The Clinical Layer: OpenAI's Epic Integration and the Risk of Automated Error
AI-generated image · US National Wire
By embedding ChatGPT into EHR workflows, OpenAI is streamlining data synthesis—but internal safety metrics may not be enough to prevent clinical harm.
OpenAI is moving beyond the consumer interface to position itself as a functional layer within the clinical environment. As TechCrunch first reported, the company is integrating its ChatGPT Health tool with Epic’s electronic health record (EHR) system. Given that Epic manages data for more than 325 million patients, the scale of this integration is significant.
Under the new system, clinicians can import patient data—including laboratory results, medications, specialist documentation, and appointment notes—to generate summaries, identify changes in patient history, and prepare for upcoming visits. TechCrunch reports that in specific deployments, these tools will exist directly within EHR workflows, allowing providers to build clinical timelines and conduct pre-visit reviews without exiting the patient chart. To maintain a boundary between AI synthesis and the medical record, OpenAI specified that the integration is read-only; the AI cannot write data back into the health records.
Beyond the EHR, OpenAI is introducing a Healthcare Public Data plugin. As reported by TechCrunch, this tool allows healthcare workers to synthesize information from PubMed, ClinicalTrials.gov, RxNorm, CMS Coverage, and DailyMed to assist with provider records, medication identifiers, coverage policy versions, and trial eligibility criteria.
However, from an evidence-first perspective, the gap between "safe" outputs and clinical accuracy is where the danger lies. OpenAI claims to have surveyed physicians across 27 clinical use cases—such as handoff summaries and medication reviews—and found that 99.1% of the 4,300 responses were safe. While that number sounds impressive in a tech demo, in a clinical setting, a 0.9% failure rate is a liability.
***Opinion:*** *OpenAI is effectively automating the synthesis of medical data at scale, yet it continues to maintain that its AI is not suitable for diagnosis or treatment. When a tool is used to summarize a patient's entire history or review medications, the line between "summarization" and "clinical decision support" blurs. Without rigorous, peer-reviewed validation of these LLM-driven summaries, we risk automating medical hallucinations that could lead to catastrophic patient outcomes.*
The real-world consequences of AI inaccuracies are already surfacing in the courts. TechCrunch reports that a pastor based in Florida recently sued OpenAI, alleging the chatbot provided a near-fatal recommendation. Additionally, OpenAI faced a lawsuit in May from the family of a user who alleged the AI provided wrongful dosage advice.
With OpenAI reporting that U.S. consumers are already submitting 300 million health-related queries per week, the push into professional EHR workflows increases the stakes. For organizations with a Business Associate Agreement, OpenAI is also permitting the use of connectors, apps, Codex, and ChatGPT Work for compliant workflows. But compliance is not the same as clinical validity. Until the "unsafe" minority of responses is eliminated, these tools remain high-risk experiments in the middle of patient care.