What the Mount Sinai research establishes is not that AI triage tools are imperfect — every clinical tool is — but that their failure mode is inverted from what safety requires. A tool that under-triages emergencies while over-alerting on low-risk cases does not make conservative errors; it makes the specific errors most likely to produce delayed treatment and alarm fatigue simultaneously. Health systems citing benchmark accuracy figures in their rollout announcements are citing a number that does not survive contact with how patients actually describe symptoms. The institutions deploying these…