A new study finds that AI-run employee interviews don't get more honest answers out of workers - they just get more sophisticated hedging.
Researchers tracked 44 first-session interviews inside a global management consulting firm, where an AI voice interview followed a survey on the same themes in a single session. Across 132 matched theme observations, 20-41% of sessions (the range depends on how "favorable" was defined) showed an employee give a positive survey rating, then raise a substantive concern on the same topic once the AI interview got going. Pulling from 158 "protective" quotes across 65 of those sessions, the researchers found employees routinely softened complaints, deflected blame away from specific people, and stopped short of saying what they meant. How much hedging happened tracked closely with whether employees believed the listening process itself was legitimate.
That matters because companies are increasingly pitching AI chat and voice tools as a way to surface what structured surveys miss - a more candid channel than a multiple-choice form. This research suggests the format doesn't bypass the underlying power dynamic. If employees don't trust that speaking up is safe, they manage their disclosure in the new channel too, not just the old one. The researchers call this pattern "bounded disclosure": talk that sounds open but is still shaped by self-protection.
A chatbot asking better questions doesn't fix an org chart that punishes honest answers.