On this page · 5 sections ▾
A clinically grounded AI answer is a response tied to a defined set of medical sources, such as peer-reviewed journals, with links back to those sources. A general chatbot writes from patterns it learned during training. A grounded system is meant to anchor its wording in documents a reader can check. Apple is now studying how people respond to this kind of answer.
What is Apple actually testing?
As The Verge reported, Apple launched the Apple Health AI Usage Study with OpenEvidence, a clinical AI platform used by healthcare professionals. Participants ask medical questions in a new iOS app and rate the replies on criteria including accuracy, detail and relevance.
TechCrunch adds that the study is meant to run for more than 12 months with up to 200,000 participants, who also complete monthly surveys. The research question is not whether AI can talk about health. It is how people understand, judge and act on answers that rest on clinical literature.
How does grounding work?
The sources do not describe OpenEvidence's internal design, so the following is general background. A common approach, often called retrieval-augmented generation, has two steps. First, the system searches a curated library for passages relevant to the question. Second, a language model writes a plain-language answer from those passages and attaches references.
The library is the key difference. Per the coverage, OpenEvidence draws on medical information from outlets like The New England Journal of Medicine, JAMA and Nature. Participants get easy-to-understand responses with links to peer-reviewed clinical sources, so a reader can open the original and see whether the summary holds up.
Grounding narrows the risk but does not remove it. A model can still misread a passage or leave out context. That is one reason Apple is asking people to rate the answers.
Why does personal data enter the picture?
Participants can choose to share HealthKit data, such as activity, sleep, resting heart rate or lab results. The aim is more personal, pointed responses. The same question can deserve a different answer depending on a person's own numbers.
That raises the privacy stakes. Apple says the data will not be used to train AI models. According to TechCrunch, information is anonymized before it goes to a secure third-party cloud provider, and OpenEvidence deletes query and response data after 24 hours. Sharing is optional.




