TL;DR

OpenAI has wired ChatGPT into Epic, the electronic record system used by many of the world’s largest hospital groups, so clinicians can interrogate an authorised patient chart in conversation rather than clicking through it. A second release lets the same workspace query nine official public medical datasets directly. Both are American launches, governed by HIPAA arrangements that have no UK equivalent.

What clinicians can actually ask

The demonstration questions are mundane, which is the point: what has changed since this person was last seen, which recent results need looking at before today’s appointment, what referrals or follow-ups are outstanding. The system assembles an answer from the record and links back to the underlying chart entries.

It runs two ways. Clinicians can pull record context into ChatGPT, or in supported deployments the assistant is embedded in the record layout itself so nobody leaves the chart. UCSF Health, a pilot partner, framed the value as time not spent synthesising a complicated history.

The separate data plugin covers PubMed, DailyMed, RxNorm, CMS Coverage and ClinicalTrials.gov among others, and is aimed at precision work — comparing trial eligibility criteria, or confirming the current label and warnings on a drug — where searching each source individually is the current cost.

The evaluation numbers

OpenAI put the record-connected behaviour in front of physicians across 27 clinical scenarios, from pre-visit review to handover summaries. Of 4,363 ratings collected, 99.1% judged the response safe. A second exercise on nuanced questions against large US datasets returned better than 93% “good” or above on accuracy for every source tested. The company says its physician network spans 26 specialties in 60 countries and has now reviewed over 700,000 responses.

Safe is not the same as correct, and neither figure is externally audited.

Looking forward

Epic is not an abstraction for the NHS — Cambridge University Hospitals and UCLH have both run on it for years — so the technical path from this announcement to an English trust is short. The governance path is not. Nothing announced here addresses UK data protection duties, national record-sharing rules, or how a trust would evidence clinical safety to its own regulator. The capability will arrive well before that framework does, which is the recurring shape of clinical AI procurement.