AI-Powered Search & Rescue Is Faster — So Why Aren't More Lives Saved?

In a simulated search and rescue environment, researchers Elahe Oveisi and Hemanth Manjunatha compared human performance under two LLM-guided conditions against a no-LLM baseline. Using eye-tracking and behavioral metrics, they found that LLM guidance enhanced task efficiency—participants achieved higher rewards and victims-per-step—but crucially did not increase the total number of victims saved. The eye-tracking data revealed an attention-guidance trade-off: visual resources shifted toward the chat interface, accompanied by increased pupil size variability, indicating cognitive load.

Expertise played a moderating role: novices tended toward passive AI reliance, often following LLM suggestions without cross-checking. Experts, by contrast, maintained a "verification loop"—they persistently scanned the environment to cross-reference AI advice with ground truth. The findings suggest that LLM-mediated teaming efficacy depends on the operator's ability to maintain situational awareness by balancing AI guidance with direct observation. This has implications for designing AI assistants for high-stakes domains like emergency response.