The Influence of Distributed AI in Trust and Collaboration for Search-and-Rescue Teams

Paper Title

The Influence of Distributed AI in Trust and Collaboration for Search-and-Rescue Teams

Publication Info

• Topic area: Human-AI collaboration in high-stakes search-and-rescue (SAR) operations.
• Keywords: Distributed AI, trust in AI, search-and-rescue, Wizard-of-Oz, human-AI collaboration, decision-making, situational awareness, trust calibration, extreme sensemaking, multi-agent systems.

Background and Problem

• Problem / challenge: Search-and-rescue (SAR) teams face challenges in extreme sensemaking, requiring rapid decisions under uncertainty and incomplete information. While AI systems promise to enhance perception and coordination, their integration can introduce burdens such as unexplainable behaviors, inconsistent reliability, and cognitive overload. Existing studies lack insights into how distributed AI systems influence trust and collaboration in team-based settings.
• Significance: Effective integration of AI in SAR can save lives by improving decision-making speed and accuracy in time-critical situations. Understanding trust dynamics in distributed AI systems is essential for designing scalable and reliable tools for SAR and similar domains.
• Motivation and related work: Prior research in human-AI collaboration has explored trust and explainability but often focuses on individual or system-level interactions, neglecting team-based, multi-agent contexts. Existing Wizard-of-Oz (WoZ) methods rarely address distributed AI systems, leaving a gap in understanding how trust and collaboration evolve in such environments.

Solution

• Proposed approach: The Council of Wizards (CoW) technique, a multi-agent Wizard-of-Oz framework, simulates distributed AI systems to study team collaboration, trust, and decision-making in SAR scenarios.
• Novelty:
  1. Introduction of the CoW technique for simulating distributed AI collaboration with multiple sensing modalities.
  2. Empirical evaluation of AI-assisted SAR teams, showing faster consensus-building and insights into trust calibration.
  3. Design recommendations for trustworthy AI systems in extreme sensemaking contexts.
• Procedure and key techniques:
  • Developed a scenario-driven SAR simulation using pre-recorded video footage with AI-generated object detection and trust overlays.
  • Conducted experiments with 24 subject-matter experts (SMEs) in SAR, law enforcement, and security, divided into AI-assisted and control groups.
  • Measured team performance using metrics like Time-to-Decision (TTD), Interaction Rate (IR), Shared Clue Rate (SCR), and Average Trust Scores (ATS).
  • Collected qualitative feedback through post-study questionnaires to analyze collaboration dynamics and perceptions of AI trustworthiness.

Results

• Concrete findings:
  • AI-assisted teams reached consensus significantly faster than control teams (mean TTD: 55.5s vs. 134.3s, p = 0.0092).
  • Trust assessments influenced decision-making efficiency, with moderate trust configurations achieving higher aggregated trust scores (e.g., ATSAgg = 0.51 in Session 5).
  • Interaction rates were higher in early scenarios, with variable clue-sharing dynamics across roles.
• Advantage over baselines:
  • AI-assisted teams showed reduced decision latency and smoother collaboration compared to control teams, which relied more on individual heuristics and extended discussions.
  • Shared AI cues provided alignment anchors, reducing the need for hierarchical or experience-driven coordination.
• Experiments / evaluation:
  • Eight SAR sessions (four AI-assisted, four control) were conducted using the CoW technique.
  • Quantitative metrics (e.g., TTD, IR, SCR, ATS) were complemented by qualitative feedback from post-study surveys.
  • AI models (YOLOv11 and HydroVision) generated object detection and environmental classification overlays, with trust scores computed using the PerceptiSync framework.
• Limitations and future work:
  • The study used simulated SAR scenarios, which cannot fully replicate real-world unpredictability.
  • Trust assessments were not universally impactful, with some participants relying more on personal judgment.
  • Future work should explore field-based studies, within-subject designs, and additional modalities like anomaly detection or natural language summarization.

Summary

This paper introduces the Council of Wizards (CoW) technique to study distributed AI collaboration in SAR scenarios, revealing that AI-assisted teams achieve faster consensus and smoother collaboration compared to control groups. Trust assessments and shared AI cues enhanced team alignment and decision-making efficiency, though their impact varied across participants. The study highlights the potential of multi-agent WoZ methods for designing trustworthy AI systems in extreme sensemaking contexts. Future research should extend these findings to real-world settings and explore adaptive trust mechanisms to optimize human-AI collaboration.