What ChatGPT in Epic Actually Changes
Alone, the models pick the right condition 94.9% of the time and the right next step 56.3%. Epic just put one in front of every authorized clinician.
What ChatGPT in Epic Actually Changes
OpenAI now connects ChatGPT to Epic, pulling a patient's labs, medications, past visits and specialist notes into the chat window for any clinician already authorized to see them.
Why it matters: The only controlled study of LLMs paired with human users found the humans made the pairing worse, not better. Epic just made that pairing standard equipment.
By the numbers (Bean et al., Nature Medicine, February 2026, 1,298 participants):
94.9% - LLMs alone correctly identified the underlying condition.
56.3% - LLMs alone recommended the correct level of care.
Under 34.5% - condition identification when real people used those same models.
No better than a web search - how the assisted group performed against the control group.
Yes, but: Those participants were laypeople, not clinicians. The finding does not transfer automatically to a physician inside an EHR. That is the point. No published study measures the configuration Epic just shipped.
What OpenAI is saying: Physicians rated 99.1% of responses safe across 27 clinical use cases and 4,363 ratings (OpenAI, January 2026). That is a vendor evaluating its own product against its own benchmark, and it measures safety, not whether the recommended next step is correct.
Between the lines: Access controls are not clinical controls. Inherited Epic permissions bound what the model can read. They say nothing about whether the summary a clinician acts on is right, and nothing about who owns the outcome when it is not.
Three questions a health system should answer this quarter:
Does the accuracy evidence cover clinicians using the tool, or only the model in isolation?
Who owns the outcome when a clinician acts on a wrong summary?
Does the HIPAA analysis cover model outputs, or only the records at rest?
What to watch: Whether OCR addresses inference outputs, and whether anyone publishes a clinician-in-the-loop trial before this reaches scale.
The bottom line: 94.9% at diagnosis and 56.3% at disposition is not one system performing unevenly. It is two capabilities, and health systems are procuring them as one.
Works Cited
Bean, Andrew M., et al. "Reliability of LLMs as Medical Assistants for the General Public: A Randomized Preregistered Study." Nature Medicine, vol. 32, Feb. 2026, pp. 609–615, nature.com/articles/s41591-025-04074-y.
"ChatGPT Connects Health Records and Healthcare Sources." OpenAI, Jan. 2026, openai.com/index/chatgpt-connects-health-records-and-healthcare-sources.
Related briefs
- AI as a Financial Copilot: Fast. Smart. Not Liable. — AI financial copilots are transforming how people manage money. But a copilot that carries no fiduciary obligation is not a financial advisor. Most users do not know the difference. Most institutions have not made it clear.
- Financial Technology, Fast Money, Slow Governance. — Convenience got customers in the door. Efficiency kept them there. Trust is what determines whether they stay when something goes wrong. Most fintech AI deployments have not built the infrastructure to protect that trust.
- Who Controls the Chips Controls the AI — DeepSeek is building its own chip. OpenAI launched its first. The AI model race gets the headlines. The chip race determines who wins it. Here is what that means for enterprise buyers.