ChatGPT Health Explained: Features, Privacy, Availability, and Who It’s For

ChatGPT Health is a context layer, not an AI doctor

ChatGPT already answered health questions. ChatGPT Health changes the input: instead of making users retype a medication list, upload the same lab report, or describe a month of sleep data from memory, it can draw on health information they deliberately connect.

That is useful. It is also easy to misunderstand.

OpenAI launched the new experience on July 23, 2026. The company says more than 300 million people ask ChatGPT health-related questions each week. That scale explains the product. It does not prove that an answer is correct for any particular person.

OpenAI’s own Health help page puts the boundary plainly: the feature is designed to support medical care, not replace it, and is not intended for diagnosis or treatment. People with urgent symptoms should seek immediate professional help.

The right mental model is therefore modest: ChatGPT Health can make scattered information easier to inspect and discuss. Medical judgment stays with qualified professionals.

What actually changes when health data is connected

Ordinary ChatGPT can explain a term or help draft a list of questions without any connection at all. ChatGPT Health adds persistent, permissioned context from supported sources. Once data is synced, ChatGPT may use it in a relevant conversation after the user approves access. Adding @Health to a prompt explicitly requests that context.

OpenAI says more than 70% of health-related conversations among early users happened outside the original dedicated Health space. The current design reflects that behavior: the Health tab manages accounts, records, trends, and past health conversations, while approved health context can help personalize other relevant chats.

ChatGPT can read this connected information, but it cannot write changes back to Apple Health or a provider’s medical record. Marking a medication as no longer current inside Health helps ChatGPT understand the conversation; it does not update the pharmacy, clinic, or official chart.

That one-way boundary matters. A polished summary may be easier to read than the original record, but the original remains the source to check before making a decision.

The jobs that fit — and the ones that do not

The most defensible uses are information tasks. They have a clear human review step and do not ask the model to assume clinical authority.

OpenAI has put substantial work into health evaluation. In a June 2026 report, the company said more than 260 physicians across 60 countries, 49 languages, and 26 specialties had helped review over 700,000 model responses. It also reported that the share of production health responses with at least one flagged factuality issue fell 71% over two months. These are encouraging company-reported evaluation results, not a guarantee for the next answer.

Independent evidence gives the caution side of the picture. A 2026 physician-led study in npj Digital Medicine evaluated 888 responses to 222 patient questions from four then-public chatbots. Depending on the model, 21.6% to 43.2% of responses were rated problematic and 5% to 13% unsafe. GPT-4o was among the older models tested; the study did not evaluate GPT-5.5, GPT-5.6, or ChatGPT Health. It therefore cannot supply a current ChatGPT Health error rate. It does show why fluent medical prose still needs verification. See the study record and methodology.

The practical rule is simple: let the system reduce reading and preparation work, not remove the person who must make the medical judgment.

Privacy is more than the sentence “we do not train on it”

OpenAI says connected Apple Health information, medical records, and conversations that use them are not used to train its foundation models or target ads, regardless of the main ChatGPT training setting. It also says ChatGPT conversations are encrypted in transit and at rest, with additional encryption protections for connected Health information.

Those are material protections. Four separate questions still need answers before “private” means anything useful:

1. What can the product read? The user chooses connected accounts and Apple Health categories. By default, ChatGPT asks before using connected Health data in a response; the user can approve once or switch to always allow.
2. Who may process the data? OpenAI’s Health Privacy Notice says a limited number of authorized personnel and trusted service providers may access Health data for model-safety improvement unless the user opts out. Vendors may also process data for hosting, cloud, support, and safety functions under OpenAI’s instructions.
3. What happens after disconnection? Synced data from a removed source is deleted from OpenAI’s systems within 30 days. Information already copied into chat history remains until the relevant conversations are deleted.
4. Which law applies? ChatGPT Health is a consumer wellness product. OpenAI states that it is not intended for clinical or covered-entity use and does not offer a Business Associate Agreement, or BAA.

That last point is often blurred. HIPAA does not automatically cover every piece of health-related data. U.S. Department of Health and Human Services guidance explains that when a person directs a covered provider to send information to an app that is neither a covered entity nor a business associate, the received data is generally no longer protected by the HIPAA Rules. See the HHS guidance on health apps and APIs.

Other consumer-protection and state health-data rules may still apply. But they are not a substitute for choosing permissions carefully.

Apple likewise advises users to review a third-party app’s privacy policy before granting Health access and notes that users can change shared categories at any time. The FTC’s consumer guidance adds a broader test: compare privacy promises, inspect default settings, keep the app updated, and decide whether the service is worth the consequence of a possible disclosure.

Encryption reduces certain risks. Permission prompts reduce accidental use. Neither makes sensitive data risk-free.

How to connect ChatGPT Health with minimum permissions

Health works without any connected source, so connection should be a decision, not a setup reflex.

1. Confirm eligibility. As of July 24, 2026, Health is gradually rolling out to logged-in U.S. users aged 18 or older on Free, Go, Plus, and Pro plans. It works on web and iOS. Apple Health syncing requires an iPhone and ChatGPT iOS version 1.2026.188 or later.
2. Open Health from the sidebar. It may sit under the “More” menu. Select “Connect Health” or “Get started.”
3. Choose one source first. Connect Apple Health, a supported provider portal, One Medical, or Function Health only when its data answers a specific need.
4. Limit Apple Health categories. The iOS permission sheet lets the user choose categories. A sleep question does not require every available health field.
5. Wait for sync and audit the result. Large histories may take longer than a few minutes. Review current medications, conditions, and family history. Mark stale items and compare important facts with the original record.
6. Keep per-use permission enabled. The default prompt offers one-time approval or “always allow.” Per-use approval makes data use visible. Permissions can be changed later under Settings → Plugins → Health.
7. Ask for an inspectable output. Request a timeline, a list of changes, questions to ask a clinician, and the record entries used. Avoid asking for a diagnosis.
8. Verify before acting. Check numbers, dates, medication names, units, and source records. Bring the summary to a clinician when the decision matters.

To leave, remove connected accounts under Health → Accounts and disconnect Apple Health from Health’s menu. Then separately delete health-related chats or memories that contain information to be removed. Disconnecting a source stops future access; it is not a universal erase button.

For an extra account-security layer, enable multi-factor authentication. Temporary Chat or disabled memory can also reduce what persists from a particular conversation, but neither replaces source-level permissions and deletion controls.

Who should use ChatGPT Health now

ChatGPT Health has a clear early audience: an adult U.S. user with scattered records or Apple Health data who wants help reading trends, organizing history, or preparing for an appointment—and who is willing to manage permissions rather than press “allow all.”

It is less compelling when the health question is general. In that case, ordinary ChatGPT can answer without a data connection. It is also a poor fit for anyone expecting diagnosis, treatment selection, emergency triage, or automatic correction of official medical records.

People outside the U.S. should wait; the current rollout does not support them. Clinicians, insurers, and other HIPAA covered entities should not route protected clinical work through the consumer product because it does not offer a BAA. Voice mode and Codex do not currently support connected Health data either.

The final verdict is conditional. Connect it when longitudinal personal context will save meaningful preparation work, and only after narrowing the permissions. Skip the connection when the task can be done with a general question or a redacted document.

ChatGPT Health does not make health decisions effortless. Its best contribution is smaller and more credible: it can turn a pile of personal data into better questions. For a wider comparison of the assistants around it, see ChatGPT vs Claude vs Gemini, browse the AI chatbot guides, or explore AI chatbot tools.