one was listening."
- Opacity about evaluation. People did not know what was being assessed — words, facial expressions, eye contact — or how their data would be used, with one likening the experience to "unpaid data entry for their algorithm."
- Accessibility gaps. Non-native English speakers hit poor accent transcription, and neurodivergent participants struggled without any reciprocal feedback.
- Defensive workarounds. Believing the system was already biased, some applicants scripted keyword-stuffed answers, used ChatGPT to polish responses, or even had "AI talk to AI." One framed it bluntly: "it's not about cheating, it's about surviving a system that already feels stacked against you."
- Some genuine upside. Not all reactions were negative — participants valued the structured format, scheduling flexibility, and reduced interviewer bias, and some with social anxiety preferred the asynchronous mode.
What does this mean for qualitative researchers?
The same forces that erode trust in AI hiring interviews threaten data quality in AI-moderated research. When a participant feels unheard or surveilled, they disengage or manage their answers strategically — the qualitative equivalent of the "deceptive practices" the study documents. The result is a transcript that looks complete but is quietly performative.
The interface study points to the fix: control and responsiveness matter more than polish. The researchers tested two design levers:
| Design lever | What was tested | Reported effect |
|---|
| Response control (re-record vs. edit vs. both) | Letting participants redo or revise an answer | Editing improved a sense of "autonomy and competence"; offering both options added "decision friction" |
| Feedback style (motivational vs. evaluative vs. combined) | How the system acknowledges each answer | Motivational feedback created "emotional safety interpreted as presence"; evaluative feedback felt "helpful but distant" |
The takeaway from Sakib, Rayasam, and Dey (2026) is that "even subtle design changes can enhance user autonomy." Translated into research practice, four principles follow:
- Frame honestly. Tell participants it is an AI interviewer, what it can and cannot do, and what happens to their data. Under-promise on "conversational" magic.
- Make it responsive. An AI moderator that reflects back and probes specifics signals listening; canned acknowledgments signal a recording booth.
- Give participants control. Let people revisit or clarify an answer rather than locking them into a one-take performance.
- Design for the margins. Test transcription on accented speech and give neurodivergent participants the feedback cues that keep them oriented.
At Qualitati, these are exactly the design questions behind our AI Interviewer, which discloses its AI role, adapts follow-up questions to what a participant actually says, and pairs with ThemeLens for thematic analysis once the conversation is done. The study is a useful reminder that trust is a precondition for good qualitative data, not a nice-to-have.
How accurate are these findings?
The study is rigorous within its scope, but two caveats matter. First, its primary context is high-stakes hiring, where applicants have more to lose than a typical research participant — some distrust is situational. Second, the Reddit phase captures people motivated to post, who may skew negative. The interface experiment (N = 180) helps balance this by testing design changes prospectively. The transferable signal is robust: opaque, non-responsive, one-take AI interviews invite disengagement and strategic answers, whatever the setting.
Frequently asked questions
Are people more honest with an AI interviewer or a human? It depends on design and topic. Other research suggests people sometimes disclose more to AI on sensitive subjects, but this 2026 study shows that when an AI interview feels opaque or unresponsive, participants instead become more guarded and strategic. Trust signals decide which way it goes.
Does this study say AI interviews don't work? No. It documents real upside — structure, flexibility, and reduced interviewer bias — alongside the trust failures. The argument is about design quality, not whether AI interviews should exist.
What's the single biggest trust lever? Honest framing plus genuine responsiveness. Participants forgive a lot when they believe they are actually being heard and they understand how their data is used.
How is research interviewing different from AI hiring interviews? Stakes and intent differ — research participants are not being judged for a job — but the mechanics of trust (transparency, responsiveness, participant control) are the same.
Last updated: June 26, 2026.
This article is an independent editorial summary of third-party research. It is not affiliated with or endorsed by the study's authors. Primary source: Sakib, M. N., Rayasam, N. M., & Dey, S. (2026). Expecting Too Much, Getting Too Little: Exploring the Challenges and Design Opportunities of Asynchronous AI Interviewers. CSCW 2026.