Ninety-nine percent of clinical responses rated safe across 4,363 physician evaluations. That’s the number OpenAI is leading with as it announced two major additions to ChatGPT for Healthcare: a live Epic EHR integration and a new Healthcare Public Data plugin connecting directly to nine official sources. For an industry that has watched AI companies make big promises with thin follow-through, these are unusually specific claims backed by unusually specific numbers.
What the Epic integration actually does
The EHR connection lets clinicians bring authorized patient records into ChatGPT without leaving their workflow. Ask what changed since a patient’s last visit, which lab results need attention before an appointment, or whether there are unresolved referrals or medication changes. ChatGPT pulls from the authorized chart, summarizes what matters, and points back to the source documentation.
There are two deployment modes. The first brings patient context into ChatGPT itself. The second embeds ChatGPT directly into the EHR layout, so clinicians never have to switch applications. UCSF Health is among the pilot partners, with their President and CEO Suresh Gunasekaran describing the goal as reducing time spent synthesizing patient data so clinicians can spend more of it with patients.
That’s a real problem worth solving. Clinician burnout driven by documentation and information-hunting is well documented. But whether ChatGPT meaningfully reduces that burden at scale, across varied patient complexity and institution types, is still an open question. Pilots are not deployments.
Nine public data sources, one plugin
The Healthcare Public Data plugin is the more immediately verifiable addition. It connects ChatGPT to nine official sources through dedicated connectors, not general web search. The list includes:
- PubMed
- DailyMed
- ClinicalTrials.gov
- CMS Coverage
- RxNorm
- And four additional official datasets
The difference between this and asking ChatGPT to search the web for drug information is significant. Teams can query specific fields, identifiers, and versioned records. A pharmacy team can confirm the current label for a medication. A research team can filter ClinicalTrials.gov for actively recruiting studies and compare eligibility criteria side by side. A population health team can combine Medicare coverage data, relevant trials, and published research in a single workspace view. That kind of structured, source-backed retrieval is where general-purpose AI has historically struggled in healthcare settings.
How OpenAI is evaluating performance
OpenAI says it works with physicians across 60 countries, 49 languages, and 26 specialties who have reviewed more than 700,000 model responses to date. For the EHR-specific evaluation, physicians rated responses across 27 clinical use cases including pre-visit summaries, medication review, and handoff notes. The 99.1% safety rating across 4,363 responses is notable, though safety and accuracy are different things.
On accuracy, a separate evaluation tested ChatGPT against five connected data sources using large U.S. healthcare datasets. More than 93% of responses received “good” or better accuracy ratings from physicians across each source. That’s a reasonable bar, but the methodology matters. OpenAI should expect close scrutiny on how these evaluations were designed, particularly who defines “good” and whether the test cases reflect real clinical complexity.
Where this fits in the competitive picture
Microsoft’s Nuance DAX and Google’s Med-PaLM have been the names most associated with clinical AI in enterprise settings. Epic itself has been building AI features into its own platform. So OpenAI is entering a space where the incumbents have deep integration histories and existing trust relationships with health systems.
But OpenAI has something the others are still building toward: a general-purpose workspace model that clinical, administrative, and technical teams can all use. ChatGPT for Healthcare includes the same Work and Codex access as the enterprise product, with role-based access controls, SSO, audit logs, and a Business Associate Agreement covering HIPAA workflows. AdventHealth’s Chief AI Officer called out the value of putting ChatGPT Work and Codex in the hands of non-clinical teams, not just clinicians.
That’s the real strategic play here. Not replacing Epic. Not replacing PubMed. Connecting them, and everything else a health system runs on, inside a single governed workspace. Whether health systems are ready to trust OpenAI with that role is a different question entirely.
