The promise of AI-powered health assistance faces a critical reality check as artificial intelligence proves unreliable precisely when medical decisions matter most. Healthcare systems increasingly turning to chatbot triage may be creating dangerous blind spots at the extremes of medical urgency. A comprehensive evaluation of ChatGPT Health's diagnostic capabilities reveals a troubling pattern: while the system demonstrates competence with routine medical scenarios, it systematically fails at the clinical decision points where human lives hang in the balance. The AI frequently minimizes genuine emergencies that require immediate intervention while simultaneously escalating minor ailments into unnecessary urgent care visits. This creates a dual hazard—patients with serious conditions may delay critical treatment, while healthcare resources become strained by false alarms. The study exposes fundamental limitations in current AI training approaches, which appear optimized for common presentations rather than edge cases that define medical expertise. Emergency medicine specialists have long recognized that clinical acumen is measured not by routine diagnoses, but by the ability to distinguish the rare heart attack from typical chest pain, or the serious headache from a minor one. These findings arrive as healthcare systems worldwide grapple with physician shortages and seek technological solutions for patient triage. However, the research suggests that deploying AI for initial medical screening may introduce systematic biases that could prove more dangerous than helpful. The implications extend beyond individual patient safety to healthcare system integrity, as unreliable AI triage could either flood emergency departments with unnecessary cases or, more alarmingly, allow serious conditions to progress untreated. This represents a sobering reminder that healthcare AI, despite impressive capabilities in controlled settings, may not yet be ready for the life-and-death decisions that define medical practice.