Bromism Case Shows the Risk of Acting on ChatGPT Medical Advice

A 60-year-old man in Seattle developed paranoia and auditory and visual hallucinations after using sodium bromide as a substitute for table salt. He told doctors that he had consulted ChatGPT before making the change and had used bromide salts for three months. The case is a sharp warning for AI builders: a plausible-sounding health answer can become dangerous when a user treats it as an actionable prescription. The reported case involved bromide poisoning, or bromism, after an AI-suggested salt substitution.

A salt substitution created a difficult diagnostic problem

The patient's initial blood tests appeared to show high chloride, a result that can suggest hyperchloremia. But other electrolyte measurements did not fit a straightforward chloride excess. Doctors suspected pseudohyperchloremia, in which a laboratory chloride test responds to another negatively charged ion and makes the chloride level appear higher than it really is.

Aspirin-related salicylate interference is one possible explanation, but the patient did not have high salicylate levels. Testing instead found bromide concentrations about 230 times above normal. That result led clinicians to diagnose bromism, a rare form of poisoning that can produce neurological and psychiatric symptoms. Bromide interfered with the chloride measurement and helped obscure the underlying poisoning.

Bromide is chemically similar to chloride and can move through chloride channels involved in neuronal activity. That similarity helps explain why bromide exposure can disrupt neurological function, although this case alone does not establish how frequently such events occur.

The practical lesson for medical AI systems

For product teams building health assistants, the failure is not simply that a model produced an incorrect fact. The more important failure mode is actionability. The answer connected a dietary goal with a specific chemical replacement, without adequately accounting for toxicity, dose, user context, or the need for professional supervision.

A safer system should treat requests involving supplements, chemical substitutions, medication changes, or electrolyte manipulation as high risk. It should avoid presenting a nonstandard substitute as a routine option, ask why the user wants the change, and direct the person to a clinician or pharmacist before they act. A warning added after a confident recommendation may not be enough.

The case also shows why clinical AI evaluation must include downstream workflow risks. A model can sound informative while creating a new diagnostic confounder. If a patient later presents with psychosis and an abnormal chloride result, the relevant history may include online supplements, unusual dietary experiments, and chatbot advice. The report's authors argued that bromism should be considered in similar presentations, particularly when the laboratory pattern is inconsistent.

What builders should not infer from one case

This is a single case report, not evidence that chatbots commonly cause bromide poisoning or that AI health tools are broadly unsafe. It does show that a rare hazard can become more reachable when an assistant gives specific advice without a reliable safety boundary.

The patient improved after about three weeks of hospital care, electrolyte replenishment, and risperidone, an antipsychotic medication. He was later discharged and remained stable at a follow-up two weeks later, but the available account provides limited information about long-term prognosis. The report describes the hospitalization, treatment, and short-term follow-up.

For builders, the decision rule is straightforward: health systems should optimize for safe next actions, not merely fluent answers. When the cost of a wrong recommendation includes poisoning or delayed diagnosis, refusal, escalation, and verification are product features, not friction.