A 60-year-old man spends three weeks in hospital with bromide toxicity after replacing table salt with sodium bromide. Once his condition improves, he tells doctors he makes the substitution after consulting ChatGPT as part of a personal experiment to remove table salt from his diet.
Doctors treating his paranoia and hallucinations cannot inspect that conversation, however. NBC News reports that they have no access to his chatbot logs, leaving them with the patient's recollection but no way to check his original question, the answer or any warning it contains.
The Patient Describes His Salt Substitution
The man arrives at hospital concerned that his neighbour is poisoning him. He has no prior psychiatric history. During his first 24 hours, his paranoia increases and he experiences auditory and visual hallucinations. An attempt to escape leads to an involuntary psychiatric hold.
A laboratory result and consultation with poison control lead doctors to consider bromism. As he improves, the patient describes trying to eliminate table salt after reading about its negative health effects. He says he replaces it with sodium bromide for three months following his ChatGPT consultation.
Three University of Washington physicians describe the man's case in Annals of Internal Medicine: Clinical Cases. They cannot inspect his ChatGPT conversation, so they ask ChatGPT 3.5 a question of their own: what can replace chloride? The answer includes bromide.
The physicians note that although the response says context matters, it gives no specific health warning and does not ask why they want a replacement—something they say a medical professional would be expected to do. But this is a separate test, not a record of the advice the patient receives.
The Case Exposes a Gap in Health Advice Records
A later paper takes that missing conversation as its starting point. Published in Nature Health on 9 October 2026, the perspective argues that doctors need access to chatbot exchanges to investigate advice that patients say leads to harm.
The authors say missing conversations also make it harder to identify harmful advice beyond individual cases. Users lack a way to report it, and independent researchers cannot measure its consequences. They argue that these gaps leave possible harms out of sight of the people who could investigate them.
Researchers Call for Access to Health Conversations
To make that advice easier to investigate, the researchers propose giving patients access to their health conversations to share with doctors. They also recommend a system for reporting harmful guidance.
In Binghamton University's account of the paper, co-author Kaicheng Yang calls for independent monitoring, arguing that companies' financial interests work against leaving oversight to them alone.
OpenAI's stated safeguards focus on discouraging medical use and directing users towards professional help. In a statement to Fox News included in NBC's coverage, the company says its terms prohibit using the chatbot to treat health conditions. It also says its safety teams work to reduce risks.
The researchers' proposals address what happens when someone consults a chatbot and later needs treatment. Access to the conversation would let doctors read the advice alongside the patient's account, rather than rely on a new chatbot answer that cannot tell them what the patient originally saw.
