Stanford's AI Coach Outperforms Human Health Experts at Persuading People to Exercise
A study of My Heart Counts, a cardiovascular research app developed at Stanford University School of Medicine in partnership with Apple, found that participants preferred motivational messages generated by artificial intelligence over those written by human health professionals.
The finding challenges a widespread assumption in digital health: that human expertise always beats algorithmic output. The AI system generated messages tailored to where each user stood in their readiness to change behavior, while the human-written messages were generic.
My Heart Counts launched in 2015 and has since enrolled over 100,000 study participants. The platform recently released a new version that includes the AI coaching system, built on a randomized trial designed to measure whether the technology actually changes behavior rather than simply tracking motivated users.
How the AI System Works
The AI draws on behavioral psychology research from the 1980s called the Transtheoretical Model of Change. The model maps how people move through distinct stages when adopting new behaviors-from not considering change at all to maintaining new habits.
Someone who has never considered exercising needs a fundamentally different message than someone who already jogs three times weekly but risks falling back into old patterns. The AI identifies where each user sits on that spectrum and adjusts its approach accordingly.
The system is not a chatbot. It generates short, structured messages based on limited inputs: a person's stage of change, basic contextual factors like age or physical limitations, preferred language, and preferred delivery time. It does not engage in open-ended conversations or accept free-text input.
"The AI isn't making medical decisions," said Anders Johnson, project manager at the Department of Medicine. "It's generating short, encouraging activity messages using a structure grounded in behavioral science."
Why AI Messages Outperformed Human Ones
Fatima Rodriguez, associate professor of cardiovascular medicine at Stanford, said the AI's advantage came from specificity. "The AI coach was fine-tuned specifically on the transtheoretical model of behavior change," she said. "It generated more personalized messages tailored to each user's stage of change."
One-size-fits-all messaging rarely works in behavior change. The AI's ability to tailor recommendations to individual context and data inputs created messages that felt more relevant to each person.
Personalized health coaching from a human professional is expensive, time-limited, and unavailable to most people who could benefit. An AI system delivering comparable results at scale changes that equation.
The Trial Design
The new version of My Heart Counts uses a randomized crossover trial to separate genuine effect from motivation bias. After a baseline period, participants randomly receive either AI-generated, stage-matched messages or generic step reminders for one week, then switch conditions.
Because each participant experiences both approaches, they serve as their own control. This design addresses a persistent problem in digital health research: apps tend to attract people already motivated to change, making it impossible to know whether the technology works or simply tracks early adopters.
The primary outcome is objective: daily step counts, analyzed with statistical models accounting for the order in which participants experience each condition.
"The initial trial is intentionally short," Johnson said, "so it's designed to measure near-term behavior change. Longer-term sustainment is a key next question the program is looking to study."
Constraints Built Into the System
The researchers stress that My Heart Counts is not a medical device. It does not diagnose disease, recommend treatments, or replace clinical judgment. Participation is voluntary, and data sharing requires explicit permission.
The AI itself operates under strict bounds. It does not access full health records or unstructured personal data. Instead, it works with a limited set of structured inputs-activity patterns, survey responses, and stage-of-change signals-keeping outputs predictable and aligned with its narrow purpose: supporting behavior change.
Rodriguez drew an important distinction often lost in broader debates about AI in medicine. "Trust really varies a lot depending on what the AI is doing," she said. "Not all healthcare AI is the same. For low-risk behavioral coaching, the downside of a less-than-perfect output is less than for a diagnostic or therapeutic decision."
The platform is built on unusual transparency for health technology. Its code is fully open-source, and aggregated, anonymized data is shared with vetted researchers through governed access pathways rather than controlled by a single commercial entity.
What's Next
The new version runs on Stanford's Spezi framework, a modular open-source architecture that allows updates without rebuilding from scratch. This approach will accelerate innovation across cardiovascular research.
Future dashboard updates could help users reduce risk for other diseases, such as breast cancer or osteoporosis.
My Heart Counts is unusual because it functions as a public research study. Adults 18 and older in the United States and United Kingdom can download the app, consent to participate, and contribute data to cardiovascular research at a leading medical institution.
The app is available on iOS and supports English and Spanish. Learn more about AI for Healthcare and how generative AI and LLM are being applied in clinical settings.
Your membership also unlocks:
