ChatGPT Health Plugs Into Epic’s 325M Patient Records

OpenAI has connected ChatGPT Health to Epic, the electronic health record system that holds data on more than 325 million patients, letting clinicians pull a patient’s chart into ChatGPT and ask questions of it. TechCrunch reported the integration on September 1.

The scope is deliberately narrow. Clinicians can import appointment notes, lab results, medication lists and specialist documentation, then have the model summarise a history or assemble a clinical timeline ahead of a visit. What it cannot do is write. The connection is read-only, so ChatGPT cannot place orders, amend documentation, or push anything back into the record. In some deployments the assistant is surfaced inside Epic’s own interface rather than in a separate window, which is where a tool like this either saves a clinician time or does not get used at all.

Read-only is the whole safety argument, and it is a defensible one. A model that misreads a lab value and tells a clinician the wrong thing is a serious problem. A model that misreads a lab value and writes it into the permanent record is a different category of problem, because the error then propagates to every downstream reader who assumes the chart is ground truth. Restricting the integration to retrieval keeps a trained human between the model’s output and the document of record.

Alongside the Epic link, OpenAI added a Healthcare Public Data plugin that reaches ClinicalTrials.gov, CMS coverage determinations, RxNorm, DailyMed and PubMed. Those are the unglamorous lookups that eat clinical time — checking trial eligibility criteria, resolving a medication identifier, working out which version of a coverage policy applies to a given claim. Organisations that have a Business Associate Agreement in place can use ChatGPT Work, Codex, apps and connectors under HIPAA.

The clinician release lands on top of a consumer product that is both enormous and contested. ChatGPT Health reached US consumers in July, and OpenAI has said the platform fields roughly 300 million health-related queries a week. It has also drawn lawsuits alleging the assistant gave harmful medical advice — litigation that remains unresolved. The professional-facing version does not settle that argument, though it does move the model into a setting where someone qualified is reading its output before anyone acts on it.

Epic is the practical reason this counts for more than a typical integration. It is the dominant record system in American hospitals, so a single connector reaches a large slice of US clinical practice at once instead of one health system at a time. The open question was never whether clinicians would try it. It is whether summarising a messy, decades-long chart turns out to be accurate enough to lean on under time pressure — and that answer will come out of audits and incident reports, not launch posts.