AI diagnoses; doctors better at management

1/ AI is getting better at one part of medicine: naming what might be wrong. Andrew Parsons, a University of Virginia physician and medical educator, wrote in The Conversation on June 1 that newer systems can now approach — and in some settings exceed — doctors on diagnostic tasks. But he drew a line between diagnosis and management, arguing that deciding what to do next is a different job. 2/ The sharpest recent data point Parsons cited was an April 2026 study in which OpenAI’s o1 posted 78% accuracy on complex diagnostic cases and outperformed experienced doctors on emergency-room cases. (medicalxpress.com) That helps explain why symptom checkers and chatbots can feel impressively capable. They are increasingly good at pattern recognition. 3/ But a correct diagnosis is not the same thing as a safe plan. Parsons’ point is that management requires prioritizing among risks, contraindications, comorbidities, and patient-specific trade-offs. (medicalxpress.com) A patient can have the “right” label and still need very different care depending on age, pregnancy status, kidney function, other drugs, or how sick they are right now. 4/ That distinction matters because diagnosis is mostly about categorizing; management is about choosing. (medicalxpress.com) Parsons described how experienced doctors use “illness scripts” built over years of practice: mental models of what a disease usually looks like, who gets it, how it progresses, and which details do not fit. Those scripts help clinicians catch the extra facts that change treatment, not just the facts that support a label. 5/ The emergency-department comparison that drove headlines in May also came with caveats. (medicalxpress.com) Jesse Pines wrote in Forbes on May 22 that the widely shared study was more nuanced than “AI beat doctors.” He noted the case set involved 76 real patients from Beth Israel Deaconess, and that the physicians used for comparison were internal medicine attendings, not emergency physicians. He also wrote that emergency medicine often prioritizes stabilization and ruling out danger, not simply landing on the single most precise diagnosis. 6/ That is exactly where human management tends to hold up better. A model may identify pneumonia, heart failure, or an ear infection. A clinician still has to decide whether the patient needs admission, imaging, antibiotics, watchful waiting, a specialist referral, or none of the above. The harder the case gets — multiple illnesses, medication interactions, unclear follow-up, social constraints — the less useful diagnosis alone becomes. 7/ The research frontier is moving in that direction. (forbes.com) Nature reported in July 2026 on conversational AI for disease management, not just one-off diagnosis, reflecting how developers are trying to build systems that reason across guidelines, patient history, and multiple visits. That work suggests the next contest is not whether AI can name a disease, but whether it can safely help manage one over time. 8/ For now, the evidence supports a narrower claim than some headlines imply. AI is becoming a stronger diagnostic assistant. (medicalxpress.com) The doctor’s comparative edge remains in deciding what to do with the diagnosis once patient-specific reality enters the picture. (nature.com)