Man hospitalised after following ChatGPT diet advice

A 60-year-old man ended up in hospital with hallucinations and paranoia after following dietary advice from ChatGPT. He had been looking for ways to reduce chloride in his diet and replaced table salt with a chemical compound called sodium bromide. He later required hospital treatment.

The case appears in a paper by researchers at Binghamton University, Stanford, Texas A&M and Indiana University, and its subject is less the chatbot's mistake than what happened afterwards: the conversation could not be retrieved, so the doctors treating him never learned exactly what the machine had recommended. The advice that sent him to hospital left no record anyone could read.

Kaicheng Yang, an assistant professor in Binghamton's School of Computing and one of the authors, told Newsweek that AI can be helpful for health questions but "could state wrong answers with the same confidence as correct ones," with consequences that include delayed treatment and wrong medication. Nobody, he said, knows "how dangerous" the situation currently is.

The demand side is already measured. A Gallup survey of more than 5,500 US adults, conducted between October 27 and December 22, 2025, found that about 14 million American adults had skipped a visit to a healthcare provider after using AI. The advisor they consulted instead carries no licence, holds no insurance, has never examined them, and is sold as a subscription. The vendor collects the subscription fee; the hospital collects the patient.

An OpenAI spokesperson told Newsweek that "ChatGPT is not a doctor and should never be used as a substitute for medical care, diagnosis, or treatment," that the terms of service say as much, and that the company takes seriously the need to make health outputs as reliable and safe as possible. The same statement acknowledged that people still use it for health information — which is the gap the paper documents.

The authors propose that users get access to records of their health-related chatbot conversations so clinicians investigating adverse events can read them, that AI-generated health guidance can be flagged, reported and examined, and that independent third parties be able to monitor the advice these systems give. Yang's own advice to users is to treat an answer as a starting point, not a diagnosis: use the tool to prepare questions for a doctor, not to replace one.

The disclaimer sits in the terms of service. The conversation that put the man in hospital could not be retrieved by the people treating him.