Harvard's latest research: Using AI large language models for medical diagnosis, 80% of initial diagnoses are incorrect

Mars Finance News, April 15 — A research team from Harvard Medical School evaluated the diagnostic capabilities of over 20 advanced AI large language models (LLMs), including ChatGPT, DeepSeek, Gemini, and Claude. The results showed that the error rate for “differential diagnosis” (identifying possible diseases) based on initial patient symptoms and signs was as high as 80%. The study further indicated that after patients provide more test results, large models can reduce the failure rate of “final diagnosis” to around 40%. The researchers stated that this means AI chatbots need comprehensive patient information to make more accurate diagnoses; when patients cannot provide complete health testing data, the results provided by AI are unreliable. The researchers also emphasized, “Artificial intelligence has not yet reached a level where it can make diagnostic decisions for patients without medical professional intervention.” (Yicai)