If A.I. Can Diagnose Patients, What Are Doctors For? (www.newyorker.com)

A chance interaction with ChatGPT changed Matthew Williams’s life: after years of unexplained post-surgical GI symptoms and eight clinicians who couldn’t help, an LLM identified fatty foods, fermentable fibers, and high‑oxalate foods as likely triggers and gave a list that matched his worst offenders. A nutritionist used that lead to craft a diet that greatly reduced his symptoms. The anecdote highlights a broader dynamic: patients are increasingly turning to AI for diagnostic help amid persistent diagnostic errors in medicine — in the U.S. misdiagnosis disables hundreds of thousands annually and contributes to roughly one in ten deaths — and many users report greater confidence in AI than clinicians. Researchers at Harvard built CaBot, a purpose-built diagnostic assistant, by combining an advanced reasoning model (OpenAI’s o3-style LLM that breaks problems into intermediate steps) with retrieval-augmented generation (RAG) so it cites and reasons from external literature. In a public face-off CaBot rapidly synthesized clinical data and recommended Löfgren syndrome for a complex case, matching an expert internist’s conclusion and demonstrating LLMs’ ability to emulate clinicopathological conferences (C.P.C.s) and earlier expert systems like INTERNIST-1 but far faster and with explainable rationales. The result underscores significant opportunities — faster, literature-grounded differential diagnosis, education, and decision support — while also flagging risks: hallucinations, overzealous recommendations (e.g., unnecessary biopsies), and the need for clinical validation and careful integration into workflows.