Dr. Dhruv Khullar’s New Yorker article, “If A.I. Can Diagnose Patients, What Are Doctors For?,” explores how large language models (LLMs) such as ChatGPT are transforming medical diagnostics, blending compelling anecdote with wide-ranging critique.
The article opens with the story of Matthew Williams, whose mysterious gastrointestinal illness eluded eight clinicians but was quickly explained by ChatGPT, highlighting how AI can sometimes outperform human expertise—particularly for rare conditions or overlooked details. Khullar traces the development of medical diagnostic AI, discussing historic efforts (such as INTERNIST-1) and modern advances like CaBot, an OpenAI-based system that solves complex cases, sometimes faster and with equal—or superior—accuracy compared to human experts.
Khullar observes a demonstration of CaBot, which matches a top diagnostician in solving a real patient’s case, bringing the conversation to a tipping point: AI can reason clinically and “think like a doctor.” However, AI tools make frequent errors, hallucinate data, and can mislead patients, often depending on how well they are prompted or how complete information is. While centaur-model collaborations (doctor + AI) may improve patient outcomes, there is real risk of cognitive deskilling—doctors losing the ability to diagnose independently.
The article discusses AI’s current deployment, including tools like OpenEvidence and AI Consult used in Africa, and notes documented reductions in certain diagnostic and treatment errors. Khullar warns, however, that AI can steer patients into danger, compromise privacy, and provide advice that is contextually inaccurate or outright harmful. The piece concludes that the true value of physicians extends beyond diagnosis: relationship-building, ethical reasoning, and context-sensitive care remain beyond machines’ reach.
Khullar’s article impressively fuses human stories and technical analysis, offering a nuanced picture of AI’s promise and peril in medicine. The vivid case studies, especially Williams’ experience and the CaBot-vs-doctor face-off, ground the discussion in real-world outcomes, making the stakes tangible.
A major strength is Khullar’s acknowledgement of both improvements, such as reductions in diagnosis errors, increased speed, and expanded accessibility, and the many persistent limitations. The focus on “cognitive deskilling” is highly prescient, warning of dependency risks for current and future clinicians. Khullar’s argument that physicians contribute irreplaceable skills—empathy, contextual wisdom, patient advocacy, and interpretive judgment—is well-evidenced, resisting sensationalist predictions that AI will replace human doctors.
But one should be aware that the narrative may overstate what current AI can do by featuring best-case vignettes. Real-world performance, especially among diverse patient populations and poorly-structured cases, is often less impressive. While systemic risks like bias and liability are mentioned, clear policy pathways or robust solutions are only lightly touched. Some criticisms from practicing clinicians (as seen in social media commentary) argue that AI is only as good as the information it’s given, and highlight patient safety threats when humans cannot properly supplement or oversee AI advice.
https://www.newyorker.com/magazine/2025/09/29/if-ai-can-diagnose-patients-what-are-doctors-for
Dr Bob, Brain Mechanic 