ChatGPT Health adds Epic integration for clinicians to import patient data
Sep 1, 2026, 10:00 AM · TechCrunch

OpenAI is wiring ChatGPT Health into Epic, which holds records for more than 325 million patients. The access is read-only. The liability is not.
Why it matters
Ivan Mehta reports that clinicians can import appointment notes, labs, medications, and specialist documentation from Epic into ChatGPT Health, then ask questions, spot changes, and prep visits. In some deployments ChatGPT sits inside the chart for pre-visit review and clinical timelines. OpenAI says the integration is read-only: the model does not write back to the record.
There is also a Healthcare Public Data plug-in that can pull ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed. Organizations with a Business Associate Agreement can use ChatGPT Work, Codex, apps, and connectors in compliant workflows. Last month ChatGPT for health opened to all U.S. consumers; OpenAI said people were asking 300 million health-related queries a week, while still saying the system is not for diagnosis or treatment.
From the desk
Read-only is the sentence OpenAI wants quoted. It is also the least interesting part of the story. The moment a clinician uses a chart-side summary to decide what to ask in the room, the model is in the care path whether or not it can click Save. Epic’s installed base is the distribution prize every health-AI vendor has been chasing. Getting a read pipe into 325 million records is how ChatGPT stops being a sidebar the doctor pastes into.
OpenAI cites more than 4,300 physician responses across 27 use cases — pre-visit review, timelines, medication review, handoff summaries — and says 99.1 percent of responses were safe. That is a marketing number dressed as a safety number. A 0.9 percent miss rate on clinical text is not a rounding error if the misses are dosage, contraindications, or missed red flags. Mehta notes the obvious: even a few unsafe answers can harm people.
The legal weather around this launch is not theoretical. A Florida pastor recently sued, alleging a near-fatal recommendation. In May, family members sued over alleged wrongful dosage advice. Shipping an Epic connector after those filings is either confidence or necessity. The likelier read is both: consumer health traffic is already at the 300-million-queries-a-week scale OpenAI cited last month, and the enterprise path is how you put a BAA and an audit log between that traffic and the next complaint.
Signal Desk’s read: this is the most consequential ChatGPT distribution move of the year, and the one most likely to be walked back by a hospital counsel after the first documented miss. Watch whether Epic customers treat it as a summarizer with a human in the loop, or as a junior resident that never writes a note.
Context
Epic is the default inpatient and ambulatory record in much of U.S. health care. Every major model lab has been trying to sit next to that chart without taking write permission. OpenAI’s consumer Health product and this clinician pipe are the same brand pointed at two different liability regimes.
Who feels it
- Clinicians
- A chart-side timeline is useful. Treat every summary as untrusted until you have opened the source notes. Read-only does not mean consequence-free.
- Health systems
- A BAA plus read-only access is the starting paperwork, not the risk review. Name who is liable when the summary is wrong and the doctor relied on it.
- OpenAI’s lawyers
- The 99.1 percent safety stat will be an exhibit. So will the pastor suit and the May dosage case. Expect discovery to ask how those numbers were scored.
What to watch
- Whether any large Epic customer enables in-chart ChatGPT beyond a quiet pilot.
- The first documented clinical miss tied to a Health-imported chart, and whether OpenAI or the hospital owns the statement.
- If write-back, order-entry, or note-drafting ever appears after the read-only pledge.
Companies: OpenAI