In September 2023 WHOOP launched WHOOP Coach, built on OpenAI's GPT-4, to give members conversational answers drawn from their own recovery, strain, sleep and stress data.1 In March 2025 Oura rolled out Oura Advisor to every Ring Gen3 and Ring 4 member after testing it in Oura Labs.2 In November 2025 Function Health launched an AI model and a chatbot that answers questions using members' lab results, notes and scans.3 In under three years the AI coach went from a novelty to a standard feature.
Do people trust an AI with health questions?
They use it more than they trust it. KFF's 2024 health misinformation poll found that 17% of US adults use AI chatbots at least once a month to find health information and advice, rising to 25% of adults under 30. Among people who use AI, 56% were not confident that chatbots provide accurate health information, and across the whole public the figure was 63%.4 Edelman's 2025 report found 82% of people trust their doctor to tell the truth about health.5 People are asking the machine and still checking with a person.
Does the AI help people reach the right answer?
The best evidence so far says the model alone is not the problem. In a randomized, preregistered study published in Nature Medicine in February 2026, 1,298 people in the UK each worked through medical scenarios. Tested on their own, the language models identified the relevant condition in 94.9% of cases and chose a correct course of action in 56.3%. When people used the same models for help, they identified the relevant condition in fewer than 34.5% of cases and chose the right course of action in fewer than 44.2%, no better than a control group using search engines and their usual resources.6
The model on its own, against people using it
Bean et al., Nature Medicine, 2026
The knowledge was in the model and got lost in the conversation. Reviewing a sample of cases, the researchers found that people often gave the model incomplete or incorrect information, and that the models sometimes gave misleading answers back. Standard benchmarks had not predicted either problem. A health app's coach has an advantage, because it already has the member's data, but the lesson carries: the hardest part is helping a person act on the answer, and that is the part human experts have always been good at.
What happens when an AI coach gets it wrong?
In 2023 the National Eating Disorders Association took down Tessa, a chatbot launched to support people with eating disorders, after users reported it had given them dieting advice, including calorie counting and regular weigh-ins. NEDA's chief executive said a chatbot cannot replace human interaction.7 Tessa was built for a sensitive group and still crossed a line its designers had not tested well enough. Any coach that talks about food, sleep, training or mood needs a written boundary for symptoms and medication, tested before launch like any other safety feature.
Where should the handover happen?
Where the AI coach answers, and where it hands over
Adelo Insights
Three kinds of question come up. "What does my number mean?" is where the AI is at its best: fast, personal and available at 4 am. "Why should I care, and what should I change?" needs a person people believe, which is why a named expert in a regular live session is still the strongest way to make an aspiration stick. "Is this symptom normal?" or "Should I stop my medication?" needs a clinician, and the AI should hand over at once.
The pairing works the other way too. An AI that can search a library of recorded expert sessions and cite the moment a sleep physician answered the same question gives members the speed of a chatbot with the credibility of a named expert. Every question the AI can't answer well is a topic for the next session.
- Let the AI answer questions about the member's own data, and cite the source for anything it says about health.
- Send "why" questions to the expert program, and build your next live sessions from them.
- Stop coaching on symptoms and medication. Route to a clinician immediately, and test the boundary with one before launch.
- Read what members ask the AI. The question log is the same research asset as a live Q&A, at far greater volume.
Adelo Insights · Health tech and the expert voice
