ChatGPT Health: How to Connect Medical Records Safely

OpenAI opened Health in ChatGPT to eligible U.S. adults on July 23, 2026, giving people the option to connect Apple Health and supported medical records. That can make scattered lab results, medications, activity data, and visit notes easier to understand—but it also raises a practical question: how do you use ChatGPT Health safely without giving an AI more data or authority than necessary?

This guide explains what the feature does, how to configure it conservatively, which questions it is good at, and where a clinician still needs to take over.

What ChatGPT Health can access in July 2026

According to OpenAI’s launch announcement, Health is rolling out to logged-in users aged 18 or older in the United States on web and iOS. It is included across Free, Go, Plus, and Pro plans, but it is not available in Codex. A staged rollout means an eligible account may not see it immediately.

You can use ChatGPT for a health question without connecting any account. If you opt in, Health supports:

• Apple Health
• Medical records from supported U.S. hospital systems
• One Medical
• Function Health

Apple Health may act as a bridge for compatible wearables, fitness apps, and nutrition apps. The available metrics depend on what each app writes to Apple Health; proprietary readiness or recovery scores may not transfer.

Once authorized, ChatGPT can use relevant medications, conditions, lab results, visits, sleep, activity, and workout information. Typical uses include comparing a new result with prior tests, summarizing changes since an appointment, explaining a visit note in plain language, and preparing questions for a clinician.

Configure ChatGPT Health with minimum necessary access

The safest default is not “connect everything.” Start with the smallest data set that solves your immediate problem, confirm its accuracy, and expand only when the benefit is clear.

1. Open Health and connect one source

Choose Health from the ChatGPT sidebar or More menu, select Get started, and connect a single source. Syncing can take a few minutes. Do not connect a hospital portal and Apple Health simultaneously unless your question genuinely needs both clinical and activity data.

For example, a question about sleep consistency may require Apple Health but not your entire clinical history. A request to summarize recent bloodwork may need a medical-record connection but not location or workout history.

2. Review Apple Health permissions at the source

Apple says third-party apps must request access to read or write specific Health categories, and users remain in control of what is shared. Review Apple’s Health app privacy explanation before approving access, then grant only the categories needed for your goal.

Pay special attention to highly sensitive categories such as medications, reproductive health, mental wellbeing, and mobility. A useful rule is: if the data would not improve the answer you plan to ask for, leave it disconnected.

3. Check the imported record before asking for analysis

A synced record is not automatically complete or current. OpenAI specifically warns that an old medication may remain listed after you stop taking it. Review conditions and medications, remove stale entries, and add important missing context such as family history only when relevant.

For any consequential question, compare the imported item with its original source. Confirm:

• The test date and units
• The laboratory’s reference range
• Medication name, dose, and status
• Whether a diagnosis is active, historical, suspected, or ruled out
• Whether wearable data has gaps caused by charging or not wearing the device

This step prevents a polished summary from hiding a bad input.

4. Keep per-request permission prompts enabled

By default, ChatGPT asks before using connected Health information. Keep that setting instead of choosing permanent access. You can manage it under Settings > Plugins > Health and can explicitly request health context by adding @Health to a message.

Per-request approval creates a useful pause: does this restaurant question really need your medical record, or would stating one allergy be enough? Convenience is not always worth broad context sharing.

5. Decide how memory should behave

OpenAI says memories can be created from Health conversations, although memories are not created directly from connected medical records or Apple Health data. If you do not want a health discussion to shape future chats, turn memory off under Settings > Memory or start a Temporary Chat.

OpenAI’s current Data Controls FAQ says Temporary Chats do not appear in history, do not create memories, are not used to train models, and are deleted from OpenAI’s systems after 30 days, subject to the stated abuse-monitoring process.

What the privacy promises cover—and what they do not

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

Those are meaningful controls, but they do not eliminate every privacy consideration.

First, disconnected does not mean instantly erased. OpenAI says data synced from a disconnected Health source is deleted from its systems within 30 days. Information already written into conversation history remains until you delete those chats.

Second, do not assume every consumer health-data flow receives exactly the same legal treatment as a hospital portal. The U.S. Department of Health and Human Services explains that HIPAA applies to covered entities and business associates; an organization outside those definitions does not have to comply with the HIPAA Rules. Coverage depends on the organization’s role and arrangement, not merely on whether the data is medical.

Third, policies and products can change. A CBS News review published July 25, 2026 noted both the potential benefit of longitudinal analysis and expert concerns about sharing deeply personal data with a chatbot. Recheck the current policy and permission screen rather than relying on what you approved months earlier.

Use prompts that produce verifiable work

ChatGPT Health is most useful as an organizer, translator, and question-preparation assistant—not as an autonomous diagnostician. Ask it to show its work in a form you can compare with original records.

Good requests include:

• “Create a dated table of these lab results. Preserve the original units and reference ranges. Flag missing or incomparable values instead of guessing.”
• “Summarize what changed since my previous visit. Separate facts in the record from possible interpretations.”
• “Translate this visit note into plain English. List unfamiliar terms and questions I should ask the clinician.”
• “Compare my recorded sleep duration across the last four complete weeks. Identify missing nights and do not treat them as zero.”
• “Draft a one-page appointment brief with current medications, recent changes, and five questions. Ask me to verify every item before finalizing it.”

Avoid requests such as “diagnose me,” “tell me whether to stop this medication,” or “decide if I can ignore this symptom.” If a response suggests changing a dose, delaying urgent care, or replacing a prescribed plan, do not act on it without a qualified clinician.

A verification workflow for every important answer

Use this four-part check before relying on an AI-generated health summary:

1. Source: Can every claimed fact be traced to a dated record or a detail you supplied?
2. Separation: Does the answer clearly distinguish recorded facts, general information, and speculation?
3. Completeness: Did ChatGPT account for units, reference ranges, missing data, medications, and relevant dates?
4. Escalation: Does the question involve diagnosis, treatment, a medication change, worsening symptoms, pregnancy, a child, or an emergency? If so, move the decision to a clinician.

OpenAI evaluates health responses with physician-designed benchmarks. Its original HealthBench contains 5,000 realistic conversations and 48,562 rubric criteria created with 262 physicians who had practiced across 60 countries. That is evidence of serious evaluation work—not proof that any individual answer is correct. OpenAI itself says ChatGPT can still make mistakes and does not replace qualified medical care.

For more context on the difference between AI information and regulated clinical tools, see NextPJ’s guide to AI medical devices and FDA oversight.

Disconnecting and cleaning up your data

When you finish a project, return to Health > Accounts and disconnect sources you no longer need. Then review the chats that used the connection and delete any you do not want retained in history. Remember that disconnecting the source does not delete those conversations for you.

Also review Settings > Memory for saved details and Settings > Data Controls for your broader ChatGPT preferences. OpenAI’s general privacy policy says users can delete chats, saved memories, or an account, and that deletion is normally completed within 30 days unless data must be retained longer for legal, safety, security, or other stated reasons.

If you reconnect later, repeat the medication and condition review. Medical records evolve; a clean setup in July may be inaccurate by your next appointment.

The practical bottom line

ChatGPT Health can reduce the work of assembling records, spotting changes, translating medical language, and preparing for a visit. Use it safely by connecting one source at a time, limiting Apple Health categories, retaining per-request permission prompts, checking imported data, and requiring outputs that point back to dated evidence.

The right mental model is a capable health-information assistant with fallible output—not a doctor, emergency service, or final decision-maker. Keep the original record visible, keep a clinician in the loop for consequential choices, and disconnect sensitive sources when the task is complete.