When Workout Buddies Are Virtual: AI Agents and Human Peers in a Longitudinal Physical Activity Study

TL;DR

This study tackles how to sustain physical activity through scalable social support by comparing four conditions: exercising alone, with a human peer, and with AI-based Simulated Exercising Peers (SEPs) in human-like (SEPH) and cyborg-like (SEPC) embodiments. Using a six-month randomized controlled design (N = $280$) and mixed methods, the authors find a partnership paradox: human peers evoke stronger social presence, while AI SEPs provide steadier, more reliable working alliances. Quantitative results show AI peers elevate working alliance and competence but do not consistently outpace human peers in activity, whereas humans produce higher social presence and more variable engagement. Qualitative data reveal believability hinges on behavioral coherence and role-fit rather than surface appearance, supporting a design emphasis on reliability and adaptive social strategies rather than fully mimicking human authenticity. The findings advocate hybrid designs that blend human authenticity with AI consistency to sustain physical activity at scale while clarifying boundaries and expectations in AI-assisted health interventions.

Abstract

Physical inactivity remains a critical global health issue, yet scalable strategies for sustained motivation are scarce. Conversational agents designed as simulated exercising peers (SEPs) represent a promising alternative, but their long-term impact is unclear. We report a six-month randomized controlled trial (N=280) comparing individuals exercising alone, with a human peer, or with a large language model-driven SEP. Results revealed a partnership paradox: human peers evoked stronger social presence, while AI peers provided steadier encouragement and more reliable working alliances. Humans motivated through authentic comparison and accountability, whereas AI peers fostered consistent, low-stakes support. These complementary strengths suggest that AI agents should not mimic human authenticity but augment it with reliability. Our findings advance human-agent interaction research and point to hybrid designs where human presence and AI consistency jointly sustain physical activity.

Figures (5)

• Figure 1: The architecture of the intervention illustrating the communication and directionality of the content between the different components.
• Figure 2: Screenshots of the Duo views (Simulated Exercising Peer as a cyborg in SEPC, and as a human in SEPH) and the Progression view from the Excero mobile application.
• Figure 3: Screenshots of the Chat view (SEPC, SEPH, and HUM conditions), the History view, and the default view displayed outside the deployment phase in the Excero mobile application.
• Figure 4: Weekly average step counts across conditions (top panel), with overall week effect (middle panel) and condition-adjusted residuals (bottom panel). Vertical dashed lines indicate the three study phases: Baseline, Deployment, and Follow-up. Human-Human pairs (HUM) consistently maintained higher step counts than cyborg peers (SEPC) during Deployment and Follow-up, while both SEP conditions showed trajectories closer to Control.
• Figure 5: Questionnaire outcomes across time and conditions. (A) IMI Perceived Competence, (B) IMI Relatedness, (C) Social Presence, (D) WAI Bond. Error bars represent ±1 SE.