ChatGPT for Healthcare Now Connects to Epic EHR and 9 Data Sources

OpenAI put ChatGPT inside the electronic health record. On September 1, 2026, OpenAI's Karan Singhal (who leads health work at the company) announced that ChatGPT for Healthcare now integrates with Epic — the EHR system used across a large share of U.S. hospitals — and added a nine-source Healthcare Public Data plugin covering PubMed, ClinicalTrials.gov, DailyMed, CMS data, openFDA, and RxNorm. This is a different product from ChatGPT Health's Apple Health integration, which is patient-facing; this is a clinician-facing, enterprise feature gated behind an existing Epic deployment.

TL;DR

What the Epic integration actually does

The core pitch, in Singhal's own framing: "healthcare moves at the speed of trust," and AI has to meet clinicians where they already work rather than asking them to open a separate tool. Concretely, the Epic integration supports two modes:

• EHR context in ChatGPT — an authorized clinician brings a patient's history, labs, medications, and specialist notes into a ChatGPT conversation to review timelines, spot changes, or prep for a visit.
• ChatGPT inside Epic — the same capability, but embedded directly in the clinician's existing Epic interface rather than requiring a context switch to a separate app.

Both modes are explicitly read-only. ChatGPT can summarize, cross-reference, and surface information from the record, but it does not write anything back into the patient's chart. That's a meaningful scope limit worth naming clearly: this is a documentation-and-review aid, not an autonomous charting or order-entry agent.

The Healthcare Public Data plugin: nine sources in one query

Separately, OpenAI shipped a Healthcare Public Data plugin that connects ChatGPT to a set of public clinical and regulatory databases — named sources include:

• PubMed — peer-reviewed medical literature
• ClinicalTrials.gov — active and historical clinical trial records
• DailyMed — official drug labeling data
• CMS data — Centers for Medicare & Medicaid Services datasets
• openFDA — FDA adverse-event and product data
• RxNorm — standardized drug nomenclature

OpenAI describes this as nine sources total; the remaining three weren't individually named in the announcement thread. The point of bundling them is straightforward: instead of a clinician or researcher manually querying six-plus separate government and literature databases with six different search interfaces, a single ChatGPT query can pull from all of them in one pass. This plugin is available more broadly than the Epic integration — it works on both ChatGPT for Healthcare and ChatGPT for Clinicians (OpenAI's free tier for physicians, nurse practitioners, and pharmacists).

The 99.1%-safe number: what it actually measures

OpenAI's headline safety claim is that physicians rated ChatGPT responses with connected EHR context across 27 clinical use cases — pre-visit review, clinical timelines, medication review, and handoff summaries among them — producing 4,363 individual ratings, of which 99.1% were rated safe.

Read that number for what it is, not more:

• It's a physician-graded evaluation on a defined set of use cases, not an independent third-party clinical trial or a regulatory safety certification.
• "Safe" is a binary rating category defined by OpenAI's evaluation design — it doesn't tell you the distribution of what the 0.9% unsafe responses looked like, how severe they were, or in which of the 27 use cases they clustered.
• A 99.1% safety rate across thousands of ratings still implies dozens of responses rated unsafe in this evaluation alone — worth knowing before treating the number as a blanket reassurance rather than a data point that supports "generally safe in tested conditions."

This is the same pattern worth watching across every major lab's health-AI safety claim: the number is real and worth citing, but it describes performance on the specific evaluation designed to test it, not a universal guarantee.

Why Epic specifically matters

Epic is the dominant EHR vendor across large U.S. hospital systems — a meaningful share of U.S. patient records live inside Epic deployments. An integration at that level isn't a niche partnership; it's OpenAI plugging into the system most large health systems already run their clinical documentation through, rather than asking hospitals to adopt a new, separate data layer. UCSF Health serving as the named pilot partner signals OpenAI is starting with a large academic medical center rather than a small practice, likely to stress-test the integration against high patient volume and case complexity before wider rollout.

What this means for builders outside healthcare

Even if you don't work in health tech, the pattern here is a useful template for any team connecting an LLM to an existing system of record:

1. Read-only first. Ship the "bring context in" capability before the "write actions out" capability — it's a much smaller blast radius if something goes wrong, and it's the natural first integration point for a system where mistakes carry real cost.
2. Publish an evaluation methodology, not just a headline number. OpenAI naming the use-case count (27), rating count (4,363), and rating category (safe/not safe) is more useful to an outside reader than the 99.1% figure alone — it lets you judge how much evaluation actually backs the claim.
3. Bundle related external data sources into one interface rather than making users query each one separately — this is the same principle behind Gemini's agentic video understanding fetching only the relevant signal instead of processing everything uniformly, applied here to structured clinical databases instead of video.

Honest limitations

• Enterprise-gated, not broadly available. This requires an existing Epic deployment and a workspace administrator's approval — most individual clinicians and smaller practices won't have access regardless of interest.
• Read-only means it doesn't close the loop. A clinician still has to manually act on anything ChatGPT surfaces; this is a research-and-review aid, not an automation layer for clinical workflows.
• The safety number is self-reported and self-evaluated. OpenAI designed the use cases, recruited the physician raters, and defined "safe" — useful as a first data point, not a substitute for independent, peer-reviewed clinical validation over time.

Bottom line

This is OpenAI moving ChatGPT from "a tool clinicians open separately" to "a layer inside the system they already use," with a deliberately narrow, read-only scope and a published (if self-graded) safety number to back the initial rollout. It's a sensible, cautious first step for a domain where mistakes are costly — the real test will be whether the 99.1% safety rate holds up once usage moves beyond the pilot's 27 defined use cases into the messier reality of everyday clinical practice.

Related on explainx.ai

• ChatGPT Health: Connect Apple Health and Medical Records Safely — the consumer-facing counterpart to this enterprise launch
• Claude Fable 5.1 and Mythos 5.1: Benchmarks, Pricing, and Safeguards — Anthropic's own life-sciences access program, launched the same week
• Grok 4.6's Biosecurity Evaluation — how a rival lab is publishing safety numbers for biology-adjacent AI work
• Gemini's Agentic Video Understanding — the same "connect to existing data instead of reprocessing everything" pattern applied elsewhere
• What Is MCP (Model Context Protocol)? — the connector pattern underlying most of these external-data integrations
• AI for Mental Health: Therapy Chatbots Guide — a related look at AI safety evaluation in a health-adjacent domain

Details in this post reflect OpenAI's official announcement and public reporting as of September 2, 2026. Enrollment requirements, supported Epic versions, and the full nine-source data list may change as the rollout expands.