Apple Health AI study: 200,000 iPhone owners, one year, no diagnosis

Apple did not ship a health chatbot this week. It shipped a study — and the study is the most concrete signal yet of how the company intends to enter consumer health AI.
Apple opened enrollment Monday for the Apple Health AI Usage Study, a research app that lets U.S. iPhone owners ask everyday medical questions and rate the answers, targeting up to 200,000 participants over more than twelve months. The app is already live on the App Store, free, and lists Apple as the developer. Eligibility is narrow: 21 or older, based in the United States, and running iOS 26 or later. The point is not to answer questions for a year. It is to find out whether clinically grounded answers actually help people understand their own health — before Apple puts the capability inside Apple Health or Siri.
The app itself is simple. You type a health question, the way you might type one into ChatGPT, and get a plain-language answer with links back to peer-reviewed clinical sources. Participants then rate responses on accuracy, detail and relevance, and are asked to fill in monthly surveys. Optionally, they can share HealthKit data — resting heart rate, activity, sleep, lab results — so answers carry personal context rather than generic advice. Apple's own App Store description is careful about the boundary: the app can help you understand symptoms, put your health data in context, prepare for a medical visit, or build an exercise and meal plan, but "neither the app nor the research study provides medical advice, diagnosis, or treatment."
The engine is borrowed, and that is the story
The answers are not generated by an Apple model. The app is built on technology from OpenEvidence, the clinical AI platform that has spent the past two years becoming the default medical reference for U.S. doctors. OpenEvidence holds content agreements with The New England Journal of Medicine, JAMA and the JAMA Network, Nature, the NCCN oncology guidelines and Cochrane — the same corpus a physician would consult, wired into a question-answering interface. It says it has supported more than 200 million AI-powered clinical consultations and that more than 100 million Americans will be treated this year by a clinician using it, per the company's own account.
The commercial scale is real. CNBC reported in January that OpenEvidence doubled its valuation to $12 billion in a round led by Thrive and DST, with CEO Daniel Nadler telling the network that 40% of U.S. physicians use the platform and that it passed $100 million in annual revenue — per CNBC's reporting. Its investor list includes Nvidia and the Mayo Clinic. It is HIPAA compliant and SOC 2 Type II certified.
That pedigree is the reason this study is worth more than the sum of its app screens. Apple is not testing whether a chatbot can produce plausible-sounding health text; the market settled that question two years ago. It is testing whether a doctor-grade evidence retrieval system, pointed at a layperson instead of a physician, produces something a layperson can act on. Those are different products with different failure modes, and no one has published a large consumer trial of the second one.
Apple is doing this in the opposite order to everyone else
OpenAI made ChatGPT Health available to all U.S. users in July, after saying people were already asking more than 300 million health-related questions a week. Anthropic announced a healthcare-specific Claude offering in January. Both shipped the consumer product first and are learning from usage. Apple, characteristically, is doing the reverse: measure, publish, then ship — or don't.
That ordering costs Apple time and buys it something the labs do not have. A 200,000-person cohort is larger than most published studies of consumer AI health tools, and it comes with something academic researchers spend years assembling: longitudinal self-reported outcomes across twelve months, tied to device-level physiological data the participant chose to share. If Apple later wants to describe a health feature as clinically grounded — to regulators, to clinicians, to the press — this is the evidence base it would point at. Apple has run large public health studies before through its Research app; this one is aimed squarely at a product decision.
The privacy terms are the strictest part of the design, and they are stated plainly. Data is anonymized before it leaves the phone for a secure third-party cloud, it will not be used to train any AI model, and OpenEvidence deletes query and response data after 24 hours, TechCrunch reports. Those two commitments are worth reading together, because they shape what the study can actually conclude. If the question text is gone within a day, the twelve-month signal lives in the survey responses and the ratings, not in a longitudinal record of what people asked. That is a privacy win. It also means the study will be much better at measuring how people feel about answers than at tracing whether their health changed.
The skeptical case is not small
Start with the sample. Two hundred thousand people who own an iPhone 15 Pro or newer, run iOS 26, and choose to download a research app about health AI are not a cross-section of patients. They are early adopters with recent hardware and enough interest in the topic to opt in. That is fine for product research and weak for epidemiology.
Then there is the rating instrument. Participants score answers on accuracy, detail and relevance — but a participant who could reliably judge clinical accuracy would not need the app. Self-reported accuracy on medical content mostly measures confidence and readability, which is precisely the axis on which consumer health chatbots have looked best while being wrong. We have covered the failure mode directly: consumer AI chatbots missed skin diagnoses in up to 98% of photo cases, and the gap between "sounds right" and "is right" is where this category gets dangerous. Apple has not said whether any clinician adjudicates a sample of the answers, which is the one methodological detail that would change how much weight the results deserve. If it publishes a satisfaction study and calls it a validation study, that distinction will matter.
There is also a small factual conflict in the launch coverage, and readers deciding whether they can join should know it. TechCrunch describes the floor as an iPhone 15 or later with Apple Intelligence enabled. The Verge reports the app requires an iPhone 15 Pro and does not use Apple Intelligence at all — dictation through the standard keyboard works, but there is no Siri or Intelligence integration, according to The Verge. Apple's own App Store listing sets only iOS 26.0 as the requirement and says nothing about Apple Intelligence, which lines up with The Verge's account. Either way, the practical answer for most people is the same: a recent iPhone, the newest iOS, and a U.S. address.
What to watch
Apple's vice president of Health and Fitness, Sumbul Desai, framed the study as the careful path — "responsible innovation starts with scientific rigor and clinical grounding," she said in the launch statement, arguing that a dedicated research study is the most thoughtful way to learn how people engage with health AI. The sentence to hold her to is the one about rigor: rigor means a published methods section, a disclosed adjudication process, and results that appear even if they are unflattering.
Three things will tell us whether this was research or a focus group with a research label. First, whether Apple publishes anything at all, and in what venue. Second, whether the eligibility widens — a study that stays locked to Pro-model owners was aimed at a launch cohort, not a population. Third, whether Apple says one word about the app between now and its next major software cycle. Silence, in this case, would be the louder signal.
If you use a chatbot for health questions today, the thing to borrow from this study is not the app. It is the habit of clicking the citation. An answer without a source you can open is a guess with good grammar.
Would you trust a clinically sourced AI answer more than your own search history — or does the source only matter if you can read it? Write to us via our contact page.







