ChatGPT can now draw on connected Apple Health information and, where supported, medical records from US providers during ordinary conversations. The feature is available to logged-in US adults across free and paid tiers on the web and iOS.
Putting years of clinical and activity data inside a familiar chatbot could help people understand fragmented records. It also makes stale information, permissions and conversational context part of a high-stakes product surface.
Health context now follows the conversation
Users can connect Apple Health, hospital records, One Medical or Function Health. ChatGPT may then use medications, laboratory results, visits, sleep and activity data outside the dedicated Health section—for example, when discussing a meal or suggesting exercise after an injury.
OpenAI changed the design after finding that more than 70% of health conversations in an early test occurred outside the separate Health area. The Health tab remains a management hub for connections, trends, prompts and previous chats, but personal context can appear wherever permission has been granted.
Early users described practical benefits: reconstructing a medical timeline, translating clinical language, comparing laboratory values over time and replacing manual spreadsheet uploads. A nurse reported discovering an unexpected chart entry that she could raise with her provider.
That example captures the useful boundary. The assistant can surface and explain information; it should not turn a record anomaly into a diagnosis. OpenAI positions Health as support rather than medical care and tells users to confirm important conclusions with a professional.
The potential audience is enormous. OpenAI says more than 300 million people make health-related ChatGPT queries each week. Even if the figure is not independently verified, the product’s existing reach gives it a distribution advantage that specialist health platforms rarely possess.
Permissions protect access but do not ensure accuracy
ChatGPT asks before using connected health data in a conversation unless the user enables an “always allow” setting. Connected records and chats drawing on them are excluded from foundation-model training and advertising, regardless of broader account preferences.
Disconnecting a source triggers deletion from OpenAI’s systems within 30 days. Information already copied into chat history remains until those conversations are deleted separately. Memories may be created from health discussions but not directly from raw records; Temporary Chat or disabled memory offers additional control.
Actions involving other plugins create a sensitive edge. Sending a training plan based on health metrics to a third party could reveal personal information. OpenAI says extra checks and explicit confirmation may apply and that it red-teams such scenarios, but it has not published failure rates.
Permission does not solve data quality. A discontinued medication may remain on a chart, wearable apps may omit proprietary measures and several sources can contradict one another. The product tells users to flag changes and check original records, placing part of reconciliation on the person least likely to understand the discrepancy.
Useful safeguards would identify source and timestamp beside every health claim, distinguish patient-reported from clinician-entered information and highlight conflicts instead of silently choosing one value.
Model evaluations cannot reproduce every personal record
OpenAI attributes the release partly to GPT-5.5 Instant and GPT-5.6 Sol. It says hundreds of physicians created scenarios and rubrics covering accuracy, safety, communication, context, completeness and escalation. GPT-5.6 models reportedly outperformed GPT-5.5 on the internal HealthBench Professional evaluation.
Physicians also tested the live product with connected data before release, but detailed methodology and results are unpublished. Controlled scenarios cannot represent every combination of chronic disease, outdated records, missing measurements and ambiguous notes.
Performance should be tracked after launch through harmful omissions, incorrect urgency, source conflicts, user corrections and successful escalation—not only response quality scores. People need a simple way to report a wrong medical interpretation and prevent it from influencing later chats.
Connected Health can make records more legible and help users prepare better questions for clinicians. Its central risk is not necessarily an obviously false diagnosis. It is a persuasive answer built from incomplete or obsolete context and delivered inside a conversation that feels personal.
The feature will earn trust when it makes uncertainty and provenance as visible as its summaries. A health assistant should help people navigate information without encouraging them to mistake convenient access for a complete clinical picture.