ChatGPT Health now plugs directly into Epic's patient charts, covering more than 325 million records -- read-only, and OpenAI says its own safety testing rated 99.1% of responses safe

OpenAI said on September 1 that ChatGPT for Healthcare now connects directly to Epic, the electronic health record system that holds charts for more than 325 million patients -- letting authorized clinicians pull a patient's notes, lab results, medications, and specialist documentation into ChatGPT, or use ChatGPT from inside Epic's own workflow, without leaving the chart. The connection is read-only: nothing a clinician does inside ChatGPT writes back to the patient's actual record.

Two distinct workflows are live now. A clinician can pull authorized patient data into ChatGPT directly to review a history and prepare for a visit, or -- in supported deployments -- use ChatGPT inside the Epic interface itself, without switching applications. OpenAI frames the target use cases narrowly: which screenings a patient is overdue for, what the clinician should review before a visit, a fast synthesis of a chart most doctors would otherwise piece together by scrolling. UCSF Health's chief executive, Suresh Gunasekaran, said the tool has the potential to reduce time spent synthesizing data and give clinicians more time with patients -- the only named health-system endorsement in the launch so far.

Alongside the Epic connection, OpenAI also launched a Healthcare Public Data plug-in that pulls from nine official sources, including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed, so a research team can check which trials are actively recruiting or a pharmacy team can confirm a drug's label and warnings without searching each database separately. The two launches are aimed at different jobs -- one reads a specific patient's own chart, the other reads the public medical literature and regulatory record -- but both are read-only lookups layered onto ChatGPT rather than new clinical judgment the model is making on its own.

Compliance is handled the same way any other regulated ChatGPT deployment is: organizations that sign a Business Associate Agreement with OpenAI can use ChatGPT Work, Codex, and connectors including this one for workflows covered by US patient-privacy law, rather than needing a separate walled-off product. That puts OpenAI in direct competition with EHR-native AI tools already selling into the same hospitals -- Microsoft's Dragonfly and Nuance DAX Copilot, and ambient-documentation vendors like Abridge, all pitch a version of the same promise: less time typing into the chart, more time with the patient. What OpenAI has that most of those competitors don't is Epic's own direct plumbing and a general-purpose assistant already familiar to millions of clinicians outside the hospital, which is as much a distribution advantage as a technical one.

OpenAI's own safety claim is specific and, on its face, reassuring: physicians rated 99.1% of 4,363 responses as safe across 27 clinical use cases, and more than 93% of responses were rated "good" or better for accuracy across the five connected public-data sources. Those numbers are OpenAI's study, using OpenAI's chosen use cases and raters, and no outside body has published an independent replication. OpenAI is explicit on the point that matters most for a health-adjacent product: the company maintains that ("AI-unsuitable-for-diagnosis" is OpenAI's own stated boundary, not this newsroom's characterization of what the tool can safely do.) ChatGPT is not suitable for diagnosis or treatment decisions, positioning the integration as a documentation and preparation aid a clinician still has to read critically, not a second opinion.

The integration lands as one of the largest single connections between a general-purpose chatbot and a live patient-records system to date -- Epic's own reach means most large US hospitals already run on it, so "connect to Epic" is a much bigger distribution move than a standalone health app could manage on its own. That scale is exactly why the read-only boundary matters: a summarization tool that misreads a chart wastes a clinician's time, but a tool that could write back into the record and misread it would risk propagating an error into the chart itself. OpenAI's design keeps the human as the last step before anything touches the patient file -- for now, and only for the deployments that stay read-only as this expands beyond UCSF.

One question the launch materials and the coverage of it both leave open: what a patient is actually told about an AI system reading their chart, or whether they can decline having a clinician use ChatGPT to review their history at a given health system. Epic and OpenAI's public materials describe controls at the clinician and institution level -- role-based access, the Business Associate Agreement, audit logging -- not a patient-facing disclosure or opt-out. That's a common gap in a day-one product announcement, but it's the one part of this rollout that neither company's own account nor the reporting around it has addressed yet.