Apple Launches Health AI Study to Examine How Americans Use Chatbots for Medical Questions

There is no live monitoring, and staff won't contact anyone over self-harm questions.
Apple's study app has safety guardrails, but the limits of its oversight are explicit.
Mark

Why does this study matter beyond Apple's own products?

Mimi

Because one in three US adults already use chatbots for health information. Apple is trying to measure what happens after the answer: do people understand it, trust it, act on it.

Mark

What's the setup?

Mimi

A dedicated app, powered by OpenEvidence, with answers linked to citations from journals like NEJM and JAMA. Up to 200,000 people for as long as 12 months.

Luke

Worth noting that the stated goals are Apple's description from a press briefing. We don't have independent findings, and nothing here says when results will be published.

Mark

And the privacy design?

Mimi

Layered. Scrubbing on the phone, then a contractor, UST Global, delinks IDs, then OpenEvidence answers and deletes data within 24 hours. HealthKit sharing takes two separate opt-ins.

Luke

Those are design claims. The reporting doesn't say anyone has audited them. And Apple says data won't train AI models, but it may inform feature development, which is a real distinction to keep in view.

Mark

What about safety?

Mimi

Guardrails nudge people toward providers when questions are concerning. But there's no live monitoring, and staff won't reach out over self-harm questions.

Luke

That's the plain gap. Also, the iPhone 15 Pro and iOS 26 requirement means the sample skews toward owners of newer, pricier phones, so it may not represent all Americans.

  • A crowded and fast-moving territory: one in three American adults already consult AI chatbots about their health, and the habit is outpacing the evidence.
  • Apple is entering that space with a 12-month study of 200,000 participants, each asked to ask real health questions and report what happened next.
  • Privacy is being layered in stages, with questions stripped of identifiers on the phone, routed through a third-party cloud, and deleted within a day by the answering platform.
  • Guardrails flag concerning questions and point people toward providers, but there is no live monitoring, and self-harm queries will not trigger a follow-up call.
  • The trajectory now turns on whether those safeguards hold and whether Apple's feature development, informed by what it learns, earns the trust it is measuring.

On a Sunday in November, Apple opened a new door onto a question that has quietly become part of ordinary life: when people ask a machine about their bodies, what do they understand, whom do they trust, and what do they do next? Through a standalone app built on its long-running Research program, the company is inviting 200,000 American adults to help answer that question over the course of a year. The work reflects a moment in which roughly one in three US adults already turns to AI chatbots for health information, and it asks whether such tools can be designed responsibly before their habits harden. Apple has pledged that the data will not train its AI models, though it may shape the features it builds next.

On Sunday, Apple released the Apple Health AI Study, a standalone app that grows out of its Research program, which has run studies on hearing, heart health, movement and women's health since 2019. The new project asks a simple but pressing question: how do people actually use conversational health AI, and can they understand what it tells them? Apple plans to look at comprehension, the quality of answers, what builds or erodes trust, and the everyday actions people take afterward, including whether these tools help them prepare for doctor's visits.

Answers come from OpenEvidence, a platform built for physicians, which draws on peer-reviewed sources such as the New England Journal of Medicine, JAMA, Nature and the Cochrane Database, with linked citations. Participants may optionally share HealthKit data like sleep, lab results and heart rate for more personal responses. Eligibility is narrow: US residents aged 21 or older with an iPhone 15 Pro or later running iOS 26.

Participation can last up to a year, with surveys at the start and monthly, and Apple Watch users receive check-in prompts. Privacy is handled in stages. Questions are scrubbed of identifiers on the phone, passed through a third-party cloud environment, and then answered and deleted within 24 hours. Only a small group of Apple researchers sees coded data. Guardrails flag worrying questions and suggest help, but there is no live monitoring, and self-harm queries won't prompt outreach from study staff. The real test ahead is whether those safeguards hold, and how Apple uses what it learns.

On Sunday, Apple released a standalone app called the Apple Health AI Study, the newest project to come out of its Research app program. Apple described it in a press briefing as an effort to learn how people actually use conversational health AI tools, and it is open to 200,000 participants.

The Research app dates to 2019. Since then, Apple has worked with medical universities and institutions on studies of hearing, heart health and movement, and women's health. Its most recent project before this one was the Apple Health Study, which covers a broad range of health areas. The new study gets its own app.

Apple says the study will look at four things: whether people understand what a health AI tells them, the quality of its responses, what drives user trust, and the everyday health actions people take afterward. It will also explore whether such tools can help people prepare for doctor's visits. The stated aim is to inform responsible design of health AI, for Apple and for the wider industry. The study arrives as one in three US adults use AI chatbots for health information. Apple says the data will not be used to train AI models, though it may inform the company's feature development.

To run the study, Apple is working with OpenEvidence, an AI platform built for doctors. Its answers draw on medical information from The New England Journal of Medicine, the Journal of the American Medical Association, the National Comprehensive Cancer Network, Nature and the Cochrane Database of Systematic Reviews. Inside the app, participants type health questions and get answers with linked, peer-reviewed clinical citations. They can also choose to share HealthKit data, such as sleep, lab results, activity and resting heart rate, for more personalized answers. HealthKit is Apple's hub for health and fitness data from the iPhone and Apple Watch.

Eligibility is narrow. Participants must live in the US, be 21 or older, and have an Apple Intelligence-enabled iPhone, meaning an iPhone 15 Pro or later, running iOS 26 or later. Apple says the app needs those phones' processing power and uses on-device Apple Foundation Models, which are part of Apple Intelligence. In the study, those models do jobs such as classifying the category of each question.

Each participant is asked to take part for up to 12 months. They can ask about general health, specific symptoms, recovery from a procedure, or their own uploaded health data. There is a baseline survey, monthly surveys, and extra surveys about preparing for doctor's visits. Participants with a paired Apple Watch will get check-in notifications with suggested topics. Someone who stops participating altogether may be removed.

Privacy is handled in stages. After mandatory onboarding that includes informed consent, participants share demographic details such as race, age and education level, which Apple says will be used to study how those factors affect preferences and communication needs. Questions are first scrubbed on the iPhone of identifying details. They then go to a secure cloud environment run by a third-party contractor, UST Global, where they are separated from participant IDs and screened again. Finally they reach OpenEvidence, which generates answers and deletes both question and response data within 24 hours. A small group of Apple researchers will see coded survey responses, conversation summaries and categories of the most common conversation types.

The app includes guardrails that detect concerning questions and suggest following up with a provider or other resources. There is no live monitoring, though, and study staff will not contact participants who ask questions related to self-harm. Sharing health records is a two-step choice: participants must first upload clinical records to Apple's Health app, then separately share them with the study app. What to watch is how those safeguards hold up and how Apple uses what it learns.

Study data will not be used to train AI models, but it may inform feature development.
— Apple, as reported by CNET
Quer a matéria completa? Leia o original em CNET ↗
Fale Conosco FAQ