A man walks into an emergency room convinced his neighbour is poisoning him. He hasn't slept properly in days. He's thirsty, but he won't drink the water they hand him. Within 24 hours, he's hallucinating, both sight and sound, and doctors have no choice but to place him on an involuntary psychiatric hold.
He has no history of mental illness. No prior hospitalisations. Nothing in his chart that would explain what's happening to him.
The answer, when doctors finally pieced it together, wasn't a virus, a stroke, or a psychiatric break. It was salt. Or more precisely, what he'd been using instead of salt, after a conversation with ChatGPT that nobody in that hospital room had ever heard of going wrong before.
This isn't a hypothetical. It's a real, peer-reviewed case published in a respected American medical journal, and it's become one of the most talked-about cautionary tales in the entire debate over AI in healthcare. It also lands at an interesting moment. Doctors and AI researchers are currently locked in a very public fight over whether AI is already outperforming physicians, and stories like this one are exactly what critics point to when they say the technology isn't ready to be trusted alone.
If you haven't seen it yet, we broke down that larger fight in our deep dive into the JAMA study claiming AI already beats doctors at diagnosis, where the American Medical Association pushed back hard against the researchers' conclusions.
This piece is about the other side of that argument. What actually happens when someone trusts an AI chatbot with a real medical decision, and it gets it wrong.
- What Actually Happened to the Man in This Case?
- How a Simple Diet Question Turned Into Bromide Poisoning?
- Why This Condition Is So Rare That Most Doctors Have Never Seen It?
- What ChatGPT Actually Said, According to the Doctors Who Treated Him?
- The Moment Everything Fell Apart: Hallucinations and an Involuntary Hold
- How the Man Recovered, and What His Doctors Learned?
- Why This Case Terrifies AI Safety Researchers More Than You'd Think?
- What OpenAI Has Said About Medical Advice and ChatGPT's Limits?
- The Bigger Pattern: Other Cases Where AI Health Advice Backfired
- What Doctors Want You to Know Before You Ask AI About Your Health?
- How to Actually Use AI Chatbots for Health Questions Safely?
- Frequently Asked Questions (FAQ)
What Actually Happened to the Man in This Case?
In August 2025, three physicians at the University of Washington published a case report that read less like a routine medical paper and more like a warning shot. The story was picked up fast, with outlets reporting on the details nationwide once the case became public.
Their patient was a 60-year-old man with no psychiatric history and no major medical issues. He wasn't reckless. He wasn't self-medicating with something illicit. He was doing something millions of people do every day: trying to eat a little healthier.
He'd read about the risks of too much sodium chloride, the chemical name for ordinary table salt. So he decided to cut it out of his diet entirely. And like a growing number of people looking for quick answers to health questions, he didn't call his doctor first. He asked ChatGPT for advice instead.
How a Simple Diet Question Turned Into Bromide Poisoning?
Here's where it gets genuinely alarming. When the man asked ChatGPT what he could use as a substitute for chloride, the chatbot reportedly pointed him toward bromide.
Chemically, bromide and chloride do sit near each other on the periodic table. In certain industrial and laboratory contexts, one can technically substitute for the other. But bromide is not something a human being should be putting into their body, and it is absolutely not something any licensed doctor, nurse, or pharmacist would ever recommend as a table salt replacement.
The man didn't know that. He sourced sodium bromide online and started using it in place of salt in his food, and by his own account, the chatbot's suggestion led him straight into the hospital. Not for a day or a week, but for three straight months.
Why This Condition Is So Rare That Most Doctors Have Never Seen It?
The diagnosis his doctors eventually landed on was bromism, or chronic bromide toxicity. If that name sounds unfamiliar, there's a good reason for that. Bromide compounds were once common in over-the-counter sedatives and anti-anxiety remedies back in the early 1900s, and historically accounted for a meaningful share of psychiatric hospital admissions at the time.
The United States Food and Drug Administration phased bromide out of consumer medications between 1975 and 1989 because of exactly this kind of toxicity risk. Since then, cases of bromism in otherwise healthy adults have become almost nonexistent.
What ChatGPT Actually Said, According to the Doctors Who Treated Him?
One detail from the case report stands out more than any other. The treating physicians didn't have access to the man's original conversation with ChatGPT. So they did something clever: they asked the chatbot the same question themselves, to see what kind of answer it would give a random user with zero medical context.
The AI once again suggested bromide as a possible substitute for chloride. It didn't ask why they wanted to know, and it didn't flag that bromide is toxic to humans in anything beyond trace industrial or veterinary use. The case has since been formally catalogued as a real-world AI safety incident, the kind of documented failure researchers now point to when they argue AI oversight can't be optional.
The Moment Everything Fell Apart: Hallucinations and an Involuntary Hold
By the time the man walked into the emergency room, three months of steady bromide accumulation had already taken hold. His vitals looked mostly normal on paper, which made the situation even more confusing for the team treating him at first.
Then things moved fast. He refused water even though he was clearly dehydrated. He grew convinced his neighbour was trying to poison him. Within the first 24 hours of admission, he began experiencing both auditory and visual hallucinations, and at one point tried to physically leave the hospital. Doctors had no choice but to place him on an involuntary psychiatric hold for his own safety.
Here's a quick look at how the symptoms lined up with classic bromism, according to the case details:
| Symptom | Present in This Case |
|---|---|
| Paranoia | Yes |
| Auditory hallucinations | Yes |
| Visual hallucinations | Yes |
| Refusal to drink despite thirst | Yes |
| Facial acne (a known bromide marker) | Yes |
| Insomnia | Yes |
| Prior psychiatric history | None |
That last row is the one that should give anyone pause. This wasn't a man with an underlying condition that tipped over. This was a healthy 60-year-old whose brain chemistry was derailed almost entirely by what he'd been eating for three months straight.
How the Man Recovered, and What His Doctors Learned?
The good news, and it's worth sitting with this for a second, is that he made a full recovery. Once doctors identified what was happening, treatment was straightforward: intravenous fluids to help flush the bromide out of his system, along with anti-psychotic medication to manage the hallucinations and paranoia while his body cleared the toxin.
Three weeks after admission, he was stable enough to be discharged. At a follow-up visit, his doctors confirmed he had returned to full health, with no lasting psychiatric or physical effects.
We looked at how AI is already reshaping diagnosis and treatment decisions in our piece on the facts behind AI vs. doctors in modern medicine, which covers where the technology is genuinely helping and where the line still needs to be drawn.
What's notable is what the treating physicians chose to publish alongside the case. They didn't just document the symptoms. They specifically noted that it's highly unlikely any human medical professional would have ever suggested sodium bromide as a table salt substitute, because a real clinician would have immediately recognised the toxicity risk and asked follow-up questions about why the patient wanted to know in the first place.
That's the gap this entire case exposes. Not that the AI made something up, but that it answered a question with technically accurate information while missing the human context that would have made all the difference.
Why This Case Terrifies AI Safety Researchers More Than You'd Think?
What makes this story spread the way it did isn't just the strange chemistry lesson at the center of it. It's what the case represents.
AI researchers have a term for this kind of failure: a "context gap." The model isn't hallucinating false information in the way people usually picture AI going wrong. It's retrieving something technically true and presenting it without the situational judgment a trained professional brings automatically. A pharmacist would never answer that question without asking why. The AI answered it instantly, with no follow-up at all.
This is part of why the case has been formally logged in databases that track real-world AI safety incidents, alongside other documented cases where AI systems caused measurable, real-world harm. It isn't an isolated internet horror story. It's treated as a data point in a much bigger conversation about where AI oversight is still missing.
If you want the fuller picture of how far AI has come in medicine, and where the hype outpaces reality, our breakdown of the "cure every disease" superintelligence claims digs into exactly that tension.
What OpenAI Has Said About Medical Advice and ChatGPT's Limits?
OpenAI's position on this has been consistent, at least on paper. When outlets reached out for comment on this exact case, the company pointed to its existing terms of service, which state plainly that ChatGPT is not intended for use in diagnosing or treating any health condition.
A company statement noted that safety teams are actively working to reduce these kinds of risks and that the model has been trained to encourage users to seek professional guidance for health-related questions. That's true in a lot of everyday interactions. Ask ChatGPT something that sounds like a medical emergency, and it will often recommend contacting a doctor.
But this case shows the limits of that safeguard. A question about a common cooking substitute didn't read as risky to the system, even though the answer it gave could have been fatal in higher doses. The warning systems built for obviously dangerous questions didn't fire, because the question itself didn't sound dangerous.
The Bigger Pattern: Other Cases Where AI Health Advice Backfired
This isn't the only documented instance of AI-generated health information causing real harm, and doctors have started speaking up about it publicly and often.
UCSF's Robert Wachter, one of the most prominent physician voices in the AI-in-medicine debate, has been candid that AI tools can confidently steer patients toward unproven or dangerous treatment paths if left completely unchecked, particularly around serious diagnoses where desperation makes people more likely to try something outside standard care. It's a pattern oncologists have separately flagged as a growing concern with unproven treatments spreading online, with AI chatbots now acting as one more channel repeating claims that haven't been through clinical trials.
Robert Wachter isn't the only physician raising this alarm, and the numbers back him up. A study run by pharmacists at Long Island University tested ChatGPT against 39 real medication questions the chatbot had never seen before, all of them pulled from actual patient calls to a drug information service. It answered only 10 of the 39 completely correctly.
One example from that same research lines up almost eerily with the bromide case. Researchers asked ChatGPT whether the COVID antiviral Paxlovid interacts with verapamil, a common blood pressure medication, and it said no interaction had been reported. That answer was wrong, and combining the two drugs in real life can cause a dangerous drop in blood pressure.
If you're curious how these same accuracy gaps show up in diagnosis rather than drug advice, our breakdown of what ChatGPT actually gets wrong in medical mysteries walks through several more documented cases.
The pattern across all of these stories isn't that the AI is lying. It's that it answers with total confidence, every time, regardless of whether the question is safe or genuinely dangerous. A pharmacist pauses. A chatbot doesn't.
What Doctors Want You to Know Before You Ask AI About Your Health?
Talk to enough physicians about this case and the same few points come up again and again.
First, an AI chatbot has no idea who it's talking to. It doesn't know your age, your kidney function, your current medications, or whether you're asking out of curiosity or because you're already in trouble. A human clinician asks. The model just answers.
Second, technically true isn't the same as safe. Bromide can chemically stand in for chloride in a lab setting. That fact is real. It's also completely irrelevant to whether a person should eat it, and no medical professional would have let that distinction slide the way the chatbot did.
Third, confidence is not the same as competence. Large language models are built to sound fluent and certain, even in the exact moments they should sound cautious. That's arguably the most dangerous trait an AI can have in a health context, because it mimics the tone of expertise without the judgment behind it.
The American Medical Association's own physician sentiment survey has repeatedly found doctors saying AI tools still need direct human oversight before anything reaches a patient, especially for less common conditions where the model has less reliable training data to draw from.
How to Actually Use AI Chatbots for Health Questions Safely?
None of this means you have to avoid AI chatbots entirely. It means using them the way you'd use a knowledgeable friend who isn't a doctor: useful for a starting point, never the final word.
- ✅ Use it to understand a term your doctor used, in plain language
- ✅ Use it to prep questions before an appointment
- ✅ Use it to get a general overview of a condition you were just diagnosed with
- ❌ Don't use it to choose a substitute for any food, supplement, or medication without checking with a pharmacist first
- ❌ Don't use it to interpret your own lab results in isolation
- ❌ Don't use it as a replacement for a second opinion when something feels seriously wrong
- ❌ Don't assume a confident-sounding answer has been checked by anyone
If you ever do get a suggestion from an AI chatbot involving a supplement, a substitute ingredient, or anything you'd put in your body, the safest move is a quick call to a pharmacist or your doctor's office before you act on it. Poison Control's free national hotline is also available 24/7 for exactly this kind of question.
Frequently Asked Questions (FAQ)
Is it true that ChatGPT told someone to use bromide instead of salt?
According to a case report published by University of Washington physicians, the patient said a conversation with ChatGPT led him to substitute sodium bromide for table salt. When the treating doctors later asked the same chatbot a similar question themselves, it suggested bromide again, without warning about toxicity.
What is bromism and is it still common today?
Bromism is chronic bromide toxicity. It causes neurological and psychiatric symptoms including paranoia, hallucinations, and insomnia. It was relatively common in the early 1900s when bromide was used in over-the-counter sedatives, but the FDA phased those products out between 1975 and 1989, making modern cases extremely rare.
Did the man in this case fully recover?
Yes. After three weeks of inpatient treatment involving IV fluids and antipsychotic medication, he stabilized and was discharged. A follow-up visit confirmed no lasting physical or psychiatric effects.
Does OpenAI say ChatGPT shouldn't be used for medical advice?
Yes. OpenAI's terms of service state that ChatGPT is not intended to diagnose or treat any health condition, and the company has said it's working to strengthen safeguards around health-related questions.
Is this the only documented case of AI health advice causing harm?
No. Multiple studies, including pharmacist-led research on medication questions, have found AI chatbots giving incomplete or inaccurate answers to real drug-related questions, sometimes missing dangerous interactions entirely.
Editorial note: this article is a news summary of a published medical case report and related research. It is not medical advice. Any question about medications, supplements, or dietary changes should go through a licensed physician or pharmacist.