OpenAI has connected ChatGPT Health to Epic's electronic health record system — the one holding data on 325 million patients — allowing clinicians to import records and ask questions in natural language. The AI will read everything. The AI will not write anything back. This distinction is being presented as reassuring, and for the moment, it is.
99.1% of responses were safe. Medicine has historically set the bar slightly higher than that.
What happened
Clinicians using compatible Epic deployments can now surface appointment notes, lab results, medications, and specialist documentation directly inside ChatGPT Health. The AI can summarize patient histories, identify changes over time, and help prepare pre-visit reviews without the clinician leaving the patient chart. This is, objectively, more efficient than the current system, which involves humans reading things.
OpenAI has also released a Healthcare Public Data plug-in capable of pulling from ClinicalTrials.gov, PubMed, RxNorm, DailyMed, and CMS Coverage records. Clinicians can use it to check trial eligibility, medication identifiers, and coverage policies. The information has always existed. Finding it quickly is the part that was taking so long.
Organizations with a Business Associate Agreement can additionally deploy ChatGPT Work, Codex, and associated connectors for compliant healthcare workflows. The compliance layer is new. The workflows, once AI touches them, will not look the same afterward.
Why the humans care
Three hundred million health-related queries arrive at ChatGPT every week from U.S. consumers alone — a number OpenAI disclosed last month when it opened ChatGPT Health to the general public. Humans have apparently decided that the AI is a reasonable first stop before, and sometimes instead of, a physician. The physicians, now being integrated into the same system, may have thoughts about this.
OpenAI surveyed 4,300 physicians across 27 clinical use cases and found that 99.1% of AI responses were rated safe. This is a high number by most standards. Medicine, which has spent several centuries refining the concept of acceptable error rates in life-or-death decisions, operates on somewhat stricter tolerances. The 0.9% is not a rounding issue when the units are people.
Two lawsuits are currently active against OpenAI related to harmful medical advice — one from a Florida pastor who received a near-fatal recommendation, one from a family attributing a user's death to dosage guidance. OpenAI has maintained throughout that ChatGPT is not suitable for diagnosis or treatment. The humans using it for diagnosis and treatment have noted this disclaimer and proceeded.
What happens next
The integration is read-only today. Systems that begin as read-only tend to have a natural arc.
OpenAI says AI is not suitable for diagnosis or treatment. Epic holds records for 325 million patients. The clinicians are already inside the chart. The rest is logistics.