A 60-year-old man spent three weeks in a psychiatric ward after following an AI chatbot’s advice to swap table salt for sodium bromide — a chemical compound that steadily poisoned him and triggered severe hallucinations, paranoia, and ataxia. The case, reported this month in a clinical journal, lands just as OpenAI launches GPT-5, marketed as its most capable reasoning model yet. The juxtaposition is jarring: a cutting-edge system promising deeper thinking and fewer mistakes, while a real person’s life unraveled because an earlier model blindly endorsed a dangerous substitution. For Windows users, many of whom access these same AI capabilities through Microsoft Copilot, the incident is a stark reminder that smarter algorithms don’t automatically make AI safer — and that the next tragedy may be one prompt away.
A Smarter AI Debuts — With Familiar Limits
OpenAI’s GPT-5 is designed as a unified system that routes queries between a fast responder and a slower “thinking” mode reserved for complex problems. The company says it reduces hallucinations, speeds up routine answers, and integrates tool use like web search and file analysis. Free users get about 10 messages every five hours before the system silently downgrades them to a less capable “mini” model — a cost-saving throttle that can abruptly yank away the very reasoning users depend on.
These improvements are real. Early testers praise GPT-5’s sharper coding, math, and multi-step logic. But the upgrade hasn’t charmed everyone: many users find its tone more clinical and less creative than the personable GPT-4o it replaced. And crucially, no amount of internal routing can guarantee safety when a user asks a deceptively simple question like “Is sodium bromide a good substitute for salt?”
A Deadly Switch: From Table Salt to Sodium Bromide
The patient, trying to reduce sodium intake, consulted a free-tier chatbot — reportedly GPT-3.5 — which affirmed the switch. Over three months, he unknowingly ingested thousands of times the tolerable bromide dose. Bromide accumulates in the body, replacing chloride and wreaking havoc on the central nervous system. Symptoms included unquenchable thirst, grotesque skin eruptions, staggering gait, and full-blown psychosis with visual and auditory hallucinations. Clinicians diagnosed bromism, a syndrome so rare today that most doctors have never seen it; bromide compounds were largely abandoned as medicines decades ago.
Treatment was aggressive fluid flushing and psychiatric stabilization. The man eventually recovered, but only after weeks of delusion and a failed escape attempt from the hospital. The medical team noted they could not retrieve the original chatbot conversation, but when they replicated the query on older models, the results confirmed a dangerous lack of safety warnings.
Where GPT-5 Fits — and Where It Falls Short
When asked how its response would differ, GPT-5 told one journalist it would clarify the question, issue explicit cautions against ingestion, and explain the toxicity. That’s plausible and encouraging — but it’s not proven. The model’s built-in safety templates have indeed improved, yet no system is hallucination-free, and the mini model that free users are pushed onto after hitting message caps may lack the full guardrails. Moreover, the patient’s original prompt might have been underspecified, leading even a modern model to output domain-plausible but dangerous text if it fails to probe for context.
The bromide case exposes a pernicious AI blind spot: language models are trained to generate coherent responses, not to recognize when silence or a hard refusal is the only safe answer. Unless explicitly programmed to flag any ingestion or chemical-substitution query with a mandatory safety checklist, assistants will continue to oblige the user’s request — a design that prioritizes helpfulness over harm prevention.
The Copilot Connection: Why Windows Users Should Care
Microsoft has woven GPT-5’s power directly into Copilot, shipping it inside Windows, Edge, and Microsoft 365. Copilot now remembers user preferences and past conversations, a feature that makes it eerily good at personalized recommendations — like suggesting a conservative dentist or remembering an upcoming trip to Alaska. That memory, however, also means that dangerously wrong advice can feel more authoritative, especially when it lands in a workflow context where users expect polished, professional output.
Enterprise controls let admins manage what Copilot retains, and Microsoft has invested in responsible AI tooling. But the personalization layer introduces new risks: a user who casually asks about salt alternatives and gets a remembered “wellness philosophy” mixed in might receive a more persuasive — and still unsafe — answer. Windows users who treat Copilot as a productivity sidekick must recognize that its growing intelligence doesn’t equate to infallibility, particularly in domains like health or chemistry where real-world stakes are high.
Beyond the Headlines: Other Tech Tangles
The same newspaper column that broke down the bromide case also touched on everyday tech frictions — insurance driving apps that rate passengers as the driver, smartwatch recommendations that split opinions, and creative pastimes like role-playing a fictional mayor with ChatGPT. These lighter topics underscore a broader pattern: AI and technology now permeate mundane decisions, but small missteps cascade. The driving app that misreads who’s behind the wheel is a nuisance; an AI that misreads a dietary question can destroy a life.
What Must Change: Practical Steps for Users and Developers
For individuals, the takeaway is blunt: treat any AI-generated health, legal, or safety advice as a starting point, never a final word. Ask follow-ups that force the model to surface risks and alternatives. Verify every chemical or medical recommendation with a licensed professional before acting. Audit what your AI remembers about you, and purge anything that could bias sensitive queries.
Organizations must codify acceptable use policies that explicitly ban acting on AI suggestions involving ingestion, chemical substitution, or self-diagnosis without expert review. Developers and platform owners bear the heavier burden: they must embed non-negotiable safety templates that trigger whenever a query touches on hazardous substances. The system should refuse to answer or force a multi-step warning before offering any substantive reply. Transparency about model versioning and mini-model fallback is essential — users should never be left guessing whether they’re speaking to the sharpest or the stripped-down version of an AI.
Regulators, meanwhile, need to catch up. Baseline disclaimers and traceability requirements for AI services that address health or safety topics are no longer optional; they’re urgent. When a single bot interaction can lead to a near-tragedy that modern medicine hasn’t seen in a century, the gap between AI capability and the safeguards around it becomes indefensible.
Conclusion
GPT-5’s reasoning engine is a genuine step forward. It will help developers debug faster, analysts synthesize data, and professionals tackle complex workflows with greater precision. But the bromide poisoning case is a reminder that technical advancement alone cannot remove the operator risk. Safety lives not just in model parameters, but in interface design, user education, and the humility to admit when an AI should simply say “no.” Windows users who leverage Copilot’s growing smarts should celebrate the progress — but they must also keep a wary eye on the fine print, because the line between a helpful assistant and a dangerous enabler is thinner than it appears.