OpenAI announced on 23 July that it is rolling out Health in ChatGPT to users in the United States, letting people connect Apple Health and supported medical records from US hospital systems, One Medical or Function Health so the assistant can reason over their own health information.

The launch is worth reading closely, but not for the reason the announcement leads with. The most revealing line is a piece of usage data OpenAI discloses about its own earlier, more cautious version — and it explains why this release looks the way it does.

The 70% that changed the design

An earlier version of Health went "to a limited group of users that required visiting a separate space in ChatGPT to get the full benefit." A walled garden: your medical data lived in one tab, and you went there to use it.

Among that group it did not work that way. OpenAI's finding, in its own words: "Among people with access, more than 70% of health-related conversations took place outside the dedicated health experience."

People were asking health questions more often outside the health tab than in it — while planning meals or looking for recipes, in OpenAI's example, where someone might want a food allergy or dietary restriction taken into account. So the design changed to match the behaviour. As the company puts it, "you can now choose to let ChatGPT draw on relevant information you've connected in Health across your conversations, without needing to start within Health."

That is a reasonable product decision. It is also a much larger change than it is presented as. The scope of what medical records can reach has gone from one screen to the whole assistant — where the user turns it on, and subject to a permission prompt that is on by default. OpenAI's own illustrations of the benefit make the expansion plain: ChatGPT "can consider a dietary restriction when helping you choose a restaurant, or a recent injury when planning lower-impact weekend activities with your family."

The clinical use case and the restaurant recommendation now run on the same connected record.

The privacy commitments, and what they actually promise

To OpenAI's credit, the announcement does not hand-wave this. The commitments are concrete enough to hold the company to later:

• Connected medical records and Apple Health information, and conversations that use them, "are not used to train our foundation models or target ads" — and that holds "regardless of the model-training setting you choose for ChatGPT." It is not something a user can accidentally opt into. Conversations that do not use this data follow the standard ChatGPT training setting, changeable in Settings.
• By default ChatGPT asks permission before using connected health information to personalise a response. Users can approve each request or switch to always-allow, which turns the prompts off.
• Disconnecting a source deletes the synced data from OpenAI's systems within 30 days. The caveat is stated plainly: information already in your conversation history stays until you delete those conversations.
• Memories "are not created directly from connected medical records and Apple Health information" — though memories can be created from Health conversations.
• All conversations are encrypted at rest and in transit; information connected in Health "receives additional encryption protections".
• Before ChatGPT takes an action in another connected plugin that could disclose health information — OpenAI's example is sending a running partner a training plan built from Apple Health data — additional safeguards "help check that the sharing matches your request", and for some sensitive actions ChatGPT "may ask you to confirm before proceeding". The company says it red-teams this specifically.

These are commitments written to be checked later, and the "regardless of the model-training setting" line is the sort of thing that is expensive to walk back.

What it can and cannot do

Availability is narrower than the announcement implies: logged-in users 18 and older, in the US only, on web and iOS, across Free, Go, Plus and Pro. It is not yet available in Codex. You can still ask health questions without connecting anything at all.

OpenAI says it works with "hundreds of physicians around the world" to build scenarios and rubrics, and that these evaluations assess qualities including accuracy, safety, communication, context awareness, completeness and appropriate escalation. Every GPT‑5.6 model, it says, outperformed GPT‑5.5 on HealthBench Professional, "an evaluation of AI performance on challenging health tasks". GPT‑5.5 Instant "brought frontier health intelligence to all free users".

The limitations are stated rather than buried. ChatGPT "can still make mistakes and does not replace the care and judgment of qualified medical professionals." And there is a data-quality warning that anyone who has read their own patient portal will recognise: "Connected information may not always be complete or current, for example, a medication may remain listed after you stop taking it, so tell ChatGPT when something changes and check important details against the original source."

OpenAI's instruction there — check important details against the original source — is the right one, and it is doing a lot of work. A model reasoning confidently over a record that still lists a drug you stopped a year ago will produce fluent, well-structured, wrong output, and the fluency is the problem, because it is the part that reads as authority. The company's answer is that the user corrects the record and verifies against the source. That is a real answer. It also puts the burden on the person least equipped to spot the gap.

Why it matters

The scale claim is the context for everything else: OpenAI says more than 300 million people a week already bring health-related questions to ChatGPT. Those conversations are happening with or without connected records. The argument for Health is that grounding them in a real record beats having them float free — and there is something to that.

But the shift here is structural, not incremental. Medical records have moved from a bounded feature into the general context of an assistant that people use for everything, on the empirical grounds that this is what users were doing anyway. The safeguards around it are specific and, on paper, strong. The surface area they have to cover just got much larger.

The thing to watch is not whether OpenAI honours the training and advertising commitments — those are clear, and clear commitments are checkable. It is the ordinary failure modes at the boundary: a stale medication list quietly shaping advice, or health context surfacing in a conversation the user was not thinking of as medical. The permission prompt is the stated mitigation for the second, and it is on by default — but it is also the first thing a frequent user will turn off.

Note on sourcing: OpenAI's own announcement page is the primary for this piece. It blocks automated retrieval from our infrastructure, so we read it via the Internet Archive's snapshot taken on 25 July; both links are below and the canonical page opens normally in a browser.