What happened
TechCrunch reports that OpenAI is integrating ChatGPT Health with Epic’s electronic health record system so clinicians can import and question patient data. The integration provides read-only access to appointment notes, laboratory results, medications and specialist documentation, while some deployments place ChatGPT inside existing EHR workflows. OpenAI said the system does not write back to patient records.
TechCrunch reports that OpenAI said ChatGPT Health is being integrated with Epic’s electronic health record system, which the publication says holds data for more than 325 million patients. The stated purpose is to let clinicians import patient information into ChatGPT and ask questions about it. The reported use cases include summarizing appointment notes, laboratory results, medications and specialist documentation, reviewing a patient’s history, identifying changes over time and preparing for future appointments. These capabilities make the AI system a record-review and synthesis layer over existing clinical data, rather than a replacement for the underlying record system.
According to TechCrunch, some deployments will place ChatGPT directly inside EHR workflows. In those settings, clinicians may use ChatGPT for pre-visit review and for building clinical timelines without leaving a patient chart. OpenAI said the integration is read-only: ChatGPT can access health-record information, but the AI does not write anything back into the record. The source does not identify the participating health systems, provide a deployment schedule, describe the technical connection between Epic and ChatGPT, or say whether Epic independently confirmed the integration.
TechCrunch also reports that OpenAI is adding a Healthcare Public Data plug-in. OpenAI said the plug-in can retrieve information from ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed, helping healthcare workers synthesize trial-eligibility criteria, medication identifiers, coverage-policy versions and provider records. OpenAI further said organizations with a Business Associate Agreement will be able to use ChatGPT Work, Codex, apps and connectors in their workspaces for compliant workflows. The article does not establish which of these functions are available to which customers or what additional controls apply to each data source.
Source details: techcrunch.com ↗
Why it matters
This is a significant move toward embedding a general-purpose AI assistant in clinical information workflows. It could reduce the time clinicians spend reviewing records and assembling timelines, but the source provides no independent validation of clinical benefits, accuracy in routine use or safety beyond OpenAI’s own reported evaluation.
The reported integration matters because it moves AI assistance closer to the point where clinicians already make decisions: the patient chart. Record review can involve fragmented notes, test results, medication lists and specialist documentation. A tool that organizes those materials could make it easier to identify changes and prepare for visits. However, the source contains no independent evidence that the integration saves time, improves care or performs reliably across different specialties, patient populations or record formats. Those questions remain open even if the product works as described.
The read-only design limits one category of risk because the model is not reported to alter the official medical record. It does not eliminate the possibility of harm. A summary can omit a relevant fact, misinterpret a medication or present an uncertain inference too confidently, and a clinician could act on that output. The public-data plug-in adds another layer of complexity because information about trials, medications, coverage and providers may change or require careful interpretation. The source does not explain how the system displays citations, uncertainty, source dates or conflicts between records.
TechCrunch reports that OpenAI collected more than 4,300 physician responses across 27 clinical use cases and found 99.1% of responses safe. That is an OpenAI-reported assessment, not an independent clinical study, and the article does not provide the underlying test design, definition of “safe,” distribution of errors or comparison with ordinary clinical workflows. The report also notes lawsuits alleging that ChatGPT provided harmful medical advice, including a Florida pastor’s allegation of a near-fatal recommendation and a May lawsuit by family members who blamed ChatGPT for wrongful dosage advice. These are allegations, not findings established by the source, but they illustrate why a small number of unsafe answers can matter in healthcare.
What to watch next
Key unknowns include which health systems will receive the integration, when it will become broadly available, how patient consent and access controls will operate, and whether Epic independently confirms the arrangement. Organizations will also need to assess whether read-only AI summaries can still create clinical risk through omissions, misinterpretation or unsafe recommendations.
The first practical question is scope. The source does not say which hospitals, clinics or clinician groups are participating, whether access is limited to the United States, how quickly the integration will roll out or whether all listed functions will be available at launch. It also does not disclose pricing, eligibility requirements, administrative controls or whether clinicians need separate ChatGPT accounts. Those details will determine whether this is a limited enterprise pilot or a broadly accessible clinical product.
Data governance will be central. Organizations will need clear answers about authentication, authorization, audit logs, retention, patient consent, data separation and the handling of information returned by public-data sources. OpenAI’s reference to Business Associate Agreements indicates a compliance framework for some organizations, but the source does not describe the agreement’s operational safeguards or explain which features and connectors it covers. The report also does not say how the system handles records containing conflicting, incomplete or especially sensitive information.
Evaluation should extend beyond model responses in controlled review tasks. Health systems and regulators will likely need evidence from real workflows showing whether clinicians notice omissions, verify summaries and avoid treating generated timelines as authoritative. They may also need specialty-specific testing for medication review, handoffs and pre-visit preparation, along with incident reporting and clear human-review requirements. OpenAI maintains that AI is not suitable for diagnosis or treatment, according to TechCrunch; how that boundary is communicated and enforced inside clinical workflows will be important as the integration develops.