OpenAI announced this week that healthcare organizations can now connect their Epic electronic health record systems directly to ChatGPT for Healthcare, letting clinicians pull a patient’s appointment notes, lab results, medication history and specialist documentation into a chat interface without leaving the medical chart. Epic’s software underpins the records of more than 325 million patients, making this one of the largest single integrations OpenAI has struck with a healthcare vendor.
The connection is read-only. ChatGPT can retrieve and summarize what’s already in a patient’s record, flag what changed since the last visit, surface new lab results and highlight follow-ups nobody’s closed out, but it cannot write anything back into the chart itself. OpenAI is also rolling out a separate Healthcare Public Data plugin that links ChatGPT to nine outside sources clinicians already rely on, including ClinicalTrials.gov, CMS coverage databases, RxNorm and PubMed.
How OpenAI says it tested this
OpenAI says physicians evaluated the system across 27 clinical scenarios, including pre-visit chart reviews, medication reconciliation, care handoffs and building out a patient’s full clinical timeline, and rated 99.1% of nearly 4,400 individual responses safe. That’s OpenAI’s own number from its own evaluation, not an independent clinical trial, and it’s worth reading with the same skepticism due any company grading its own homework on a product it’s trying to sell into hospital systems.
The overlap problem
Epic already has its own AI features built into its records system, and hospitals have spent the last few years layering on additional AI tools from Microsoft, Google and smaller health-tech vendors. ChatGPT plugging directly into that stack adds a third or fourth AI layer doing overlapping work, and hospital IT departments are now stuck figuring out which tool a clinician should trust for which task, and who’s accountable when two AI systems summarize the same chart differently.
The upside case is straightforward: fewer clicks, faster chart reviews, less time hunting through tabs for a lab result that’s already sitting in the record. The tradeoff is that every additional AI layer sitting between a doctor and a patient’s actual chart is one more place for something to get summarized wrong, and read-only access doesn’t eliminate that risk, it just limits the damage to bad information rather than a bad edit.
Sources: OpenAI and TechCrunch.
