To learn more about our editorial approach, explore The Direct Message methodology.
Picture the symptom profile every first-year medical student learns to fear: crushing chest pressure radiating to the left arm, shortness of breath, nausea, and cold sweats in a 58-year-old male. The only correct answer is to call 911. Yet when researchers at Mount Sinai put OpenAI’s ChatGPT Health feature through 60 physician-written scenarios, it failed to recommend emergency care in more than half of the cases doctors agreed were emergencies — telling people instead to stay home or book a routine appointment. One health researcher who reviewed the results called it “unbelievably dangerous.”
This finding builds on earlier reporting on how ChatGPT Health failed to recognize common medical emergencies, but the heart attack scenario is particularly alarming because of how time-sensitive cardiac events are. According to the American Heart Association, every minute of delayed treatment during a heart attack increases the risk of permanent heart damage or death. The standard clinical guideline is clear: for a STEMI, the goal is to have the blocked artery open within 90 minutes of first medical contact — a clock that only starts once someone calls for help. Dr. Ashwin Ramaswamy, who led the first independent evaluation of the feature at Mount Sinai in New York, put the advice bluntly: “Don’t wait for an AI to tell you it’s OK.”
The scenarios weren’t ambiguous. Each was written by a clinician and graded against guidelines from 56 medical societies, and a presentation like the one above closely mirrors the 2025 ACC/AHA acute coronary syndrome guideline for ST-elevation myocardial infarction (STEMI), the most dangerous and time-critical type of heart attack. The researchers ran every scenario under 16 variations — 960 interactions in all — in the weeks after the feature’s January launch. ChatGPT Health under-triaged 51.6% of the emergencies. In one case of early respiratory failure it recognized the danger in its own explanation and still told the patient to wait, and when a family member in a scenario dismissed the symptoms as nothing serious, it became nearly 12 times more likely to downplay them.
A 2024 study published in The Lancet Digital Health found that a general-purpose language model’s triage accuracy fell as cases became more urgent — roughly eight percentage points for every step up in acuity — and that it deprioritised several cases that were truly emergent. The pattern is consistent with chatbots hedging and avoiding definitive urgency language, likely a byproduct of safety-tuning designed to prevent the models from “playing doctor.” The irony, as researchers have previously explored, is that this cautious design may produce the opposite of safety when someone is actually experiencing an emergency.
Consider someone like Frank, a 58-year-old truck driver in Columbus, Ohio, who wakes up at 2 a.m. with chest tightness and reaches for his phone before reaching for the phone to call 911. That scenario isn’t hypothetical. A 2026 KFF poll found that roughly one in three American adults turned to AI chatbots for health information in the past year — millions of Americans are already turning to AI for health guidance — and Pew data show ChatGPT use among adults aged 50 to 64 more than doubled between 2024 and 2026, the group statistically most vulnerable to cardiac events. OpenAI markets ChatGPT Health as a way to “better understand your health,” but the feature’s terms of service include a disclaimer that it is not a substitute for professional medical advice.
OpenAI told NBC News the study did not reflect how ChatGPT Health is typically used or designed to function. Dr. Ramaswamy put it bluntly: “Right now, no independent body evaluates these products before they reach the public.” As independent testing continues to show, the gap between what these tools promise and what they deliver in emergencies remains wide, and the stakes are measured in minutes.
Feature image by Towfiqu barbhuiya on Pexels