ChatGPT Health can read Apple Health: what to know before connecting your data

OpenAI is rolling out Health in ChatGPT in the United States, allowing the assistant to use Apple Health information and supported medical records with permission. The stated goal is to compare results over time, prepare for an appointment or relate sleep and activity trends to daily routines without repeatedly importing the same documents.

Personal context can make answers more relevant, but it also brings some of a person's most sensitive data into a general-purpose assistant. OpenAI emphasizes encryption, permissions and a commitment not to use connected information to train foundation models or target ads. Those controls do not turn ChatGPT into a clinician, guarantee that synced records are current or make the feature available outside the U.S.

The short answer

People can still ask general health questions without connecting anything. The integration is optional and intended to add personal context, not to become a requirement for using ChatGPT.

What Health in ChatGPT can connect

Current sources include Apple Health, supported U.S. provider portals, One Medical and Function Health. Availability varies by provider and account. Electronic medical record connections are restricted to adults in the United States.

Apple Health can act as a bridge for information produced by an iPhone, Apple Watch or another app. Not every metric transfers. Proprietary scores, rankings or calculations may remain inside the original application. Missing information in ChatGPT does not prove that it is absent at the source.

After synchronization, a user can ask for comparisons across tests, a summary of changes since an appointment or a review of sleep and activity patterns. The Health area also provides access to recent information and lets users add conditions, medication or family history that may be missing.

ChatGPT cannot write back to Apple Health or a connected medical record. The announced access is read-only. That reduces the chance of an interpretation directly altering a prescription or source measurement, but a mistaken answer can still influence a human decision.

Synced records may be incomplete

A medical record is not always a perfect representation of the present. A discontinued medication may remain listed, a recent result may not have arrived and different providers may use different units or labels. OpenAI tells users to check important details against the original source.

Wearable data deserves the same caution. A watch estimates sleep, heart rate and energy expenditure through sensors and algorithms. These values do not all have the accuracy of a clinical device. A pattern may help prepare a question; it does not establish a condition by itself.

Before asking for an interpretation, verify the time period, unit, origin and context of each measurement. Exercise heart rate is not directly comparable with a resting value, and a laboratory may change its method or reference interval.

OpenAI's privacy commitments

OpenAI says connected medical records and Apple Health information, as well as conversations that use them, are not used to train its foundation models or target advertisements, regardless of the account's general model-training setting. Ordinary conversations that do not use this connected information continue to follow standard data controls.

ChatGPT conversations are encrypted at rest and in transit, with additional encryption controls announced for synchronized health data. These measures protect storage and transport. They do not remove risks from a compromised account, an unlocked device, a malicious browser extension or voluntary sharing with the wrong recipient.

By default, ChatGPT asks before using medical records or Apple Health to personalize an answer. Users can approve a single request or allow ongoing access, removing later prompts. Keeping per-request approval initially provides more visibility, particularly on an account used for unrelated topics.

Disconnecting does not automatically erase chats

A source can be removed from Health under Accounts. OpenAI says synced data from that source is deleted from its systems within 30 days. The connected source and conversation history have separate lifecycles, however.

If a medical value was quoted in a question or repeated in an answer, it can remain in that chat after disconnection. Those conversations must be deleted separately. Anyone running a short trial should understand both deletion paths before connecting an account.

Memory is another control. OpenAI says memories may be created from Health conversations to personalize future chats, but not directly from connected records or Apple Health data. Temporary Chat or disabled memory is more appropriate when a topic should not affect later interactions.

A preparation assistant, not a clinician

The most defensible use is turning a large amount of information into questions that can be checked. ChatGPT may rewrite a report in clearer language, explain differences between units, produce a timeline or prepare a list for a professional who knows the patient.

It should not independently decide to start, stop or change medication. A relationship between sleep and a symptom does not prove causation. A reassuring response is not a reason to ignore an urgent sign, and an alarming response is not a diagnosis.

For chest pain, difficulty breathing, sudden neurological symptoms, a severe allergic reaction or another possible emergency, contact emergency services or a qualified professional immediately. A model conversation must not delay care.

OpenAI explicitly states that Health supports rather than replaces medical care and is not intended for diagnosis or treatment. That boundary still matters when an answer appears precise and correctly references several personal measurements.

Useful lower-risk workflows

Several tasks can reduce preparation work without delegating clinical judgment:

1. summarize important changes since the previous appointment;
2. turn technical notes into understandable questions;
3. compare results while preserving dates, units and reference intervals;
4. identify missing information before a consultation;
5. prepare a symptom timeline for professional review;
6. explore an activity or sleep trend without claiming a cause;
7. restate post-appointment instructions, then check them against the original document.

A useful prompt asks the assistant to separate facts, hypotheses and unknowns. It can also require the exact values used, flag contradictions and produce questions instead of a definitive conclusion.

When not to connect anything

Someone who only wants to understand a medical term does not need to expose a complete history. A general question can be asked without a connected source or identifying clinical detail. Data minimization means providing only what is necessary for the immediate objective.

Do not analyze another person's records in your account without consent. OpenAI recommends a separate account or isolated space for someone else's information. Mixing patients can create context errors and expose data to people who should not see it.

U.S. healthcare organizations should not treat this consumer feature as a regulated clinical tool. OpenAI says Health in ChatGPT is not intended for covered-entity use and does not offer a Business Associate Agreement. Separate healthcare products exist for clinical settings.

Checklist before connecting

1. verify that the account and feature are officially eligible;
2. protect ChatGPT and Apple accounts with strong authentication;
3. update iOS, the app and browser;
4. review the exact sources and data categories being authorized;
5. keep per-request permission enabled at first;
6. verify medication, allergies, dates and units after syncing;
7. ask the assistant to separate facts, assumptions and missing data;
8. confirm any decision with a qualified professional;
9. use Temporary Chat when the topic should not feed memory;
10. remove conversations and connected sources separately after a trial.

What this changes

Health in ChatGPT reduces the effort required to assemble scattered information. Its strongest potential is not a dramatic answer but better preparation: finding a change, understanding a document and arriving at an appointment with more precise questions.

The cost of convenience is further concentration of personal data in an online service. The announced safeguards are substantial, but users still need to choose sources, control permissions, verify accuracy and remove information that is no longer useful. For people outside the United States, this remains a development to watch rather than a feature they can currently enable.