See how ChatGPT patient health records now surface in any chat, why OpenAI dropped the sandbox, and what it means for regulated industries.
OpenAI has started rolling out a Health feature inside ChatGPT that lets users connect Apple Health data and medical records directly into the chatbot, effective now for logged-in users aged 18 and over on web and iOS. The feature is live across ChatGPT's Free, Go, Plus, and Pro tiers, and users can sync data from Apple Health as well as, where supported, US hospital systems, One Medical, and Function Health, according to OpenAI's rollout announced in late July 2026.
Once connected, ChatGPT can draw on a person's medications, lab results, recent visits, sleep data, and activity logs inside any conversation — not just a dedicated health section. That is the core change. OpenAI's earlier pilot required users to open a specific health area to get answers grounded in their own records, but the company found that more than 70 percent of health-related conversations in that test group happened outside that space entirely, buried inside meal planning chats or unrelated symptom questions. That usage pattern is what drove the redesign.
Why OpenAI moved data out of a sandbox and into the main chat
The shift matters because it changes what "connected" actually means in practice. A dedicated health tab is easy to govern and easy to ignore. A chatbot that pulls medical context into any conversation, unprompted by a mode switch, is a different kind of product — closer to an assistant with standing access to sensitive records than a tool a user visits on purpose.
OpenAI has kept the Health tab in the sidebar, but its function has narrowed to account management: connecting sources, reviewing synced data and trends, and returning to past health-related chats. The intelligence layer now sits underneath the entire product, not beside it.
That distinction is where OpenAI's own published research becomes relevant. The company has previously used its HealthBench evaluation framework to measure how its models perform against realistic medical conversations, an effort built specifically because general-purpose benchmarks don't capture the risk profile of health-related answers. Folding personal records into every conversation raises the stakes on exactly the kind of accuracy and judgement HealthBench was designed to test.
What early testers are actually reporting
OpenAI published accounts from its early access group, and they split fairly cleanly into two categories.
The first is pattern-recognition across scattered records. Blake, a technical program manager cited by OpenAI, described using the tool to turn "disconnected diagnoses, imaging results, and surgery notes" into a single timeline he could explain to a physical therapist. Reweti, a portfolio manager, made a similar point about longitudinal analysis: connected labs let the model "infer patterns that aren't obvious," which he compared to having a research analyst on hand.
The second category is more cautionary. Shannon, a nurse in the same early access group, found an unexpected entry in her own chart through the feature — not because the AI created an error, but because it surfaced something she hadn't noticed, which she then followed up on with her own healthcare providers. That is the line OpenAI is trying to hold: a tool that flags things for human review, not one that makes a clinical call. Whether that line survives contact with millions of ChatGPT's Free-tier users, most of whom have no clinical training to catch the model when it gets something wrong, is the open question analysts on health AI have been circling for years.
The workflow lesson for regulated industries outside the US
Australian health providers won't get this exact feature — the integrations are built around Apple Health, US hospital systems, One Medical, and Function Health, none of which map directly onto Australian health infrastructure or My Health Record. But the underlying pattern is transferable: OpenAI didn't build a chatbot that answers health questions in isolation. It built a pipeline that pulls structured data from multiple systems into a single conversational interface, then measured how people actually used it before deciding where the intelligence layer should sit.
That is the same problem facing any regulated Australian sector — allied health administration, aged care compliance, financial services reporting — where the data already exists in multiple systems but nobody has connected it to a workflow. OpenAI's redesign is a reminder that the hard part of this kind of deployment isn't the model. It's deciding where in the process the AI gets access, what triggers it, and what a human is still required to check.
Diagnostics
Executives weighing what this kind of integration means for their own regulated data are asking themselves a narrower set of questions:
Where does sensitive client or patient data currently sit disconnected across separate systems, waiting for someone to manually stitch it together before it's useful?
If an AI tool had standing access to that data inside every conversation, is there a documented process for what it's allowed to flag versus what still requires a qualified person to confirm?
Has anyone tested how staff actually use an AI tool in practice, the way OpenAI tracked its 70 percent figure, or is the deployment plan still based on how the tool was designed to be used rather than how it's being used?
