For years, healthcare artificial intelligence has existed in an awkward position. The technology has become increasingly capable, yet much of the information clinicians actually need remains buried inside electronic health records. An AI model may know an extraordinary amount about medicine, but if it cannot efficiently understand what happened to the patient sitting in front of you, its usefulness inside a real clinical workflow remains limited.
OpenAI's integration of ChatGPT Health with Epic could begin changing that. The integration will allow clinicians to import patient information from Epic and use ChatGPT to retrieve, organize, and summarize information including appointment notes, laboratory results, medications, specialist documentation, and previous clinical history. In certain deployments, clinicians will also be able to access ChatGPT directly within the EHR workflow without leaving the patient's chart.
That may sound less exciting than announcing a dramatically more powerful AI model. As a physician, I think it could be much more important.
Modern medicine has created an extraordinary amount of patient data. A medically complex patient may accumulate years of primary care notes, emergency visits, hospitalizations, imaging reports, medication changes, laboratory results, procedures, specialist consultations, and discharge summaries. Technically, clinicians have access to all of it. Practically, finding the right information at the right moment can be incredibly difficult.
This is one of the most immediate problems healthcare AI could help solve. Imagine preparing for an appointment and asking ChatGPT to summarize the patient's major clinical events over the past year. A physician could potentially ask what medications changed after the last hospitalization, whether kidney function has been progressively worsening, what cardiology recommended six months ago, or how frequently the patient has been admitted for heart failure. Instead of manually reconstructing the story from dozens of documents, AI could create a starting point.
The value here is not necessarily that AI discovers something no physician could discover. It is that the information already exists, but extracting it requires time clinicians increasingly do not have.
OpenAI says the Epic integration will have read-only access. ChatGPT can retrieve and synthesize information, but it does not write information back into the patient's health record through this integration. That is an important boundary. There is a major difference between allowing AI to summarize a medication history and allowing an autonomous system to change medications, enter orders, or modify clinical documentation.
Read-only access does not eliminate risk. AI could still omit an important detail, confuse historical information with current information, misunderstand contradictory documentation, or create a summary that sounds more certain than the underlying record supports. But keeping the technology primarily within an information retrieval and synthesis role creates a clearer place for human verification.
Healthcare AI conversations often jump immediately to diagnosis. Can AI diagnose cancer? Can it outperform radiologists? Can it interpret an ECG better than a physician? Those questions matter, but they can distract us from much more immediate opportunities.
A significant amount of clinical work happens before a diagnosis or treatment decision is ever made. Clinicians gather information, reconstruct timelines, compare previous results, review medications, read consultant recommendations, and determine what has changed. For a patient with years of fragmented documentation, this can consume a meaningful portion of an encounter.
If AI can reliably reduce that burden, it does not need to diagnose the patient to create enormous value.
OpenAI is extending this idea beyond individual patient records with a Healthcare Public Data plug-in that can retrieve information from sources including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed. That potentially creates an even more interesting workflow. Patient-specific information traditionally sits inside the EHR, while medical literature, drug information, clinical trial criteria, and coverage policies exist across separate databases. AI could increasingly become the interface connecting those worlds.
A clinician might eventually review a patient's history, investigate relevant medical literature, identify potential clinical trials, examine medication information, and review coverage policies through the same conversational interface. Anyone who has practiced modern medicine understands how significant reducing that digital navigation could be.
OpenAI says it collected more than 4,300 physician responses across 27 clinical use cases, including pre-visit review, clinical timelines, medication review, and handoff summaries, and found that 99.1% of responses were safe.
That sounds reassuring, but healthcare requires us to look at percentages differently. At scale, even a small failure rate can become meaningful. If millions of AI-assisted clinical interactions eventually occur, a fraction of a percent can represent a substantial number of potentially problematic outputs.
More importantly, "safe" is not necessarily the same as clinically reliable. We need to understand how these systems fail. Does the AI occasionally omit a critical medication? Can it distinguish copied-forward documentation from a genuine new diagnosis? What happens when two specialists document conflicting recommendations? Does it recognize uncertainty? Can clinicians easily trace every important statement in an AI-generated summary back to its original source?
These questions matter because an AI system does not need to independently diagnose a patient to influence a clinical decision. If it summarizes the wrong information, misses an important result, or incorrectly characterizes the patient's history, that output can shape what the clinician thinks next.
Recent allegations involving harmful health-related ChatGPT recommendations also reinforce why the distinction between assistance and clinical authority matters. OpenAI continues to state that AI is not appropriate for independently diagnosing or treating patients. Bringing AI deeper into professional healthcare workflows makes strong guardrails, traceability, monitoring, governance, and human accountability even more important.
Perhaps the biggest significance of the Epic announcement is what it tells us about the next phase of healthcare AI.
The first phase was largely about capability. Companies wanted to demonstrate that their models could answer medical questions, interpret complex information, summarize documents, and perform well on examinations and benchmarks.
The next phase is about integration.
Healthcare professionals do not need another disconnected AI application requiring another login, another browser tab, and another place to copy and paste patient information. They need useful intelligence where the work already happens. In much of American healthcare, that means inside the EHR.
This is why workflow integration may become one of the most important competitive advantages in clinical AI. The companies that successfully connect AI to clinical information, permissions, security requirements, existing infrastructure, and actual physician workflows may ultimately create more value than companies whose standalone models perform marginally better on benchmarks.
The best clinical AI may not be the model that produces the most impressive demonstration. It may be the system that saves a physician 15 minutes before a complicated appointment, finds the important specialist recommendation buried in a year-old note, or reconstructs a patient's clinical history without requiring the physician to open 30 different documents.
That is much less dramatic than an "AI doctor."
It is also much closer to what clinicians actually need.
There is enormous potential in giving clinicians better tools for navigating patient information. Imagine opening tomorrow's schedule and already having concise pre-visit summaries available for medically complex patients. Imagine being able to identify meaningful laboratory trends across several years, reconstruct previous hospitalizations, review major medication changes, and surface relevant specialist recommendations in seconds.
That could give clinicians something healthcare technology has repeatedly taken away from them: time.
But the goal should not be to make physicians passive consumers of AI-generated conclusions. The safest and most useful workflow remains one in which AI retrieves information, organizes it, summarizes it, and presents it to a clinician who can verify the source and make the final decision.
In other words:
AI retrieves and synthesizes. The clinician verifies and decides.
That distinction becomes more important as AI gets better, not less important. Highly convincing AI output can create automation bias, where users become increasingly inclined to accept recommendations because the system has historically performed well. Healthcare cannot afford to confuse fluency with correctness.
The ChatGPT Health and Epic integration therefore represents something bigger than another product feature. AI is moving closer to the information environment where clinicians actually make decisions. If these systems can reliably reduce information overload, reconstruct clinical histories, surface relevant evidence, and save meaningful time without creating new layers of risk, the impact could be enormous.
But once AI enters the patient chart, the standard changes.
Accuracy is no longer simply a product metric. Reliability is no longer simply a software issue. And a wrong answer is no longer merely an annoying chatbot response.
It becomes a clinical responsibility.
For years, some of the biggest names in artificial intelligence have made extraordinary predictions about what AI could do for medicine. Faster drug discovery. Personalized trea...
