When the Interviewer Is a Bot: Behavior, Breakdowns, and Trust in MLLM-Led Interviews

A study explored the behavior, breakdowns, and trust dynamics when real-time multimodal Large Language Models (MLLMs) conduct semi-structured interviews. Researchers developed 'InterviewBot' to observe default MLLM interviewing behavior in a practice study with 15 participants. The preliminary analysis indicates that the MLLM interviewer's behavior is heavily reliant on acknowledgments.

Why it matters

This research explores the application of advanced AI in traditionally human-centric processes like qualitative research. Understanding the capabilities and limitations of MLLMs in interviewing roles is crucial for strategic resource allocation, process automation, and maintaining research integrity in an evolving technological landscape.

Key insights

• Semi-structured interviews, though foundational for qualitative research, are labor-intensive.
• The study deployed 'InterviewBot', a voice-based interviewing system utilizing an off-the-shelf real-time MLLM.
• The MLLM was observed as a research instrument, not a novel architectural solution, to understand its default interviewing behaviors.
• A practice study (N=15) involved participants completing bot-led interviews, followed by human-led reflection sessions.
• Turn-level behavioral analysis (N_turns=428) of the MLLM interviewer showed an acknowledgment-heavy interaction style.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.10412