Real-Time AI Interview Assistants: What a 2026 Study Found

Qualitati Research Team · 2026-09-30 · 7 min read

Short answer: Real-time AI interview assistants can help human interviewers probe more. In a 2026 study of 18 interviewers, AI suggestions raised follow-up questions from about 6 to 9–10 per session without replacing the interviewers' own questions. But interviewers only accepted the AI as a quiet assistant, not as an evaluator, competitor, or co-interviewer.

Most research on AI in qualitative interviewing asks whether an AI can run the interview itself. A new preprint asks a different question: what happens when the human stays in the chair and an AI whispers suggestions on the side? According to Liu, Dai and McGrenere (2026), the answer is more probing, slightly longer interviews, and a set of social tensions that tool designers and research teams should plan for.

What did the study test?

The study tested a prototype called ProbeAssist that shows an AI panel next to a normal video call, visible only to the interviewer. The authors describe it in The Interviewer's Perspective: Unpacking the Impact of Real-Time AI Interviewing Assistance on Social Dynamics (arXiv, submitted September 24, 2026, under review for CHI 2027). The interviewer uploads the interview guide and research objectives, and the system transcribes the conversation live and analyses each answer against the guide.

Two variants of assistance were compared against a no-AI baseline:

Condition	When help appears	What it shows	Mean follow-up questions	Mean length (min)
No AI	Never	Nothing	5.83	10.12
Restrained	On demand (interviewer clicks)	Summary of the answer plus probe keywords	9.83	13.76
Expressive	Proactively, when the answer ends	Ready-to-use follow-up questions	8.72	12.65

Eighteen people with semi-structured interviewing experience (1–10 years, 3.6 years on average) each ran three short simulated interviews of 10–15 minutes: one without AI, then one with each variant in counterbalanced order. The interviewee was a trained research assistant following a role-play protocol, so every interviewer faced a similar conversational partner. The authors then ran a debriefing interview with each participant and analysed those transcripts with reflexive thematic analysis. The quantitative measures are supporting evidence; the study is qualitative-first.

Does real-time AI assistance improve probing?

Yes, in this setting it increased the amount of probing, and the extra questions were added on top of the interviewer's own. Participants asked 9.83 follow-up questions with the restrained variant and 8.72 with the expressive variant, compared with 5.83 without AI (p = .005 for both comparisons). Of those, about three per session were AI-assisted (3.39 and 3.00). The number of self-formulated follow-ups stayed close to the no-AI level (6.44 and 5.72 versus 5.83). In other words, the AI did not crowd out the interviewer's own questions.

Participants also rated interview quality higher with AI support (6.04 restrained and 5.90 expressive, versus 5.56 without AI, on a 7-point scale), and reported a high sense of control in both variants (6.28 and 6.21). Self-reported cognitive load did not differ significantly across the three conditions.

Two caveats matter. Interviews got longer (13.76 and 12.65 minutes versus 10.12), which is expected if you ask more questions. And "more follow-ups" and "higher self-rated quality" are not the same as richer data: the study did not have independent judges assess the depth of the transcripts.

What roles do interviewers accept from an AI assistant?

Interviewers accepted the AI as a tool or assistant and rejected it as an assessor, a competitor, or a co-interviewer. Sixteen of the 18 participants framed the prototype as a tool; those used to interviewing with a human partner compared it to a team member or research assistant.

Assessor. Unsolicited evaluation, such as a suggestion to move on when the interviewer wanted to keep probing, made participants feel judged and pulled their attention away from the interviewee.
Competitor. Oddly, very good suggestions could feel threatening. Some participants felt out-performed and questioned their own role.
Co-interviewer. The firmest line: the AI should never speak to the interviewee or direct the interview. Participants cited limited trust in AI judgement and their duty of care, for example not pushing a follow-up when an interviewee seems upset.

What are the hidden costs of AI interview assistance?

The study found costs that do not show up in accuracy benchmarks: evaluation effort, social friction, and threats to ownership.

Good suggestions cost more attention than bad ones. A poor suggestion is dismissed in a glance. A good one must be compared with the interviewer's own plan, which participants described as a context switch. An interesting but off-topic idea could also pull them away from the guide.
Latency becomes a social problem. Average generation time was 3.3 seconds for the restrained variant and 1.9 seconds for the expressive one. That is short technically, but participants felt the pauses as awkward silence in front of an interviewee who could not see why they were waiting. Late suggestions were often useless because the conversation had already moved on.
Ownership and creativity. One participant coined the term "creative shadowing" for an AI idea that covers up the interviewer's own emerging line of inquiry. Others worried that widespread use could make interviewing more uniform.

Preferences also split. Some wanted on-demand help and keywords to keep their own phrasing; others found full questions and proactive prompts easier. The same person could want different modes at different moments in one interview.

What this means for researchers and research teams

The authors derive three design implications: keep AI assistance subordinate and low-visibility, let interviewers shift it between "follow my thinking" and "show me something new", and judge it by whether it protects the interviewer's presence with the interviewee. For teams using live AI support today, that translates into practical rules:

Start on demand. Default to help that appears when asked, and let experienced interviewers switch on proactive prompts.
Prefer summaries and keywords when you want interviewers to keep their own voice; ready-made questions are faster but more likely to displace the interviewer's idea. See our guide to spotting leading questions before reading any AI-written probe aloud.
Never let the assistant grade the interviewer mid-session. Save feedback for afterwards.
Log what was suggested and what was used, so the influence of AI on your data can be reported.

These principles match how Qualitati's Active Listener mode is designed: the researcher runs a semi-structured interview while the tool transcribes and offers optional follow-up suggestions, without ever speaking to the participant. When you want the AI to lead instead, for example in large pilot rounds, the AI Interviewer runs text or voice interviews on its own.

Limitations of the study

It is a preprint and has not yet completed peer review.
The sample is small (18 interviewers) and interviews were short, simulated, and online.
A single role-playing interviewee was used, which may have introduced carry-over effects, and interviewees' own experience was not studied.
Stakes were low because participants did not analyse the data they collected.

FAQ

Do AI interview assistants make interviewers ask better questions?

In this 2026 study they made interviewers ask more follow-up questions (roughly 9–10 instead of about 6) and interviewers rated their interviews higher. Whether the resulting data are deeper was not independently measured.

Should an AI suggest full questions or just keywords?

Both have trade-offs. Keywords preserve the interviewer's phrasing and control; full questions are quicker to use but more likely to override the interviewer's own idea. Participants split on this, so make it configurable.

Is it acceptable for the AI to ask the participant questions directly?

Not in a human-led interview, according to these participants. They saw it as taking over control and responsibility for the interviewee's well-being. If you want the AI to ask questions, use a fully AI-moderated design and tell participants.

How fast does real-time AI assistance need to be?

Faster than it sounds. Average delays of 1.9 to 3.3 seconds still created awkward pauses, because the interviewee cannot see why the interviewer is waiting.

Last updated: September 30, 2026

This article is an independent editorial summary of third-party research by the Qualitati Research Team. It is not affiliated with or endorsed by the study's authors. Consult the original paper for full methods and results.