At the launch of GPT-5 on August 7, 2025, OpenAI chief executive Sam Altman brought a couple on stage. The wife, who is being treated for more than one type of cancer, described using ChatGPT to interpret her biopsy results and to weigh whether to undergo radiation therapy. The demonstration was meant to show that the new model can help people through the most delicate decisions of their lives. It also showed how far the company has moved from the heavily hedged medical disclaimers that once accompanied every health question put to a chatbot.
A warning published days earlier
Two days before the launch, a case report in Annals of Internal Medicine: Clinical Cases described a 60-year-old man in the United States who, after asking ChatGPT how to eliminate chloride from his diet, replaced table salt with sodium bromide bought online. Over three months he developed bromism, a toxic syndrome that was common a century ago when bromide sedatives were widely sold. He arrived at a hospital with paranoia, hallucinations and the belief that a neighbor was poisoning him, and was treated in a psychiatric unit for about three weeks before recovering, according to the report as covered by The Guardian and NBC News.
The physicians who wrote the case could not see the patient’s original chat logs. When they asked ChatGPT a similar question themselves, they reported that it mentioned bromide as a substitute without a clear health warning or any question about why the user wanted to know, the kind of context a clinician would ask for.
Better, but not infallible
OpenAI argues that GPT-5 is its most capable model for health questions, and it has built a benchmark to back the claim. HealthBench, released in May 2025, was developed with input from more than 260 physicians across 60 countries and scores models on thousands of realistic health conversations using physician-written criteria. OpenAI reports that GPT-5 outperforms its earlier models on the benchmark, with notably fewer errors on the hardest cases.
A higher score is not the same as reliability. Several AI researchers who tested GPT-5 after launch described it as an incremental improvement rather than a leap: the same categories of reasoning errors persist, delivered in a warmer and more confident tone. Altman has compared talking to GPT-5 to consulting a PhD-level expert. The bromism case is a reminder that an expert-sounding answer, taken without context, can still send someone to the hospital.
What this means for users
The lesson is not to avoid AI tools for health questions. Used to understand terminology, prepare questions for a doctor or make sense of a test report, they can be genuinely helpful, which is what the patient on OpenAI’s stage described. The lesson is that launch events and benchmark charts are marketing as well as evidence, and that a model which cannot ask why you want to remove salt from your diet is not a substitute for a clinician who can.
A second opinion from a human doctor remains essential for any decision about treatment, medication or diet. Until AI systems can reliably recognize when a question is dangerous and when a user needs more than an answer, that is unlikely to change.
