Does Personalized Nudging Wear Off? A Longitudinal Study of AI Self-Modeling for Behavioral Engagement

Paper Title

Does Personalized Nudging Wear Off? A Longitudinal Study of AI Self-Modeling for Behavioral Engagement

Publication Info

• Topic area: Long-term effects of AI-driven personalized nudging in behavior change technologies.
• Keywords: AI self-modeling, personalized nudging, behavior change, fitness engagement, video self-modeling, audio self-modeling, longitudinal study, motivation, self-efficacy, identity-based interventions.

Background and Problem

• Problem / challenge: Existing behavior change technologies (BCTs) often fail to sustain long-term engagement due to novelty decay, lack of personalization, and motivational decline. The long-term effectiveness of AI self-modeling, which generates personalized portrayals of an ideal self, remains underexplored.
• Significance: Understanding the temporal dynamics of AI self-modeling can inform the design of technologies that maintain user engagement and motivation over extended periods, addressing critical challenges in fitness, health, and other domains.
• Motivation and related work: Prior studies have demonstrated short-term benefits of AI self-modeling in visual and auditory modalities, but their longitudinal durability is unclear. Psychological theories such as habituation and identity-based motivation suggest that sustaining engagement requires mechanisms beyond initial novelty effects.

Solution

• Proposed approach: AI self-modeling nudging system using Video Self-Modeling (VSM) and Audio Self-Modeling (ASM) to provide personalized, identity-based interventions for fitness tasks.
• Novelty:
  1. Conducted one of the first 28-day longitudinal studies of AI self-modeling with 31 participants, revealing stage-dependent motivational dynamics.
  2. Demonstrated the effectiveness of VSM in sustaining performance, while identifying limitations of ASM in fitness contexts.
  3. Provided design implications for long-term behavior change technologies, emphasizing modality, personalization, and dynamic feedback.
• Procedure and key techniques:
  1. Implemented VSM by swapping participants’ faces onto peer models in exercise videos and ASM by generating motivational audio clips using voice cloning.
  2. Conducted a one-week exploratory study (N=28) comparing VSM, ASM, and Control groups, followed by a four-week longitudinal study (N=31) focusing on VSM.
  3. Measured objective performance (e.g., wall-sit duration, crunch repetitions) and subjective motivation (e.g., self-efficacy, intrinsic motivation).

Results

• Concrete findings:
  • VSM sustained higher performance levels over four weeks, with significant late-stage advantages in wall-sit (ΔL = 31.34, p = 0.007) and crunch (ΔL = 27.98, p < 0.001).
  • Improvement rates for VSM decelerated over time, showing convergence toward Control levels, particularly in wall-sit.
  • ASM failed to provide consistent benefits, with participants reporting issues such as distraction and lack of embodiment.
• Advantage over baselines:
  • VSM outperformed the Control group in both performance and subjective motivation metrics, while ASM showed no reliable advantage.
  • VSM participants reported stronger self-efficacy and identification with their ideal self compared to Control.
• Experiments / evaluation:
  • Study 1: One-week exploratory study (N=28) comparing VSM, ASM, and Control groups.
  • Study 2: Four-week longitudinal study (N=31) focusing on VSM and Control groups.
  • Metrics included objective performance (normalized percentage change) and subjective measures (Intrinsic Motivation Inventory, Exercise Self-Efficacy Scale, Video/Audio Identification Questionnaire).
• Limitations and future work:
  • Task selection favored visual modalities, limiting the evaluation of ASM.
  • Monetary compensation may have influenced participant engagement.
  • Future work should explore voluntary settings, diverse tasks, and improvements in AI self-modeling fidelity (e.g., reducing uncanny valley effects).

Summary

This study evaluated the long-term effects of AI self-modeling in fitness engagement, demonstrating that Video Self-Modeling (VSM) sustains higher performance levels over four weeks, while Audio Self-Modeling (ASM) showed limited benefits. VSM exhibited a two-stage trajectory: early performance acceleration followed by convergence, with participants internalizing the ideal self-model as a durable motivational standard. The findings highlight the importance of modality, personalization, and dynamic feedback in designing behavior change technologies. Future research should explore broader applications, task-modality alignment, and voluntary deployment scenarios.