ChatGPT Health launches: connected data raises both usefulness and risk

On July 23, OpenAI began rolling out ChatGPT Health in the United States. It can connect Apple Health, supported hospital records, One Medical, and Function Health so questions can be answered with the user's own context.

OpenAI does not present it as a diagnostic product and says it supports rather than replaces professional care. The shift is still substantial: lab results, medications, visit notes, sleep, and activity can sit inside the same conversation.

Why move health context into general chat?

In early testing, OpenAI says more than 70% of health-related conversations among people with access happened outside the dedicated Health space. Allergies during meal planning or a recent injury during activity planning naturally appeared in ordinary chats.

Users can allow ChatGPT to draw on connected health information in other conversations. The default is to ask permission each time. A user can approve once or switch to persistent access and change the setting later.

Who gets access?

For now it is limited to logged-in U.S. users aged 18 and older, rolling out on web and iOS across Free, Go, Plus, and Pro. Apple Health and supported U.S. medical records can be connected. Health is not currently available in Codex.

OpenAI says more than 300 million people ask ChatGPT health-related questions every week. That makes the feature strategically important, but the number is a company-reported metric rather than an independent measure.

Privacy commitments and limits

According to OpenAI, connected records, Apple Health data, and conversations that use them are not used to train foundation models or target ads. Synced data is deleted within 30 days after a source is disconnected, but health information already included in chat history can remain until those conversations are deleted.

Health data is not always current or complete. A discontinued medication may remain listed, and a wearable measurement may not mean the same thing as a clinical measurement. When a model combines those fragments fluently, the result can look more certain than the evidence allows.

Important decisions require returning to the original record, making uncertainty visible, and verifying with a qualified professional. A good summary does not prove its data is correct or current. Permission, provenance, freshness, deletion, and professional escalation have to be designed together.