Invisible Impact of Empathy on Behavioral Change: Isolating the Effect of Empathy in Long-term Physical Activity Coaching Chatbot Interactions

Synthesis: Siyan et al. (2026) conduct a carefully controlled experiment isolating the effect of empathetic language in LLM-powered physical activity coaching chatbots over a longitudinal deployment. While the empathy condition did not directly increase exercise behavior, it significantly improved users' sense of being understood, which in turn predicted sustained engagement with the coaching system. This finding has direct relevance to Affective Tutoring research in education: AI tutors that express empathy may not directly boost learning outcomes, but may sustain engagement long enough for learning to occur. The work connects to Personalized Learning system design by showing that affective features like empathy serve a relationship-maintenance function rather than a direct instructional one. The study also contributes to Large Language Models (LLMs)-based educational tool design by demonstrating that careful experimental isolation is needed to understand which conversational features actually drive outcomes. While the study's education implications are indirect (health coaching rather than classroom learning), the mechanism of empathy-sustained engagement transfers to Student Experience in any long-term AI-mediated learning relationship, particularly in Feedback Loop contexts where sustained interaction is critical.

What this means for practice

• Instructors. Use empathy to hold learners in the relationship, not to move outcomes directly. In this study empathetic language did not raise step counts, but it produced significantly faster growth in intention to follow advice and self-efficacy, so treat it as affective scaffolding around the work rather than the instruction itself.
• Designers. Start a long-term AI coaching or tutoring relationship with the empathetic variant: initiating the six-week period with the empathetic chatbot produced the least attrition.
• Designers. Do not tune a chatbot from users' ratings of how empathetic it feels. Participants repeatedly failed to tell the three versions apart, and the non-empathetic version scored well on engagement and usefulness, so perceived-empathy self-reports are an unreliable design target.
• Researchers. Measure change within each condition rather than raw satisfaction levels. Because a within-subject design lets the previous version contaminate the next, weekly deltas were more informative than raw check-in ratings in this study.
• Instructors. Keep the long horizon in view when evaluating conversational support: the effect of empathy appeared only as an interaction with time (β = 0.011 on both intention to follow and self-efficacy), not in any single session.

Limitations

• Thirteen participants produced the analyzed data out of 33 enrolled. Sixteen dropped out before interacting with all three chatbot versions, three did not complete surveys consistently, and one was excluded for an incorrect initial setup.
• The 13 participants were aged 19–30 and mostly women (10 of 13), recruited through social media and by emailing computer science students at the authors' own institution — the authors flag the gender and age skew themselves; the pilot that preceded the main study drew 21 Prolific participants, unevenly distributed (8 non-empathetic, 4 standard, 9 empathetic), and its results are reported descriptively without statistical testing.
• Engagement and progress rested on self-report: intention to follow, self-efficacy, and step counts, with no objective activity tracking, and the authors note they emphasized tracking steps partly to improve the accuracy of self-reported counts.
• The within-subject design means time spent with one chatbot version may carry over into how the next is perceived, and the weekly deltas showed no statistically significant differences under linear mixed models; the intervention also targeted physical activity coaching rather than classroom learning, so its education relevance is by analogy only.

Citation

Li Siyan, Kai-Hui Liang, Shopnil Shahriar, Yilin Ye, Shiyoh Goetsu, Wei-Wei Du, Masahiro Yoshida, Tsunayuki Ohwa, Xuhai Xu, Zhou Yu (2026). Invisible Impact of Empathy on Behavioral Change: Isolating the Effect of Empathy in Long-term Physical Activity Coaching Chatbot Interactions.