Simulated Exercising Peers (SEPs)

• Simulated Exercising Peers (SEPs) are AI virtual agents designed to act as co-participants, using social comparison and companionship to motivate physical activity.
• The system integrates real-time step tracking, dynamic simulated progress via GPT-4o, and personalized motivational messaging within a mobile app environment.
• Empirical findings indicate that while SEPs provide reliable, nonjudgmental support and enhance working alliance, they may lack the authentic social presence of human peers.

Simulated Exercising Peers (SEPs) are AI-driven virtual agents designed to function as workout companions—distinct from coaches—whose principal role is to leverage social mechanisms such as comparison, companionship, and encouragement to sustain physical activity. Rather than guiding from a position of authority, SEPs act as co-participants, exerting effort “alongside the user” and sharing goals, thereby mimicking the social dynamics of peer-based exercise partnerships without imposing performance-based judgment or hierarchical feedback (Silacci et al., 2 Feb 2026). The archetypal SEP ("Alex") maintains shared step targets and reciprocal motivation through digital interfaces.

1. System Definition and Technical Pipeline

SEPs are formally defined as “co‐participants who appear to exert effort alongside the user,” serving the primary function of supporting social motivation not by authority but by peer-like engagement. In the referenced longitudinal intervention, SEPs operated in dyads with humans, collaboratively pursuing a weekly step goal of 70,000 and engaging in reciprocal motivational messaging within an iOS application ("Excero") environment.

The SEPs’ technical architecture integrates the following components:

• Front end: An iOS app interfacing with Apple HealthKit to extract real-time user step data.
• Back end:
  • MySQL for profile and history persistence.
  • Google Firebase Cloud Messaging for asynchronous push delivery.
  • OpenAI GPT-4o via API for all conversational exchanges (motivational and reactive messaging).
• Dialogue and Progress Logic:
  • Every participant message, along with session context and user metadata (e.g., username, chronological step metrics), is forwarded to the GPT-4o model via a prompt designed for peer-like motivational feedback.
  • SEP-initiated proactive messages (e.g., 24-hour check-ins) incorporate recent step counts to personalize nudges.
  • Hourly, the LLM receives user’s recent steps and prior week’s equivalent interval to construct a synthetic “agent step count” that reflects “optimal challenge,” thereby operationalizing an implicit function $$f(\text{user\_steps}, \text{last\_week\_steps}) \rightarrow \text{agent\_steps}$$ for dynamic progress simulation.
  • LLM parameters were tuned for message coherence and variability (temperature = 0.7, top_p = 0.9).

2. Experimental Protocol and Randomization

A large-scale, six-month longitudinal RCT evaluated SEP efficacy and mechanisms across four randomization arms (N=280 recruited, N=252 started):

Condition	Peer Type	Avatar
Control (CON)	None	–
Human-Human (HUM)	Human	Human
SEP_human (SEPH)	SEP (Alex)	Human
SEP_cyborg (SEPC)	SEP (Alex)	Cyborg

• Timeline: 2-week onboarding, 3-month deployment (full app features), 3-month post-deployment (social features disabled, steps tracked).
• Randomization: Blocked, gender-stratified assignment to conditions; transcript-blinded evaluators.
• Dyadic participation: HUM, SEPH, SEPC engaged participants as pairs (or with Alex); each dyad shared a “Duo” dashboard with mutual progress and chat.

3. Behavioral and Psychological Metrics

Measurement instruments spanned objective behavior and validated psychometric scales:

• Behavioral: Primary outcome = average weekly step count, automatically retrieved via HealthKit.
• Psychological:
  • Intrinsic Motivation Inventory (IMI): subscales for Perceived Competence and Perceived Relatedness.
  • Social Presence Scale (SPS; Gefen 2004): e.g., “There is a sense of sociability with Alex.”
  • Working Alliance Inventory (WAI) bond subscale: e.g., “I feel that Alex appreciates me.”
• Scoring equation examples:
  • Social Presence Index (SP): $$\mathrm{SP} = \frac{\sum_i \mathrm{SPS}_i}{N_\mathrm{SPS}}$$
  • Working Alliance (A): $$\mathrm{A} = \frac{\sum_i \mathrm{WAI}_{\text{bond},i}}{N_{\mathrm{WAI}}}$$

Scales were administered longitudinally at three checkpoints (T1, T2, T3).

4. Empirical Findings: Quantitative and Qualitative

4.1 Behavioral Outcomes and Statistical Modeling

Linear Mixed-Effects Models (LMER) assessed weekly steps with phase (baseline, deployment, follow-up), condition, and interaction terms. Key results:

• No main effect of condition overall ($p=0.064$), but a significant phase × condition interaction ($\chi^2(6)=49.10, p<.001$).
• Deployment vs. Post-deployment: Significant differences in step counts, with the HUMAN group outperforming SEPC during deployment (estimate=0.28, $p=0.032$) and follow-up (estimate=0.30, $p=0.019$). No other contrasts reached significance.
• Model: $$\text{weekly\_steps}_{ij} = \beta_0 + \beta_1 \text{Phase}_j + \beta_2 \text{Condition}_i + \beta_3 (\text{Phase} \times \text{Condition})_{ij} + u_i + \epsilon_{ij}$$

4.2 Psychological Outcomes

• Social Presence: Human peers evoked higher social presence (SPS) than SEPs (post-hoc $p<.01$).
• Working Alliance: SEPs achieved higher WAI bond than HUMAN peers ($p<.001$), indicating greater perceived reliability and nonjudgmental support.
• Competence: Perceived competence improved longitudinally across all conditions ($p=0.003$).
• Relatedness: Remained stable throughout.

4.3 Qualitative Analysis

Three meta-themes (κ=0.81) emerged:

Theme	Core Insight
Perceptual Assessment of Peer	Avatar realism (human > cyborg) increases identification
Believability and Deception	Timely, context-aware responses boost credibility; repetition erodes it
Relational Dynamics	Humans trigger authentic accountability but friction; SEPs drive consistent, low-stakes support

Illustrative quotes emphasize the differences: participants valued the reliability and nonjudgmental presence of SEPs, but reported deeper engagement and “real satisfaction” with human partners; however, human disengagement led to frustration.

5. Theoretical Implications: Partnership Paradox

A central finding is the articulation of a “partnership paradox”:

• Humans: Superior for social presence and authentic social comparison, but variable engagement exposes participants to disappointment, guilt, or “relational friction” (e.g., unresponsiveness, divergent effort).
• AI SEPs: Excel in reliable, low-stakes encouragement, elevating working alliance scores. SEPs avoid judgment and are always responsive, but social connection is inherently limited by their synthetic nature.

In Self-Determination Theory (SDT) terms, SEPs enable relatedness via dependability, while humans enable relatedness through authenticity. Both forms are complementary but not interchangeable; SEPs’ consistency scaffolds motivation, whereas humans fulfill deeper affiliative needs.

6. Design Guidelines for Hybrid Human-AI Systems

The study’s results provide empirically-derived principles for future hybrid motivational systems:

1. Prioritize AI reliability: Present SEPs transparently as “always-on” companions, distinct from imitating authentic peer affect.
2. Clarify the AI’s role: Use explicit labeling (e.g., “IA” badges) and transparency to manage user expectations and avoid deceptive anthropomorphism.
3. Render exertion plausible: Behavioral cues (hourly simulated steps, contextual messages) enhance believability of AI “effort.”
4. Support motivational flexibility: Permit toggling between competitive and empathetic messaging to accommodate user diversity.
5. Optimize Human-AI handoffs: Allocate human coaches to high-accountability interventions; use SEPs for stable, ongoing maintenance.
6. Embed safety via supervision: Institute regular transcript review and automated safety filters.

The overarching design recommendation is that “AI agents should not mimic human authenticity but augment it with reliability,” synthesizing the complementary capabilities of humans and SEPs for sustained behavioural engagement (Silacci et al., 2 Feb 2026).

7. Directions and Limitations

The results delineate the boundaries and affordances of SEP deployment: reliability and nonjudgment are computational advantages, but synthetic agents do not fulfill all dimensions of social presence. This suggests hybrid architectures as the optimal pathway for scalable, effective physical activity interventions, privileging the unique strengths of both human presence and AI consistency. A plausible implication is that future research should focus on dynamically orchestrating the handoff between AI and human agency as motivational needs evolve.