OpenAI announced on 1 September that healthcare organisations can connect their Epic electronic health record environments to ChatGPT for Healthcare, with launch organisations including Cedars-Sinai, Memorial Sloan Kettering, HCA Healthcare, UCSF Health, Boston Children's Hospital, AdventHealth and Baylor Scott & White. Clinicians can pull appointment notes, lab results, medications and specialist documentation and ask questions across a record; some deployments place ChatGPT inside the EHR interface itself.
What the safety number actually is
The evidence offered is a vendor-run physician rating exercise, not a trial and not a clinical study: 4,363 physician ratings across 27 clinical use cases, of which 99.1% were rated safe. That leaves roughly 39 responses below the safety bar. There is no published methodology, no peer review and no independent replication, and the raters were engaged by the vendor. A separate accuracy range of 93.2% to 98.6% is reported for the five public-data connectors — ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed.
What the common framing gets wrong
Two distortions run in opposite directions. “325 million patients” is Epic's total footprint, not the reachable population — the connection is per-organisation, opt-in, and needs a HIPAA Business Associate Agreement, so day-one reach is whatever subset of Epic customers switches it on. And 4,363 is a count of ratings, not of physicians, which several write-ups conflated. Nothing here was cleared or approved by the FDA; clinical decision support of this shape sits outside device regulation, so “cleared for hospital use” language is simply wrong.
The read-only boundary cuts both ways
Access is read-only — the model writes nothing back to the chart. That is the governance design, and it is also a workflow limit: every output remains an unrecorded suggestion a clinician must re-key. It is the constraint most likely to come under commercial pressure first, because a suggestion nobody can save is a suggestion someone has to retype.
Where the residual risk sits
The ~39 responses that failed the bar arise precisely in the workflow being sold — a clinician reading a summary of a chart they have not opened themselves. OpenAI maintains, in the same announcement, that AI remains unsuitable for diagnosis or treatment. Epic has issued no statement of its own.