OpenAI has connected ChatGPT Health to Epic, the electronic health record system that holds charts for 325 million patients, letting clinicians pull visit notes, lab results, medication lists and specialist documents into a chat window. The connection runs one way: ChatGPT reads the chart and writes nothing back into it. It went live days after a Florida pastor sued OpenAI, alleging that ChatGPT gave him a recommendation that nearly killed him.
Inside the record, the model summarises what a clinician would otherwise page through — encounters, labs, medications, documents from specialists. It can walk a patient's history, surface changes between visits, and prepare for the next one. In some deployments the physician runs a pre-visit review and assembles a clinical timeline without leaving the patient chart.
Alongside the Epic link, OpenAI shipped an extension called Healthcare Public Data, which pulls from ClinicalTrials.gov for trial data, CMS Coverage for insurance rules, RxNorm for drug identifiers, DailyMed for drug information and PubMed for the literature. The company says it helps clinicians line up trial eligibility criteria, drug identifiers, versions of coverage rules and provider records against each other.
Organisations that have signed a healthcare data agreement with OpenAI can now also use ChatGPT Work, Codex, apps and connectors inside their workspaces for compliant workflows. That is the quieter half of the announcement and the part with the longest tail: it turns a clinical feature into a procurement relationship covering a hospital's general-purpose software, not just its charts.
The safety case rests on one test. OpenAI gathered more than 4,300 clinician ratings across 27 clinical scenarios — pre-visit review, clinical timelines, medication checks, handoff summaries — and reports that 99.1 percent of responses were safe. The company also repeats, as it has all along, that AI is not suitable for diagnosis or for prescribing treatment.
Run the arithmetic the other way and 99.1 percent of roughly 4,300 rated responses leaves somewhere near forty that clinicians did not call safe. In a consumer chat about a headache, forty bad answers out of 4,300 is a support problem. In a medication check inside a live chart it is a different category of number, and it is the only category that matters, because a summary is not consulted the way a search result is — it is read once, believed, and acted on. The 99.1 percent is also OpenAI's own figure, from 27 scenarios OpenAI chose. Nothing in the announcement says what the twenty-eighth scenario would have been.
The other number doing heavy lifting is 300 million — the health questions people ask ChatGPT every week, a figure OpenAI first gave a month ago when it opened ChatGPT Health to all US users. That is a consumer number, and it is being used here as evidence of clinical demand. They are not the same thing. Notably absent from the announcement: how many health systems or clinicians actually have the Epic integration running, and whether patients are told that a third-party model read their chart.
The lawsuits track the consumer product, not this one. The Florida pastor's claim concerns advice ChatGPT gave him directly; in May, relatives of another user sued over what they say was an erroneous medication dosage recommendation. Both involve a person talking to a chatbot with no clinician in the loop.
That distinction is exactly what the read-only wall is built to preserve. It keeps the model's output out of the legal record and keeps a licensed human between the summary and the patient. It does nothing about the failure mode that actually threatens a doctor using this product: a summary that quietly omits an interaction is already wrong in the physician's head, and it never has to be typed into Epic to do harm.