TANDE: Disentangling Verbal and Nonverbal Backchannels in Emotional AI-Avatar Conversations with Young Adults

In plain English

A within-subjects study of 36 adults aged 18-25 compared a static avatar against one that nodded, and one that nodded and said "mhm", while people talked about stressful experiences. None of the backchannel conditions significantly improved rapport, empathy or engagement, and the static avatar scored similarly or slightly higher on some measures, with the verbal cue described by several participants as intrusive. The authors' own conclusion is worth sitting with: spend the effort on what the agent says rather than on adding listening behaviours, and consider whether embodiment is needed at all. Small sample, short sessions.