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

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.

💡 Research Summary

This paper reports a six‑month randomized controlled trial (N = 280) that examined how large‑language‑model (LLM)‑driven simulated exercising peers (SEPs) compare with human workout partners and a no‑partner control in sustaining physical activity. Participants (young adults, average age 22.4) were randomly assigned to one of four conditions: (1) human‑human dyads (HUM), (2) an AI SEP embodied as a human‑like avatar (SEPH), (3) an AI SEP embodied as a cyborg‑like avatar (SEPC), or (4) a control group with no partner (CON). All participants wore smartwatches that recorded daily steps and active minutes, and they set weekly activity goals through a study app. Human partners exchanged real‑time progress updates and encouragement messages, while the AI SEPs delivered daily, personalized motivational messages, feedback, and goal adjustments generated by an LLM. The study measured objective activity data, weekly self‑report scales of social presence, working alliance, and relatedness satisfaction, and conducted post‑study interviews for qualitative insight.

Quantitative analysis used mixed‑effects models to assess time, condition, and their interaction effects. Results revealed a “partnership paradox.” Human partners generated the highest social presence scores (M = 4.7/5) and produced a sharp activity boost in the first two months (+23 % over baseline), but their impact waned after month four as partner availability and engagement declined. AI SEPs showed lower initial boosts (+10 %) but maintained a steady increase throughout the study; by month six, their total activity gain (≈+18 %) was comparable to the human condition. Working alliance scores were highest for the AI conditions (SEPH = 4.5, SEPC = 4.3) and significantly exceeded both the human and control groups, indicating that participants perceived the AI as a reliable, task‑focused collaborator. The human‑like avatar performed slightly better than the cyborg avatar on social presence and trust, but both outperformed the control.

Qualitative findings complemented the numbers. Participants described human partners as sources of authenticity, competition, and accountability—key drivers of early motivation. However, they also noted variability in partner responsiveness and occasional demotivation when the human peer was unavailable. AI partners were praised for their consistency, 24/7 availability, and non‑judgmental tone; users grew to trust the AI after repeated, predictable interactions despite initial skepticism about the agent’s lack of “real effort.” The cyborg avatar was perceived as futuristic and interesting but less warm than the human‑like avatar.

The authors introduce the concept of the partnership paradox to capture these complementary strengths and propose a hybrid intervention model: leverage human peers for the initial motivation phase (capitalizing on authenticity and social comparison) and transition to AI SEPs for sustained support (leveraging reliability and continuous alliance building). They discuss design implications, emphasizing that AI agents need not mimic human authenticity perfectly; instead, they should focus on behavioral consistency, contextual responsiveness, and transparent goal‑management.

Limitations include the homogenous sample of university students, which restricts generalizability across ages and cultures, and the lack of detailed discussion on AI transparency, potential hallucinations, and ethical safeguards. The visual design was limited to two avatar styles; future work should explore multimodal cues (voice, gestures) and test the hybrid model in broader health outcomes (e.g., weight, cardiovascular markers).

In sum, this study provides robust longitudinal evidence that LLM‑powered AI workout buddies can sustain physical activity as effectively as human partners, especially by delivering a reliable working alliance. The findings suggest that future digital health interventions should strategically combine human authenticity with AI consistency to address the enduring challenge of motivating long‑term physical activity.