Summary
OpenAI’s ChatGPT for Healthcare integration with Epic puts multiple AI systems inside the clinical workflow—raising new questions about trust, verification, and patient safety. Here’s why health systems need to rethink AI training as clinicians learn which AI to trust, when to verify it, and how to recognize confidently wrong information.
Anne Hyland, VP of EHR Learning, Amplifire
Recently the AI-in-healthcare story got more complicated. OpenAI announced that ChatGPT for Healthcare now connects directly to Epic—letting authorized clinicians pull patient data into a chat window or use ChatGPT embedded right inside supported Epic workflows. At least one pilot health system is already using it (one of my favorite clients, actually), with others reportedly exploring it.
It’s a significant moment, and it’s worth thinking about the real impact—because it’s not just “another AI feature.” It’s a second AI system, built by a different company, now living inside the same chart as Epic’s own AI tools.
What does the integration really mean?
The integration supports two workflows: clinicians can pull authorized patient data—appointment notes, lab results, medications, specialist documentation—into ChatGPT to review a history or prepare for a visit, or they can access ChatGPT directly within supported Epic workflows without leaving the chart. The connection is explicitly read-only: ChatGPT can retrieve the record, but it cannot write anything back into it.
Alongside the Epic connection, OpenAI also shipped a Healthcare Public Data plugin linking ChatGPT to a handful of official sources—things like PubMed, ClinicalTrials.gov, CMS coverage data, etc.—so clinicians can pull in medical evidence and coverage information without leaving the workspace.
This lands on top of Epic’s own recent AI push—Art, Penny, Emmie, and Agent Factory—which means the chart a clinician opens tomorrow may have two independent AI layers in it: Epic’s native tools and now OpenAI’s, sitting alongside or inside the same workflow.
The new problem this introduces: Which AI do I trust for what?
This is the part that actually changes the training conversation, and it’s genuinely new. Up to now, “verify the AI” has meant checking one system’s output against your own clinical judgment. As I’ve written before, AI literacy increasingly depends on knowing when to trust an AI response, when to question it, and when to verify it. Once there are two AI systems drawing on the same underlying record, potentially summarizing, emphasizing, or framing the same information differently, clinicians face a question they likely never had to address before: not just “is this AI right?” but “which AI’s read on this patient do I trust, and why?”
That’s not a hypothetical edge case. Two models built by two different companies, trained differently and optimized for different things, will not always converge on the same emphasis from the same chart. A clinician moving quickly between an Epic-native suggestion and a ChatGPT-generated summary needs more than baseline AI skepticism. They need to understand what each tool is actually built to do well and where each one may miss something.
Is read-only risk-free?
Because the integration can’t write back into the record, it’s tempting to file this under “low stakes,” but I’d push back on that a little. A wrong or subtly incomplete analysis of real data, delivered fluently and confidently, can be more dangerous than no analysis at all—especially when it looks authoritative, feels like it “came from Epic,” and saves the clinician the step of checking the primary sources themselves.
This is the same Confidently Held Misinformation™ problem Amplifire has been talking about, just showing up in a new place. The risk was never really about the AI writing the wrong thing into the chart. It’s about a clinician acting on a wrong read of the data because the read arrived polished and plausible—or because the clinician over-trusted the analysis.
Is it standardized for everyone?
Health systems can control what data and workflows are exposed through the integration. Frontline staff won’t just need to learn “how ChatGPT works” in the abstract; they’ll need to know specifically what’s turned on, what data it can see, and what workflows it touches in their own environment. That’s a communication and training challenge distinct from the tool’s mechanics.
What does this mean for training?
The core shift from AI in the EHR still holds here: training content is moving away from rote navigation and toward oversight and verification. But this integration sharpens that shift into something more specific—and more teachable—than generic “AI literacy”:
- Know which AI system you’re looking at in a given moment and what data it drew from.
- Understand each tool’s intended use case and likely blind spots. Go beyond “AI can be wrong” to recognize what information a tool may be missing, when to check its analysis against the patient record, and how to catch an answer that sounds right but isn’t.
That’s a harder, more specific skill set than “learn the new feature.” It’s also the kind of gap that shows up in KLAS Arch Collaborative satisfaction data long before it shows up anywhere else. It starts as a quiet erosion of confidence, and no one traces it back to its source until it’s already shaping how clinicians use the tools in front of them.
The stakes are high. This is exactly where Amplifire does its best work: moments when confidence has to be matched by accuracy and there’s no time to slow down.