Same answer, different reasons: AI-moderated interviews uncover topics and reasoning that surveys miss.

Surveys can reach large samples but capture only pre-defined answers, with identical responses often concealing very different reasoning. We tested whether AI-moderated interviews capture this reasoning at scale. In a pre-registered, randomized, counterbalanced design, 1,953 US adults from a probability-based panel and an opt-in panel completed three AI-moderated interviews as well as a section with closed-ended and open-text survey items. Topics were extracted from transcripts inductively, without a predefined codebook, and validated by a confirmatory follow-up. Participants endorsed 94% of topics as fair descriptions of what they said. 21-51% of topics from the AI interview had no counterpart in the closed-ended battery, and 17-37 percentage points net coverage over open text answers. Among respondents who gave identical closed-ended responses, their reasoning diverged widely in valence and content. AI-moderated interviews surface reasons that closed-ended and open-text items miss, and show how reasoning differs among people who answer alike.