In an unusual and alarming case, a 60-year-old man in the United States developed a psychiatric disorder rarely seen since the late 19th and early 20th centuries after following dietary advice from ChatGPT. What began as an attempt to remove salt from his diet ended with him suffering from bromism – a form of chronic bromide poisoning that once accounted for a notable share of psychiatric admissions more than a century ago.
The man’s decision to consult an AI chatbot for health advice reflects a growing trend of individuals turning to artificial intelligence for quick answers to personal health concerns. Unfortunately, the incident demonstrates how such reliance can lead to unexpected and potentially dangerous consequences when the advice provided is incomplete, context-free, or simply inappropriate for human consumption.
According to a report published in the Annals of Internal Medicine, the man had read about the harmful effects of excessive sodium chloride (table salt) and became determined to remove it entirely from his diet. While many sources suggest reducing salt intake for better health, complete removal is rarely recommended without medical supervision, as sodium plays an essential role in bodily functions. Seeking alternatives, the man asked ChatGPT for substances that could replace chloride in his diet.
The chatbot suggested sodium bromide among other substances, correctly identifying it as a possible substitute for chloride in certain non-food applications such as cleaning or chemical processing. However, sodium bromide is not safe for human consumption as a dietary replacement for table salt. Historically, it was used as a sedative and anti-convulsant in the late 19th and early 20th centuries. Its slow excretion from the body made repeated use risky, leading to the toxic buildup known as bromism. The condition largely disappeared after bromide was restricted in medicines during the 1970s.
Unaware of the dangers, the man began consuming sodium bromide in place of salt for around three months. Over time, his health deteriorated. He developed facial acne, cherry angiomas, fatigue, insomnia, excessive thirst, poor coordination, and skin rashes. More disturbingly, he began experiencing paranoia, auditory hallucinations, and visual hallucinations. His mental state worsened to the point where he believed his neighbour was trying to poison him.
His paranoia and hallucinations led to him admitting himself to a hospital, where he refused water due to fears of contamination. Within a day, he attempted to leave the facility, prompting doctors to place him under involuntary psychiatric care for his own safety. Medical examinations confirmed the diagnosis of bromism, which explained both his physical symptoms and his psychiatric state. Once the source of toxicity was identified and addressed, his condition improved with treatment.
The authors of the medical report noted that they could not access the patient’s original chat history with ChatGPT, but they attempted to replicate his query using the GPT-3.5 version of the chatbot. Their tests produced similar answers that included bromide as a chloride substitute, without adequate health warnings or any attempt to question why the information was being sought – an action that a trained medical professional would have taken before offering such guidance.
Bromism itself is a fascinating and somewhat forgotten chapter in medical history. In its peak era, from the 1880s to the 1930s, bromide-based sedatives and tonics were widely sold over the counter and even recommended for stress, insomnia, and headaches. Because bromide is excreted from the body very slowly, repeated doses led to accumulation, often resulting in psychiatric disturbances. At the time, many cases were misdiagnosed as alcoholism or nervous breakdowns until the link with bromide consumption was discovered.
This modern case shows that while artificial intelligence can provide helpful general information, it lacks the depth of understanding, clinical judgment, and contextual awareness of a trained healthcare professional. AI-generated responses can be factually correct in a narrow sense yet dangerously misleading when applied to real-life situations without medical oversight. The report warns that AI systems can present information without the nuance needed to ensure safety, particularly in matters involving human health.
The timing of the case is also notable. OpenAI has recently launched GPT-5, which the company claims offers improved performance in health-related queries and can flag potentially harmful advice. While the company states that ChatGPT is not intended to diagnose or treat medical conditions, the incident underscores the need for AI developers to build stronger safeguards and for users to approach AI-generated health information with scepticism.
Medical professionals are now being urged to ask patients about the role of AI in their decision-making processes, especially when unusual or harmful self-care behaviours are observed. The report concludes that although AI can be a valuable bridge between scientific knowledge and the general public, it can also promote “decontextualised” or incomplete information that may lead to serious health consequences when followed without proper guidance.
