An 82-year-old man with Parkinson's Disease was living in a senior residential facility when a serious Norovirus outbreak struck his building. Immunocompromised and at heightened risk, he used ChatGPT to research what the outbreak meant for someone in his situation. Based partly on what he learned, he made the decision to temporarily leave the facility and isolate himself. He believes that decision may have protected his life.
His account illustrates something that rarely surfaces in AI tool coverage: for a significant portion of the population, an AI that explains medical risk in plain language isn't a convenience. It's filling a gap that the healthcare system leaves open.
What AI Does Well in This Role
ChatGPT can't diagnose anyone. It doesn't have access to your labs, your chart, or your history unless you tell it. But it can explain what Norovirus is, why it poses heightened danger in communal living environments, what meaningful isolation looks like in practice, and what symptoms require emergency care. For someone making a risk assessment rather than seeking a diagnosis, that's genuinely useful information - delivered faster and in clearer language than many patients get from overloaded providers.
The pattern that consistently works: AI as a research tool when you have a specific, bounded question and need context to make a more informed decision. An elderly man weighing whether to stay in a Norovirus-hit building fits that profile well. He understood his own conditions, knew his environment, and used ChatGPT to fill in the clinical context he was missing.
The Problem With the Same Tool
The quality that makes ChatGPT useful here - confident, clear explanations - is also what makes it dangerous in other situations. Someone asking whether their chest pain requires emergency care gets the same confident tone from the model. Someone asking whether it's safe to stop a prescription medication gets the same confident tone. ChatGPT doesn't calibrate its certainty to the stakes involved in a question.
OpenAI encourages users to consult healthcare professionals, but there's no triage layer built into the interaction - no mechanism that distinguishes a relatively safe risk-assessment question from a genuinely dangerous one. The entire burden of evaluating AI health information falls on the person asking. That works when they're medically literate enough to know what they're asking, and fails badly when they're not.
ChatGPT is already functioning as informal health infrastructure for people who lack alternatives: elderly adults with mobility constraints, rural communities far from specialists, people who can't afford frequent care. That role is real and growing. How to make it safer - through better model behavior, clearer scope limits, or something else - remains genuinely unresolved.