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

Paper Title

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

Publication Info

• Topic area: Longitudinal evaluation of AI-driven virtual peers in promoting physical activity.
• Keywords: Physical activity, social support, AI agents, human-agent interaction, simulated exercising peers, motivation, self-determination theory, social presence, working alliance, longitudinal study.

Background and Problem

• Problem / challenge: Physical inactivity is a global health issue, and scalable, long-term motivational strategies are limited. While conversational agents show promise as exercise companions, their long-term effectiveness and relational dynamics remain unclear.
• Significance: Addressing physical inactivity could improve global health outcomes. Understanding how AI agents and human peers differently sustain motivation can inform the design of effective interventions.
• Motivation and related work: Prior research highlights the importance of social support in physical activity, with human peers offering authentic accountability but inconsistent availability. AI agents provide scalable support but face challenges in authenticity and believability. This study explores these gaps by comparing human and AI peers over six months.

Solution

• Proposed approach: Simulated Exercising Peers (SEPs) powered by Large Language Models (LLMs) designed as co-participants rather than coaches, evaluated alongside human peers and a control group.
• Novelty:
  1. Long-term empirical evaluation of LLM-powered SEPs for physical activity.
  2. Identification of the "partnership paradox": humans excel in social presence, while AI excels in reliability.
  3. Insights into how visual appearance and conversational style affect perceptions of AI peers.
• Procedure and key techniques:
  • Conducted a six-month randomized controlled trial with 280 participants in four conditions: no peer (Control), human peer (HUM), SEP with a human-like avatar (SEPH), and SEP with a cyborg-like avatar (SEPC).
  • Used a custom iOS app (Excero) to track step counts and facilitate interactions.
  • Measured outcomes using step counts, validated psychological scales (IMI, SPS, WAI), and semi-structured interviews.
  • Analyzed data using linear mixed-effects models, mixed ANOVAs, and thematic analysis.

Results

• Concrete findings:
  • Human peers (HUM) recorded higher step counts during deployment and follow-up phases compared to cyborg SEPs (SEPC).
  • AI peers (SEPH, SEPC) achieved stronger working alliance scores than human peers but lower social presence scores.
  • Participants in SEP conditions appreciated consistent encouragement but struggled to perceive AI as authentic exertion partners.
• Advantage over baselines:
  • Both human and AI peers outperformed the no-peer control in promoting physical activity during the deployment phase.
  • AI peers provided steadier, non-judgmental support, while human peers offered authentic but inconsistent accountability.
• Experiments / evaluation:
  • Quantitative: Step counts and psychological scales (IMI, SPS, WAI) analyzed across baseline, deployment, and follow-up phases.
  • Qualitative: Thematic analysis of 30 post-study interviews, identifying themes of relational dynamics, believability, and visual assessment.
• Limitations and future work:
  • Sample limited to young adults in an academic setting; results may not generalize to other populations.
  • Attrition and reliance on step counts as the primary behavioral measure.
  • Future research should explore longer-term effects, diverse populations, and richer interaction designs.

Summary

This study investigated the effectiveness of AI-driven Simulated Exercising Peers (SEPs) and human peers in supporting physical activity over six months. Human peers excelled in fostering social presence, while AI peers provided reliable and consistent support, forming stronger working alliances. These findings highlight complementary strengths: humans offer authentic connection, and AI provides steady encouragement. The study suggests that AI peers should not mimic human authenticity but instead augment it with reliability, paving the way for hybrid designs that combine human and AI strengths to sustain long-term motivation.