OpenAI has announced that ChatGPT Health will connect to the electronic health record system built by Epic. Clinicians will be able to pull in clinical records from a platform holding data on more than 325 million patients and question ChatGPT about them on the spot. In some deployments, ChatGPT can be called up without leaving the patient chart. What the integration allows, however, is reading only. The AI does not write anything back into the record.

Asking Questions Without Leaving the Chart

The material clinicians can pull in covers what they already review day to day: appointment notes, laboratory results, current medications, and documentation from specialists. ChatGPT summarizes these together, traces a patient's course, and points out where things have changed. It can also be used to prepare for the next visit.

In certain deployment formats, ChatGPT can be invoked while the patient chart stays open. Pre-visit review can be completed in place, and a patient's history can be organized into a clinical timeline without switching windows. For many practices, moving back and forth between the EHR and a separate window has been exactly where consultation time was leaking away.

The Line Drawn at Read-Only

The second pillar of the design is permissions. The integration grants ChatGPT read-only access, and the AI cannot write back into clinical records.

That distinction looks minor but it settles the question of responsibility. A chart is both a record of care and evidence used later to review what was done. Introducing a system that appends to it automatically inevitably raises the question of who stands behind those entries. Limiting the AI to reading keeps the final act of editing the record in human hands.

A Plugin That Spans Public Databases

Alongside the integration, OpenAI is adding a plugin called Healthcare Public Data. It lets ChatGPT query public sources directly, including the clinical trial registry ClinicalTrials.gov, coverage information from CMS (the Centers for Medicare and Medicaid Services), the drug terminology standard RxNorm, the drug labeling database DailyMed, and the literature index PubMed.

The intended uses include sorting out trial eligibility criteria, checking medication identifiers, confirming which version of a coverage policy applies, and verifying provider records. All of these previously meant opening several databases and cross-checking them by hand, so the aim is clearly to compress research time.

Organizations With a BAA Also Get ChatGPT Work and Codex

OpenAI announced an expansion of scope at the same time. Organizations that have signed a Business Associate Agreement can now use ChatGPT Work, Codex, apps, and connectors inside their workspace as part of compliant workflows.

A BAA is the contract signed with outside vendors that handle protected health information under HIPAA, the US health data privacy law. Clearing that hurdle lets healthcare organizations bring general-purpose features that were previously hard to adopt into daily work while staying compliant.

How to Read the 99.1 Percent Figure

OpenAI says it gathered more than 4,300 evaluations of responses from physicians. The evaluations covered 27 clinical use cases, including pre-visit review, clinical timelines, medication review, and handoff summaries, and 99.1 percent of responses were judged safe.

The number looks high, but in medicine what matters is what sits in the remainder. Even a rate below 1 in 100 leaves no room for recovery when it touches dosing or a clinical decision. OpenAI itself has consistently held that AI is not suited to diagnosis or treatment.

Real litigation sits in the background. A few days before this rollout, a pastor in Florida sued OpenAI, alleging that ChatGPT gave him advice that could have cost him his life. In May, family members of a user filed suit over what they said was incorrect guidance on dosage. Consumer health features were opened to all US users last month, and health-related questions reportedly run at 300 million per week. The wider the scale, the higher the odds that an exceptional failure lands on someone.

Summary

The ChatGPT Health and Epic integration is shaped less as a way to push AI into clinical judgment than as a tool for shortening the time spent reading documents. Restricting it to read-only, splitting public database lookups into a separate plugin, and opening work features only to organizations with a BAA are all defensive choices. Even so, whether to take a summary at face value remains a call made at the point of care. What clinicians do with the time they save from reading charts is what will decide how this integration is judged.

※ The thumbnail image is an AI-generated illustration.