Can AR Embedded Visualizations Foster Appropriate Reliance on AI in Spatial Decision-Making? A Comparative Study of AR X-Ray vs. 2D Minimap

Paper Title

Can AR Embedded Visualizations Foster Appropriate Reliance on AI in Spatial Decision-Making? A Comparative Study of AR X-Ray vs. 2D Minimap

Publication Info

• Topic area: Human-AI collaboration in spatial decision-making using AR visualizations.
• Keywords: Augmented Reality, AI reliance, spatial decision-making, embedded visualizations, X-ray visualization, Minimap, cognitive biases, human-computer interaction, time-critical tasks, spatial cognition.

Background and Problem

• Problem / challenge: Traditional 2D visualizations for spatial decision-making impose cognitive load due to reference frame translation, leading to inappropriate reliance on AI suggestions. AR embedded visualizations promise to reduce this load but may introduce new challenges.
• Significance: Improving human-AI collaboration in spatial tasks is critical for applications like emergency response, indoor navigation, and crisis management, where poor decisions can have severe consequences.
• Motivation and related work: Prior research has focused on desktop-based human-AI collaboration and AR navigation but has not extensively studied AR embedded visualizations for spatial decision-making. This paper addresses the gap by exploring how AR visualizations affect AI reliance and decision-making quality.

Solution

• Proposed approach: Comparative study of AR X-ray (embedded visualization) and 2D Minimap in AI-assisted spatial decision-making tasks.
• Novelty:
  1. Empirical investigation of AR embedded visualizations in human-AI collaboration for spatial tasks.
  2. Identification of perceptual challenges and cognitive biases unique to AR X-ray.
  3. Analysis of decision accuracy, AI reliance, response time, spatial mapping, and confidence across visualization methods.
• Procedure and key techniques:
  • Conducted a 2 × 2 within-subjects study with 32 participants using Apple Vision Pro.
  • Compared Minimap and X-ray visualizations under AI-assisted and non-AI conditions in a time-pressured spatial target selection task.
  • Measured decision accuracy, AI reliance, response time, confidence, and pointing error rates.
  • Used simulated AI with 75% accuracy to provide decision suggestions.

Results

• Concrete findings:
  • Decision accuracy was higher with Minimap (median = 0.88) compared to X-ray (median = 0.70).
  • X-ray visualization improved spatial mapping, evidenced by lower pointing error rates (median = 0 vs. 0.08 for Minimap).
  • X-ray visualization led to greater inappropriate reliance on AI, primarily as over-reliance.
  • Response times were longer in X-ray conditions but shorter when initial search time was excluded.
  • No significant differences in self-reported confidence across conditions.
• Advantage over baselines:
  • Minimap outperformed X-ray in decision accuracy and appropriate AI reliance.
  • X-ray excelled in spatial mapping, reducing post-trial pointing errors.
• Experiments / evaluation:
  • 1024 trials conducted across 32 participants in a real two-floor building environment.
  • Controlled spatial arrangements and randomized trial orders ensured experimental rigor.
  • Quantitative measures complemented by qualitative thematic analysis of post-study interviews.
• Limitations and future work:
  • Fixed parameters (e.g., AI accuracy, spatial arrangements) may limit generalizability to real-world scenarios.
  • Sample size and participant demographics may not reflect specialized user groups.
  • Future studies should explore alternative visual encodings, task modalities, and AR-induced cognitive biases.

Summary

This study compared AR X-ray and 2D Minimap visualizations in AI-assisted spatial decision-making tasks. While the X-ray visualization enhanced spatial mapping, it also induced over-reliance on AI due to perceptual challenges and cognitive biases like visual proximity illusions. Minimap outperformed X-ray in decision accuracy and appropriate AI reliance. These findings highlight the need for careful integration of AI assistance in AR systems to mitigate biases and leverage AR’s strengths in spatial tasks. Future research should explore richer visual encodings, embodied tasks, and strategies to counteract AR-induced biases.