Study finds empathetic chatbots boost step counts despite user preference for blunt versions

A new study from Columbia University and other institutions challenges the assumption that empathy is always essential in conversational AI, especially for behavior change. The researchers built three WhatsApp-based physical activity coaching chatbots that were identical except for their empathy levels (low, medium, high). In a six-week within-subject study with 13 participants, each user interacted with all three versions. Surprisingly, participants could not reliably distinguish which chatbot was more empathetic, and the non-empathetic version was actually rated as more engaging and useful on subjective scales.

However, objective metrics told a different story. Higher-empathy variants were associated with a larger overall increase in daily step counts and faster improvement in users' intention to follow the chatbot's advice. The authors explain this mismatch through the Elaboration Likelihood Model's peripheral route: users may not consciously perceive empathy, but it still influences motivation and trust through subtle, non-conscious cues. This suggests that next-generation coaching chatbots should not simply maximize perceived empathy, but balance it with effectiveness. The study highlights design implications for long-term behavior change interventions where the invisible impact of empathy can drive real outcomes.