On July 13, 2025, Mr. Winters asked ChatGPT about pain in his groin, which the chatbot assured him was “very likely another minor piece of the long story.” Hours later, he was admitted to the intensive care unit because of a massive pulmonary embolism, according to the lawsuit. The doctors said the dizzy spells may have been caused by a series of smaller pulmonary embolisms, and that the lung clots were likely brought on by the weeks he had spent sitting, according to the lawsuit.

In the weeks afterward, Mr. Winters needed help standing, eating, dressing and using the bathroom, according to the lawsuit, which added that he was facing “years of intensive physical and psychological recovery.”

Hundreds of millions of people ask ChatGPT Health and wellness questions every week, according to OpenAI. A 2025 analysis of searches on Microsoft’s Copilot found that health was “consistently the most common topic” users asked about on mobile.

Yet recent research has questioned the safety of chatbots for health advice. The first randomized study evaluating health advice from general chatbots like ChatGPT found that none of the models studied was “ready for deployment in direct patient care.” Even when scientists stress tested ChatGPT Health, a service that was tailor made for health questions, they found it missed emergencies and inconsistently activated guardrails.

Several aspects of Mr. Winters’ story made it “a nightmare scenario” for safety researchers, said Dr. Adam Rodman, a medical A.I. researcher at Beth Israel Deaconess Medical Center in Boston. For one, he said, Mr. Winters was using GPT-4o, an older model that was notoriously eager to please. It was implicated in several high-profile controversies, including the death of Adam Raine, a teenager who discussed plans to end his own life with ChatGPT.

And for technical reasons, the fact that Mr. Winters had long conversations with the chatbot made it more difficult for the safety features to work. Dr. Rodman said newer models are much better at detecting and responding to concerning symptoms, which, he said, he believed would make stories like Mr. Winter’s less likely. But he cautioned that there may be other ways the chatbots fail at providing health advice that “we haven’t realized yet.”