Abstract
Timely recognition and initiation of basic life support (BLS) before emergency medical services arrive significantly improve survival rates and neurological outcomes. In an era where health information-seeking behaviors have shifted toward online sources, chatbots powered by generative artificial intelligence (AI) are emerging as potential tools for providing immediate health-related guidance. This study investigates the reliability of AI chatbots, specifically GPT-3.5, GPT-4, Bard, and Bing, in responding to BLS scenarios. A cross-sectional study was conducted using six scenarios adapted from the BLS. Objective Structured Clinical Examination (OSCE) by United Medical Education. These scenarios covering adult, pediatric, and infant emergencies, were presented to each chatbot on two occasions, one week apart. Responses were evaluated by a board-certified emergency medicine professor from Tehran University of Medical Sciences, using a checklist based on BLS-OSCE standards. Correctness was assessed, and reliability was measured using Cohen's kappa coefficient. GPT-4 demonstrated the highest correctness in adult scenarios (85% correct responses), while Bard showed 60% correctness. GPT-3.5 and Bing performed poorly across all scenarios. Bard achieved a correctness rate of 52.17% in pediatric scenarios, but all chatbots scored below 44% in infant scenarios. Cohen's kappa indicated substantial reliability for GPT-4 (k = 0.649) and GPT-3.5 (k = 0.645), moderate reliability for Bing (k = 0.503), and fair reliability for Bard (k = 0.357). While GPT-4 showed the highest correctness and reliability in adult BLS situations, all tested chatbots struggled significantly in pediatric and infant cases. Furthermore, none of the chatbots consistently adhered to BLS guidelines, raising concerns about their potential use in real-life emergencies. Based on these findings, AI chatbots in their current form can only be relied upon to guide bystanders through life-saving procedures with human supervision.