ChatGPT for Healthcare now reads a specific patient’s chart instead of only answering medical questions in the abstract. OpenAI’s new Epic integration, announced September 1, pulls “authorized patient context” — notes, labs, medications, specialist documentation — into ChatGPT so a clinician can ask what changed since the last visit. A companion Healthcare Public Data plugin adds structured access to nine official sources. The value OpenAI is selling is synthesis, and synthesis is exactly the task where an error is hardest to catch: the clinician asked for the summary to avoid reading everything underneath it.
Two ways in: inside ChatGPT, or inside the chart itself
OpenAI describes two modes. In the first, ChatGPT pulls the authorized Epic record into the conversation and, per the announcement, “points back to supporting chart information” for each answer. In the second, available only “in supported deployments,” ChatGPT sits directly inside the EHR layout the clinician already uses, rather than in a separate window.
Nine public datasets, open to individual clinicians too
The Healthcare Public Data plugin connects nine official sources, including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed. Access to it splits along the same line as everything else in this launch: a ChatGPT for Healthcare admin turns on the Epic integration for a whole organization, and an Enterprise customer arranges a “Regulated Workspace” through its account team, but the EHR integration itself “is not available for individual accounts.” Eligible individual clinicians in the US can install only the public-data plugin — the chart stays behind the organizational gate.
A pilot at UCSF, and a detail OpenAI’s own page skips
UCSF Health is named as a pilot partner, and its CEO describes the hospital as “exploring” the integration and “engaging frontline teams to validate these capabilities”: language that describes evaluation, not a live rollout. On what the integration can do to the record, OpenAI’s announcement is silent; it is TechCrunch that reports the connection is read-only, quoting that “AI doesn’t write anything back.” OpenAI’s own page never makes that claim.
The number behind the reassurance, and its other half
OpenAI backs the launch with role-based access, single sign-on, audit logs, and a Business Associate Agreement that lets customers run HIPAA-compliant workflows, the standard enterprise healthcare checklist. Its clinical evidence is a physician evaluation: across 4,363 ratings spanning 27 clinical use cases, physicians rated 99.1% of responses safe. Read whole, that same evaluation means around 40 ratings were not judged safe, and OpenAI has not published who the raters were, how “safe” was defined, or what those responses looked like. A separate two-round evaluation reported more than 93% “good or better” accuracy for each of five connected data sources; again OpenAI’s own measurement, without an outside audit. The company says it works with physicians across 60 countries, 49 languages and 26 specialties who have reviewed more than 700,000 model responses, a scale claim that does not itself describe this evaluation’s method. TechCrunch adds context the announcement omits: lawsuits, reported by the New York Times, already allege harmful ChatGPT medical advice, and its own framing is that “even a few unsafe answers can lead to harmful results” once a tool reaches clinical scale.
What the announcement does not say
No statement from Epic appears in the sources reviewed, and neither describes the technical path: whether it runs through FHIR APIs, Epic’s app marketplace, or a direct arrangement. OpenAI has not published a price; ChatGPT for Healthcare is sold through its sales team. And the announcement does not say whether patient data passing through the integration is excluded from model training; the announcement leaves that open.