The structural problem this study exposes is not that ChatGPT Health performed poorly on a benchmark — it is that OpenAI launched a consumer-facing triage product in January 2026 and the first independent safety data arrived in Nature Medicine the following month, after the tool had already reached a mass user base. The researchers at Mount Sinai did not find an edge case; they found a systematic pattern where the tool failed in both directions across a structured test of triage recommendations. A product that errs toward under-triage on emergencies…