Doctors are missing two-thirds of AI errors in patient messages. That's not a guess. A 2024 Lancet Digital Health study tested 20 primary care physicians reviewing AI-drafted patient portal responses. The AI produced drafts that were empathetic, well-written, and sometimes dangerously wrong. Here's what caught our attention: → About 7% of unedited AI drafts carried a risk of severe patient harm → One draft carried a risk of death → Physicians missed an average of 2.67 out of 4 intentionally flawed messages → Between 35% and 45% of those flawed drafts were sent to patients completely unedited Why? Because the writing sounded so good that doctors trusted it. Ninety percent reported confidence in the tool's performance, even while critical errors slipped past them. Our team calls this the "eloquent hallucination" problem. The better AI writes, the harder it becomes for a busy clinician to spot when it's confidently wrong. The fix isn't removing AI from healthcare. Physicians are drowning in admin work. Some spend 10 hours a month just answering patient messages. They need help. But help has to be built differently. Systems grounded in verified medical knowledge, not just next-word prediction. That's what our latest research digs into. Honest question for this community: if you received a message from your doctor's office, would you want to know whether AI helped write it? #HealthcareAI #PatientSafety #ClinicalAI