OpenAI Expands ChatGPT Health Beyond Its Dedicated Space, With Consent as the Gate

The revised design removes the need to begin in Health. A separate account of the launch likewise describes the feature as accessible from the sidebar while allowing connected data to inform relevant questions elsewhere in ChatGPT.

That makes this primarily a product-distribution change. It reduces the friction of moving into a separate health tab, but it does not remove the separate decisions to link data, grant access and check whether the context is accurate.

More context is optional—and may still be partial

By default, OpenAI says ChatGPT asks before it uses connected medical records or Apple Health information to personalize a response. A user can approve an individual request, set continuing permission, or invoke the context with @Health. The setting can later be changed, and connected accounts can be disconnected.

OpenAI says synced information from a disconnected source is deleted from its systems within 30 days. Material already included in a ChatGPT conversation remains in that conversation history until the user deletes it. The company also says connected medical records, Apple Health data and conversations that use them are not used to train its foundation models or target advertising, regardless of the user’s usual model-training setting; it describes those as additional privacy and security protections.

Those are the company’s stated controls, and they matter because the feature is designed to carry health context into non-health-specific requests. They do not establish that every data source is complete. OpenAI explicitly warns that a medication can remain listed after someone stops taking it. Information from a wearable, fitness or nutrition app is available only to the extent that app shares it through Apple Health, and some proprietary scores may not transfer.

Users can correct or add context, but that puts an important limit on any promise of a complete longitudinal record. The company says people should check important details against the original source and discuss medical decisions with a healthcare provider.

OpenAI’s company-reported HealthBench Professional methodology: physician-reviewed tasks were selected from 15,079 examples for a 525-task benchmark. Source: OpenAI.

The model claims are not a clinical handoff

In its announcement, OpenAI says more than 300 million people ask ChatGPT health-related questions each week. It says it has used dedicated training and worked with hundreds of physicians worldwide to create health scenarios and scoring rubrics covering accuracy, safety, communication, context awareness, completeness and appropriate escalation.

The company presents GPT-5.5 Instant, available on the Free plan, as having improved recognition of situations where urgent care may be needed, and calls GPT-5.6 Sol, available to paid users, its strongest health model for more complex questions. It also says every GPT-5.6 model outperformed GPT-5.5 on HealthBench Professional, an evaluation of challenging health tasks. These are OpenAI’s descriptions of its training and evaluation results, rather than evidence that the product can replace a clinician’s judgment.

OpenAI says physicians tested Health in ChatGPT before release, while also cautioning that ChatGPT can make mistakes and does not replace qualified medical professionals. The rollout is limited to U.S. adults using web or iOS and is not yet available in Codex.

What would show whether the expansion is useful

The immediate decision is still with users: whether to connect records and permit their use outside the Health section. The pilot statistic makes a case that health questions arise inside ordinary chats, but it does not measure connections, continued use, answer quality or outcomes.

The evidence that would resolve the launch’s central claim is therefore more specific than account availability: how often people connect data, how often they authorize access, how they correct stale or missing records, and whether OpenAI can show that its permission controls and responses remain understandable in those mixed-purpose conversations. Until then, the release is a broader availability test of permissioned context, not proof that putting health data inside a general chatbot improves healthcare decisions.