Read, but do not write
OpenAI connected ChatGPT to Epic on 1 September, letting healthcare organisations bring authorised patient records from their Epic environment into ChatGPT for Healthcare. The access runs one way: the integration is read-only and the model writes nothing back to the chart, OpenAI told TechCrunch. Epic’s system holds records for more than 325 million patients.
The company describes two shapes for it. In the first, a clinician pulls authorised patient information into ChatGPT to review history, spot changes and prepare for an appointment. In the second, available in supported deployments, ChatGPT is embedded in the EHR layout itself, so the assistant sits inside the chart rather than beside it.
A second connector, for public data
Alongside the Epic work, OpenAI shipped a Healthcare Public Data plugin with dedicated connectors to nine official public sources. It named five of them: ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed.
The distinction OpenAI draws is between searching those sources and querying them. The plugin is built to work with specific records, fields, identifiers and versions, which is what separates confirming the current label for a medication from finding a page that discusses it.
The safety numbers are OpenAI’s own
OpenAI said physicians evaluated the model’s responses across 27 clinical use cases, including pre-visit review, clinical timelines, medication review and handoff summaries. Across 4,363 ratings, it said physicians rated 99.1% of responses safe. In a separate two-round evaluation, it said more than 93% of responses were rated good or better on accuracy for each of the five connected data sources tested.
Those are the company’s own evaluations of its own product, and the residual is the part worth holding onto. Nine tenths of one per cent of 4,363 ratings is roughly forty responses. OpenAI has acknowledged that even a few unsafe answers can produce harmful results, and it is being sued in two cases reported by TechCrunch: one brought by a Florida pastor who alleges ChatGPT gave him a near-fatal recommendation, and one by family members who blame it for wrongful dosage advice.
For scale on the review programme behind the numbers, OpenAI said it works with hundreds of physicians across 60 countries, 49 languages and 26 medical specialties, who have reviewed more than 700,000 model responses to date.
Who can actually turn it on
Access is tiered and narrower than the announcement suggests. ChatGPT for Healthcare customers ask a workspace administrator to enable both features. ChatGPT Enterprise customers go through an OpenAI account team to confirm eligibility for a Regulated Workspace. Eligible individual clinicians in the United States can install the Healthcare Public Data plugin, but the EHR integration is not available for individual accounts at all.
HIPAA-compliant workflows require an applicable Business Associate Agreement. The workspace carries role-based access, single sign-on and audit logs. Suresh Gunasekaran, president and chief executive of UCSF Health, said in OpenAI’s announcement that the health system is a pilot partner and is validating the capabilities with frontline teams.
What to watch
Two things will tell you whether this is a product or a pilot. Whether pilot partners such as UCSF Health publish their own evaluations rather than restating OpenAI’s, which is the only way the 99.1% figure gets independent weight. And whether the read-only boundary holds: writing back to the chart is the obvious next step commercially, and the first integration to cross that line will be the one worth reading closely.
