ChatGPT Health connects medical records and Apple Health in the United States

OpenAI has begun rolling out Health in ChatGPT to logged-in users age 18 and over in the United States, on the web and iOS. The feature lets people connect Apple Health, supported medical records, One Medical, or Function Health so that the assistant can, with permission, use that context when explaining changes in lab results, preparing for an appointment, or relating sleep and activity. It is not available in Codex or outside this initial rollout.

What matters is not only that a chat can summarize a lab result. The earlier version of Health lived in a separate space; OpenAI now allows connected information to be used in ordinary conversations when relevant. The company says that more than 70% of health conversations among early users took place outside that dedicated space. That responds sensibly to a real habit—an injury question may arise while planning the weekend—but it makes the boundary between a general conversation and one that uses health data much more consequential.

Three permissions, not one

The announcement describes an architecture with several decisions that should not be collapsed into one. By default, ChatGPT asks permission before using connected medical records or Apple Health in a response; a person can grant it once or allow continuing use. They can disconnect a source, and OpenAI says data synced from it is deleted from its systems within 30 days. But information already brought into a chat remains in the history until that conversation is deleted. Memories may also be created from a Health conversation—though not directly from connected records—unless that feature is turned off or a temporary chat is used.

Those distinctions are not fine print. “Not using connected records to train foundation models or target ads,” as OpenAI promises, answers a specific question about that data category. It does not on its own answer the others: which context is authorized in each chat, what remains written there, what memory is created, and what might be shared if another connector is asked to act. The company says it applies additional checks before actions that could disclose Health information, but good design here cannot rely on a person remembering a single choice made during setup.

There is a clear human benefit if those controls are understandable. Medical histories are often spread across portals, notes, apps, and devices; turning them into a readable timeline can help someone ask a better question of the professional who knows their case. There is an equally clear limit: the system can work from incomplete or outdated information, and a fluent explanation neither verifies that a medication is still current nor determines what care someone needs. OpenAI itself says ChatGPT can make mistakes and does not replace the judgment of qualified professionals.

The company links this launch to evaluations designed with physicians, including rubrics for accuracy, communication, context, and appropriate escalation. That is a useful signal, but evaluating a response does not solve the new problem created by the integration: managing sensitive data over time and across features. The most revealing test will not be a summarization demo. It will be whether a person can answer four questions without ambiguity: what data was connected, when it was used, where it is still stored, and how to stop using it.

In health, an interface that makes it easier to ask must make it just as easy to narrow, correct, and withdraw context. That symmetry of control is part of product quality, not an added privacy feature.

Sources

• OpenAI, Launching Health in ChatGPT, July 23, 2026.
• OpenAI, Introducing ChatGPT Health, updated July 23, 2026.
• OpenAI, Improving health intelligence in ChatGPT, June 18, 2026.