OpenAI has connected ChatGPT Health to Epic, the electronic health record system that holds the medical files of more than 325 million patients. That is not a pilot with a clinic; it is a bridge between the most-used consumer AI on earth and the records infrastructure behind a huge share of American healthcare. The promise is enormous. So are the questions, and both deserve saying out loud.
What actually happened
Epic is the software most large US hospitals use to store patient records: notes, test results, medications, history. By wiring ChatGPT Health into it, OpenAI is aiming at a real and painful problem, that patients cannot understand their own records and clinicians drown in admin. An AI that can read your chart and explain it in plain language, or draft the paperwork a doctor spends hours on, would genuinely help. Healthcare is one of the few places AI could save lives and hours at the same time.
Why this is a big deal
Most AI in medicine so far has been walled into narrow tools. Connecting a general-purpose assistant to the actual record system for 325 million people is a different scale of ambition. If it works, it could make the maddening experience of navigating your own care dramatically simpler, and hand clinicians back time they currently lose to typing. The prize is real.
The friendly sceptic’s corner
Now the part that should make everyone slow down. Medical records are the most sensitive data most people will ever generate, and the history of tech plus health data is not reassuring. Three worries stand out. First, privacy: who can see what, under whose consent, and how tightly is that enforced. Second, accuracy: a language model that invents a plausible wrong answer about your holiday plans is a nuisance; one that does it about your medication is a danger, and these systems still hallucinate. Third, consent that is real rather than buried in a form nobody reads. None of this makes the project wrong. It makes the guardrails the whole story.
What to watch
The tell will be in the details OpenAI and Epic publish about data handling: whether records are used to train models, who holds the keys, what a patient can see and switch off, and how errors are caught before they reach a decision. Watch for independent clinical validation rather than demos, and watch what regulators say. In healthcare, “move fast” is precisely the instinct that has to be resisted, and the responsible version of this is slow on purpose.
What this means
This is one of the most consequential AI deployments of the year, aimed at a genuinely broken experience and carrying genuinely serious risk. The upside is patients who finally understand their own care and doctors who get their evenings back. The downside is a privacy or accuracy failure at a scale of hundreds of millions. Both are on the table, which is why the guardrails, not the ambition, are what will decide whether this is remembered as a breakthrough or a cautionary tale.
Related on Top Tool Stack: Gemini aimed at Big Law · Who controls your AI logs
Did you know: Epic’s software touches the records of more than 325 million people, most of the US population. Connecting a consumer chatbot to it is one of the largest single expansions of AI into private data yet attempted.