200,000 people test Apple's health AI before it ships
Apple will recruit up to 200,000 Americans to test a health AI chatbot - before it ships anything like it in the Health app or Siri.
OpenAI already has the opposite problem: more than 300 million people a week ask ChatGPT health questions, and Apple just decided to study the question first.
What Apple actually announced
Today, November 15, Apple announced the Apple Health AI Usage Study, a virtual research study with its own app, built with clinical AI company OpenEvidence.
The facts, from Apple's own newsroom announcement (I pulled the page myself to check the hardware line):
- Up to 200,000 participants, age 21 and older, US only.
iPhone 15 Proor later, runningiOS 26, with Apple Intelligence enabled. (TechCrunch wrote "iPhone 15or later"; Apple's announcement says 15 Pro. Apple's version wins.)- Participants ask everyday health questions and get answers grounded in peer-reviewed clinical sources - NEJM, JAMA, Nature - with linked citations on each response.
- They can optionally share
HealthKitdata: resting heart rate, activity, sleep, lab results. The study measures whether personal context changes how people understand the answers. - The privacy terms are the strictest part of the whole announcement: study data is anonymized before going to a secure third-party cloud, it will not be used to train any AI models, and OpenEvidence deletes the query and response data after 24 hours.
- The app is not meant to diagnose or treat anything, and does not replace a doctor.
What Apple measures: trust, response quality, comprehension, and everyday health actions.
What Apple did not say: whether any of this ever lands in the Health app or Siri.
TechCrunch reports the study runs more than 12 months; Apple's own announcement gives no duration at all.
The number that makes this a strategy, not a feature
200,000 participants sounds big until you put it next to the incumbent. OpenAI said in July 2026 that more than 300 million people a week ask ChatGPT health questions - up from 230 million a week when ChatGPT Health launched in January.
Apple's entire study cohort is 0.067% of one week of ChatGPT's health traffic. That is the whole story.
Apple is not launching a health chatbot. It is launching a 200,000-person experiment to find out whether people should trust one - ten months after its biggest rival already shipped one to hundreds of millions of people and started connecting their medical records.
Why OpenEvidence, and why that choice is the tell
Apple did not build this model. It rented the one doctors already use.
OpenEvidence raised $250 million in September 2026 at a $15 billion valuation, and says nearly two-thirds of US physicians use its free, ad-supported clinical search.
A JAMA study published October 15, 2026 put harder numbers on the adoption: 271 million queries from 952,008 verified US clinicians between January 2024 and June 2026, with monthly active users growing from 355 to 703,265 - including 556,200 verified physicians cumulatively.
Read the fine print on that study, though, because it matters for what Apple is buying. It measures adoption, not accuracy. It used OpenEvidence's own de-identified data, three of its seven authors were OpenEvidence employees, and it did not evaluate whether the answers were right or whether patients did better.
A Nature Medicine comparison from June 2026 found general-purpose models outperforming OpenEvidence on standard clinical benchmarks.
So Apple picked the most-used clinical AI in America, and the one whose accuracy record is the least settled - then wrapped it in the strongest privacy terms it could write: anonymized, no training, 24-hour deletion.
The privacy architecture is the product decision. Apple is not confident enough in the answers to ship them, and not confident enough in the data loop to train on them, so it is buying a year of evidence first.
The uncomfortable comparison
There are two ways to read "study first, ship later":
- Caution as a feature. Health answers that are wrong in a confident tone are worse than no answers. A 200,000-person study with cited sources, optional
HealthKitcontext, and monthly surveys is how you find out whether people act on bad advice before you scale the bad advice. - Lag dressed as rigor. ChatGPT Health has been live since January. Claude, Amazon, Copilot and Perplexity all shipped health products in the first quarter of 2026. Apple's study will still be running in late 2027. By the time Apple publishes findings, the behavior it is studying - hundreds of millions of people asking general chatbots about their symptoms - will be two years old.
Both can be true. The number that decides which one wins is not the 200,000. It is the 24 hours: OpenEvidence deletes the data a day after the question.
Apple is studying trust in a system designed to forget everything it learns. That is either the most honest health AI design on the market, or an admission that the company does not yet believe its own answers enough to keep them.
Sources
Sources: Apple Newsroom, "Apple launches research study to explore the use of AI for health" (November 15, 2026); TechCrunch, "Apple launches new app to study how people use AI for health" (November 15, 2026); JAMA OpenEvidence usage study, via Employed AI's analysis (published October 15, 2026; 271M queries, 952,008 clinicians, Jan 2024-Jun 2026); OpenAI's 300M weekly health users figure, announced July 23, 2026.
The iPhone 15 Pro requirement, 200,000 cap, and anonymization/no-training terms were confirmed against Apple's announcement directly. TechCrunch's "more than 12 months" duration is attributed to TechCrunch - Apple's announcement states no duration.
Comments
No comments yet. Start the discussion.