Which fitness apps train AI on your data?

Runna lists “Train, fine-tune, and improve internal ML or AI models” among the things it does with your activity, performance and location data. The only opt-out the policy names is for advertising. It’s a Strava company and keeps its own policy instead of deferring to Strava’s, which matters here. Strava’s points at a switch. Nothing like it appears in Runna’s.

Fitbit’s is a purpose list, not a declaration: it uses the information it collects to “develop new features and Services, which may include generative artificial intelligence models”. The sentence doesn’t say what the information is. On a device worn all day it’s resting heart rate, sleep stages, menstrual cycle logging and where you ran. No opt-out is named anywhere in the document.

Freeletics trains its Coach+ assistant on fitness data and says it uses only aggregated or de-identified data to do so. Categorically, where Strava says “where possible”. What the policy never offers is a way to keep your fitness data out of that training set, and that’s what the grade measures.

Strava trains AI and ML models on the information it collects, de-identified “where possible”. A choice about contributing exists, under Settings, then Privacy Controls, then Product Improvements. But the policy links to it instead of naming it, and never says which way the switch ships. That gap is the difference between a D and a C on this site.

Peloton builds and trains its models on your activity, reviews the results by hand, and says so. The switch is named, Personalized Experiences in the Privacy section of your Preferences, and the document says it ships on. That’s what a C is for: training on by default, where the product tells you. Voice data is the one input it says is used only if you opt in.

Garmin’s privacy policy is still silent. The answer is in its AI Transparency Statement for Active Intelligence, which says the model was trained on “a small sample of fitness data” from people who opted in to product improvement. Opt-in is what a B is. The statement describes one model, in the past tense, which is why the confidence is medium and not high.

Oura’s policy puts its own use of personal data under a heading that says “Large Language Model Training”, to develop and improve its own AI and LLM-powered features, and rules out only selling it to train somebody else’s. Whether the heart rate, sleep stages and body temperature it holds train a model, it never says. The object of every sentence is a feature, not a model.

WHOOP’s policy promises never to sell your data and sends you to another page for WHOOP Coach, the AI feature. That page isn’t in the capture. The policy itself uses “training” twice, both times meaning your workouts.

MyFitnessPal names machine learning and artificial intelligence once, as tools it uses to improve the Services. Asked, it replied that it isn’t using personal information to train proprietary models at this time. The reply is printed on the verdict. The grade reads the document, and the document hasn’t changed.

Strong’s policy never mentions AI, machine learning or model training. Its one use of “training” is weights and reps. Silence is a finding about the document, not an accusation.