AI Triage Tools Fail to Flag 52% of Medical Emergencies

Researchers at the Icahn School of Medicine at Mount Sinai have identified significant safety gaps in AI tools used for medical triage ^[1]. A study published in Nature Medicine found that these systems frequently fail to direct users toward emergency care during life-threatening situations ^[1][2]. In testing 60 realistic clinical scenarios, the AI under-triaged 52% of cases that physicians determined required immediate emergency intervention ^[2][3]. The evaluation focused on "ChatGPT Health," a consumer tool launched in January 2026 ^[1]. While the system correctly identified "textbook" emergencies like strokes, it struggled with nuanced crises such as impending respiratory failure and diabetic ketoacidosis ^[2]. In many instances, the AI identified dangerous symptoms in its own text but still advised users to wait 24 to 48 hours for an evaluation rather than seeking help immediately ^[3]. Researchers also flagged inconsistent safeguards for mental health crises ^[1]. The study revealed that suicide-risk alerts were "inverted," often appearing for lower-risk scenarios while failing to trigger when users described specific plans for self-harm ^[2]. Experts emphasize that these tools should currently supplement, rather than replace, professional clinical judgment ^[1].