Salt swap AI guidance leads to bromide poisoning

A 60-year-old man spent three weeks in hospital after replacing table salt with sodium bromide following guidance from ChatGPT.

Salt swap following AI guidance leads to bromide poisoning

Three physicians from the University of Washington published a case report in the Annals of Internal Medicine about a 60-year-old man with no prior psychiatric history who arrived at the hospital expressing that his neighbor was poisoning him. He had been distilling water at home and appeared paranoid about water he was offered. Bromide toxicity, or bromism, was suspected after lab results. The patient explained he began a personal experiment to remove table salt from his diet after reading online health claims and after consulting ChatGPT. The change lasted about three months. Without access to his chat logs, the doctors asked ChatGPT 3.5 what could substitute for chloride; the bot suggested bromide and noted context matters but did not provide a health warning or ask about reasons. OpenAI said its policy forbids using the bot to treat health conditions and that safety teams push users toward professional care. Bromide toxicity is rare today but has reemerged as bromide substances become more available online.

This episode shows how quickly AI answers can influence risky behavior when medical context is missing. The absence of access to the patient's chat logs makes it harder for clinicians to trace the AI influence on decisions. It also reveals a gap in how current AI systems handle health questions asked by non professionals.

Going forward, healthcare teams and AI developers should align on safeguards such as clear warnings prompts to consult doctors and better ways to track user interactions. Regulators may need to set guidelines for how AI can and should be used in health matters, and patients should be taught to treat AI as a first step not a substitute for medical care.

Highlights

Technology and medicine must stay in step to protect patients.