OpenAI just opened ChatGPT Health to every adult in the US, on every plan, free included. You can now plug in Apple Health and your hospital records, and ChatGPT will use that context in any conversation, not just a walled-off health tab. It shipped one day after a former pastor sued OpenAI over health advice that he says nearly killed him. Here is what the feature actually does, how to set it up in about ten minutes, and where the guardrails are thin.
On July 23, 2026, OpenAI began rolling out Health in ChatGPT to logged-in US users aged 18 and over, on web and iOS, across Free, Go, Plus, and Pro. It is not in Codex.
That is a big jump from January, when the same feature launched as a limited pilot for a handful of testers.
Here is the number that explains why OpenAI keeps pushing on this: more than 300 million people ask ChatGPT health questions every week. In January that figure was 230 million. Seventy million more weekly health conversations in six months.
People are already doing this. OpenAI is just deciding how.
What actually changed this week
The January version had a design problem. Health lived in its own sandboxed corner of ChatGPT, and you had to go there deliberately for it to work.
OpenAI’s own testing found that more than 70% of health-related conversations happened outside that dedicated space. Someone planning dinner wants ChatGPT to remember a food allergy. Someone booking a weekend hike wants it to remember a bad knee. Nobody switches tabs for that.
So OpenAI removed the wall. Your connected health information can now flow into any conversation, with your permission, wherever you happen to be chatting.
Three other things shipped alongside it:
- Broader record connections. Apple Health, supported US hospital systems, One Medical, and Function Health. Press reports also name Epic and Oracle Health among the supported record systems.
- Better models on health questions. GPT-5.5 Instant powers free accounts. GPT-5.6 Sol handles paid accounts and is OpenAI’s strongest health model so far.
- A permission layer. By default ChatGPT asks before touching your records. You can approve each time or switch to always-allow.
Set it up in 10 minutes: step by step
Step 1: Check that you qualify
You need to be 18 or older, physically in the US, and logged in on web or iOS. The medical records connection is US-only. If you are outside the US, skip to the global section further down, because there is still something here for you.
Step 2: Open Health
In ChatGPT, look for Health in the left sidebar, or under the More menu. Click Get started. If you do not see it yet, the rollout is staged, so give it a few days.
Step 3: Connect your sources
ChatGPT walks you through linking accounts. Start with Apple Health if you own an iPhone or Apple Watch, then add your hospital portal, One Medical, or Function Health. Syncing takes a few minutes. You can close the tab and come back.
Step 4: Fix what the sync got wrong
This step matters more than people expect. Health records are messy. A medication you stopped taking two years ago will still be sitting in your chart.
Go through the conditions and medications list. Delete what is no longer true. Add what is missing, including family history, which almost never syncs automatically.
Step 5: Connect your wearable properly
If you use a fitness or nutrition app, check that it shares data into Apple Health first. ChatGPT reads what Apple Health has, not what your app has. Proprietary scores that some apps calculate internally often will not transfer at all.
Step 6: Choose your permission setting
Default behaviour is a prompt every single time ChatGPT wants to use your records. You can flip that to always-allow under Settings > Plugins > Health.
Worth doing: leave the prompts on for the first two weeks. You will learn a lot about when ChatGPT thinks your medical history is relevant, and some of those moments will surprise you.
Step 7: Pull health context into any chat
Type @Health in a message to explicitly ask ChatGPT to include your health context. Otherwise it decides on its own when the topic warrants it, then asks.
Step 8: Know how to leave
Go to Health > Accounts and disconnect any source. Data synced from that source is deleted from OpenAI’s systems within 30 days. Anything already sitting in your chat history stays there until you delete those conversations yourself.
Five prompts worth running on day one
Generic prompts get generic answers. These are shaped around what the tool is actually good at.
- Trend reading: “Compare my last three lipid panels. What moved, by how much, and which changes are outside normal range?”
- Appointment prep: “I see my doctor Thursday about persistent fatigue. Based on my records, draft eight questions I should ask, ranked by importance.”
- Translation: “Rewrite my last visit note in plain language at a level a smart teenager would follow. Flag anything I should have asked about.”
- Everyday context: “Suggest four dinners for tonight that respect my dietary restrictions and current medications.”
- Timeline building: “Build a chronological timeline of my diagnoses, procedures, and medication changes over the past five years.”
That last one is what early testers keep praising. One tester quoted by OpenAI, a technical program manager named Blake, described turning scattered medical history into a clear timeline he could actually explain to a physical therapist.
The lawsuit sitting underneath this launch
One day before the rollout, Scott Winters, a 55-year-old former Florida pastor, sued OpenAI and Sam Altman in San Francisco County Superior Court. The claims are negligence and unauthorised practice of medicine.
According to the complaint, Winters consulted ChatGPT-4o for months about dizziness and blood pressure instability. He says the chatbot stopped recommending a doctor, told him to rest, and characterised his symptoms as not dangerous. In July 2025 he was admitted to intensive care with a massive pulmonary embolism. A treating physician attributed the clots partly to the prolonged immobility he had been advised to maintain.
OpenAI’s position is that ChatGPT is not a substitute for medical care, and its terms state the service is not intended for diagnosis or treatment of any condition.
Both things can be true at once. The tool can genuinely help you understand a lab result, and it can also sound authoritative while being wrong about something urgent. That gap is the entire risk.
What the research says, plainly
Benchmark scores and real-world performance are not the same thing, and there is now solid evidence for that.
A University of Oxford study published in Nature Medicine ran a randomised trial with nearly 1,300 participants working through doctor-written medical scenarios. Half used an LLM. Half used ordinary methods like web search and their own judgment.
The LLM group did not make better decisions.
The researchers found three specific failure points. Users did not know what information to give the model. Small changes in how a question was phrased produced very different answers. And responses often mixed good advice with bad advice in a way users could not separate.
Dr Rebecca Payne, the study’s lead medical practitioner, put it bluntly: AI is not ready to take on the role of the physician, and asking a model about your symptoms can be dangerous.
Meanwhile OpenAI reports that every GPT-5.6 model beat GPT-5.5 on HealthBench Professional, its own benchmark for hard health tasks, built with hundreds of physicians worldwide.
Both statements are accurate. Models are improving on structured tests. Structured tests do not capture what happens when a scared person types four vague sentences at 2am.
The privacy question nobody should skip
OpenAI’s commitments here are specific and worth reading closely.
| What OpenAI says | What it means for you |
|---|---|
| Connected records and Apple Health data, plus conversations using them, are not used to train foundation models or target ads | This holds regardless of your general training setting. Conversations that do not touch health data follow your normal preference |
| Health data gets encryption on top of standard ChatGPT encryption | Layered protection, not a separate promise of anonymity |
| Disconnect any account and synced data is deleted within 30 days | Chat history is separate. You have to delete those conversations yourself |
| Memories can form from Health conversations, but not directly from connected records | Use Temporary Chat or switch memory off for anything you want left untouched |
| Extra safeguards run before ChatGPT shares health info through another connected app | Sending a training plan built on your Apple Health data to a friend triggers a confirmation step |
Now the part OpenAI cannot promise. Consumer AI health tools largely sit outside HIPAA. That law covers doctors, hospitals, and insurers. It does not automatically cover an AI company holding a copy of your chart, as the Center for Democracy and Technology has pointed out.
You are trusting a company policy, not a federal protection. Decide with that clearly in view.
If you want a refresher on keeping sensitive prompts off the record entirely, our guide to ChatGPT Temporary Chat covers the mechanics.
If you are outside the US
Record connections are US-only, and there is no announced timeline for other countries. At the January pilot, users in the European Economic Area, Switzerland, and the UK were excluded outright.
You can still get most of the practical value without connecting anything:
- Paste, do not connect. Upload a lab PDF or type your numbers directly. You get the same interpretation without a standing data link.
- Use it for appointment prep. Describing symptoms and asking for a ranked question list works in any country and does not depend on records access.
- Check the alternatives. Anthropic’s Claude for Healthcare launched in January with Apple Health and Android Health Connect integrations. Microsoft shipped Copilot Health in March. Google has a Gemini-powered health coach. Availability varies by market, so check your own.
- Mind your local privacy law. GDPR, PIPEDA, PDPA, and similar frameworks treat health data as a special category. What is permitted in the US may not be permitted where you live.
Limitations you should not ignore
- Records go stale. Discontinued medications linger. OpenAI says so directly. Verify anything that matters against the original source.
- Wearable data is partial. Only what your app pushes into Apple Health arrives. Proprietary readiness or recovery scores often do not.
- No Android record sync, no Codex. Web and iOS only for now.
- Confident tone, uncertain accuracy. The Oxford finding stands. Fluent answers mix good and bad guidance in ways non-experts cannot reliably separate.
- It will not escalate the way a nurse would. This is precisely what the Winters complaint alleges failed. Treat urgent symptoms as urgent regardless of what any chatbot says.
- Adults only, US only. Under-18 users are excluded from the medical records feature entirely.
Who this is genuinely useful for
People managing chronic conditions. Years of scattered results turned into one readable timeline is real value.
Caregivers. If you manage a parent’s appointments, being able to summarise a visit note in plain language saves hours.
Anyone who freezes in the exam room. Walking in with eight ranked questions changes the quality of the visit.
People who leave appointments confused. Feed the visit summary in afterwards and ask what you should have asked.
Where it is a poor fit: acute symptoms, anything involving chest pain or breathing difficulty, mental health crises, and any moment where the honest answer is that you need a human clinician now.
The bottom line
ChatGPT Health is a well-built comprehension tool wearing a slightly dangerous costume. Reading your chart, tracking trends, translating jargon, prepping questions: it does all of that well, and 300 million weekly health questions say people want exactly this.
The danger is not the feature. It is the drift. You start by asking what a lab value means, and six weeks later you are asking whether you need to go to hospital. That is the path the Winters lawsuit describes, and the Oxford study explains why it is so easy to walk down without noticing.
Use it to understand your health. Do not use it to decide your health. Draw that line before you connect anything, because it is much harder to draw once ChatGPT knows your entire chart and sounds certain.
The best question to ask an AI about your health is not “what do I have?” It is “what should I ask my doctor?”
Sources
- OpenAI, “Launching Health in ChatGPT,” July 23, 2026 openai.com
- OpenAI Help Center, “Health in ChatGPT” help.openai.com
- OpenAI, “Introducing ChatGPT Health,” January 7, 2026 openai.com
- TechCrunch, “OpenAI makes ChatGPT Health available to all U.S. users,” July 23, 2026 techcrunch.com
- CBS News, “ChatGPT’s medical advice nearly killed a Florida man, lawsuit against OpenAI claims,” July 23, 2026 cbsnews.com
- University of Oxford, “New study warns of risks in AI chatbots giving medical advice,” February 9, 2026 phc.ox.ac.uk
- Bean et al., “Clinical knowledge in LLMs does not translate to human interactions,” Nature Medicine nature.com
- HealthBench Professional arxiv.org
- Fierce Healthcare, “OpenAI taps b.well to connect health data for ChatGPT Health,” January 8, 2026 fiercehealthcare.com
- Fierce Healthcare, “JPM26: Anthropic launches Claude for Healthcare,” January 11, 2026 fiercehealthcare.com
- Engadget, “OpenAI once again makes the case for giving ChatGPT your health records,” July 23, 2026 engadget.com
Just Being Resourceful! 
Leave a comment