All sections
Overview

Openness

Privacy and your data

What the AI assistant receives when you write to it, what it does not, and who processes the requests.

11 min read

In this article

To work out your load and the status of the day, the platform stores some fairly sensitive things about you: heart rate, variability, sleep, answers about how you feel, your health questionnaire. Next to all that lives the AI assistant, an assistant that explains decisions in plain language. This article says what part of that data the assistant receives, what it does not, and who processes it.

What is stored and where

Your profile, your health questionnaire answers and their outcome, any training restriction and the rule that fired, symptoms you have flagged, your cycle log and pregnancy flag, injury history, heart rate, variability, sleep, your morning check-in answers, the calculated readiness and your conversation with the AI assistant all sit in the platform's database on servers in Russia.

Strava and analytics do not receive health data. intervals.icu receives some only when plan sending is on there. Then the platform writes your threshold heart rate for running and cycling into your intervals.icu settings. If a workout's target is Heart rate, the heart rate targets of its intervals go there too, as percentages of that threshold. intervals.icu is outside Russia. The threshold written stays there after you turn sending off, and you can change it in your intervals.icu settings. Section 10 of the Policy has the details. A language model receives part of it when you write to the AI assistant. The next section covers that. We rent our working servers, our email and our cloud from contractors in Russia. Backup copies of the database, which include health data, are also kept for up to 90 days on a server in Germany. The contractors act on the operator's instructions. The list is in the Policy.

What the language model receives

The AI assistant uses language models in Yandex AI Studio. Yandex.Cloud LLC processes requests on infrastructure in Russia.

When a message is sent to the assistant, the model receives it, the recent conversation and training data. It also receives daily readiness, its reasons and the signals used to calculate it. These include resting heart rate and heart rate variability. The model also receives your last recorded weight and your threshold heart rate. The request includes current load limits, their reasons and duration. When you ask the assistant about your plan settings, thresholds, races, strength training, connections or team, the model receives this data too. Your name, email, nickname and Telegram from your account are not included in the request.

There is a second case. The platform itself writes short assistant messages to your feed: a review after a session, a reminder the evening before a key session, the plan for race week. For this it sends the model a request without a message from you. It carries the same data as a conversation, except morning check-in answers.

Morning check-in answers from recent days are sent when a check-in consent is recorded. During the pilot, the system accepts both the earlier checkin-v2 and the new checkin-v3.

New users sign the health data consent core-v12. It names new data for the model: the heart rate of your sessions and how it rose by the end of a long session. Also heart rate zones and interval targets in beats per minute, health questionnaire and screening answers, the symptoms you flagged and medical confirmations with dates. And internal codes of limit reasons and the platform's suggestions under its safety rules. The checkin-v5 consent is signed separately when the morning check-in is enabled.

In a strength session you can connect a chest heart rate monitor over Bluetooth. The model then receives only the average and maximum heart rate of that session. The second-by-second recording is not sent to it. Only core-v11 names such a monitor, earlier versions do not.

In the iPhone app you can connect Apple Health. Workouts, resting heart rate, heart rate variability and sleep are then read from the Health app, and the model receives the wake-up date and sleep duration. Only core-v12 names this source and sleep from it, earlier versions do not.

For existing pilot users with consent from core-v3 to core-v7 the model receives the same data, the new data included. Those versions do not name the new data, and core-v3 promised not to send health data to language models at all. This transfer goes beyond the earlier consents, and the missing consent has not been obtained. The system does not ask them to sign again for now. Existing users will sign core-v9 after the pilot ends. Continuing to use the service does not replace a separate consent to this processing.

We do not add your profile name or contact details to the request.

Your text is sent as written. If it contains health information or your name, the model receives it. The same applies to workout and race names.

You may send a message to the assistant yourself. An administrator may also send it when handling your support request. Content logging for improving Yandex AI Studio is turned off.

The Privacy Policy describes the actual processing. The scope of your consent is determined by the text you signed.

A voice message is first turned into text by the Yandex Cloud speech recognition service. That is not a language model. From then on the text is handled the same way as typed text.

What the platform answers itself

Some answers in the chat come from the platform itself, without the model. If you write about chest pain, fainting or another alarming sign, you get a text written in advance: stop the load and seek help.

If you write that you are tired and there is a session today, the platform builds a card itself. There you can ease the session, move it or remove it.

A chat message changes neither the readiness light nor medical restrictions. Symptoms are flagged on the day card, and how you feel goes into the morning check-in.

Explanations on screens are written by a template

When the home screen or the day card explains why a day is red or why the plan has been paused, the text is assembled by a template written in advance, with your values slotted into it. The model does not write it.

Two consequences follow. The explanation is identical under identical conditions, and it can be checked. And it keeps working even if the language model is unavailable entirely.

In the chat the assistant retells the same reasons in its own words. It takes the colour of the day, the reason and the numbers from the platform.

What the AI assistant can do

The assistant sees only the part of your data listed above and cannot change the plan directly. It does not lift restrictions, does not diagnose and does not prescribe treatment. Its answer may be inaccurate and does not replace a doctor.

What it can do is ask the engine: move, shorten, swap or remove a session, rebuild the week. The engine prepares a proposal showing before and after, you accept or reject it with one tap, and only then is the change applied, with a fresh check on the server.

Hard safety rules, a full stop and a paused plan, are applied and lifted by the calculation itself, with no involvement from the model and no way around it.

The rule is simple: the engine calculates and decides, the assistant explains. The numbers in its answer are the same ones on your session card, not invented afresh.

Consents and your control

Ordinary personal data, health data and the daily wellbeing check-in are three separate consents, not one shared tick box: for health data the law requires a separate consent, and we ask for it separately. The wellbeing check-in and the cycle log are off by default; the check-in is switched on by its own separate consent.

Any consent can be withdrawn, with no reason given. Here is what happens to the calculation when you do: that channel goes to “no data”, not to “normal”. A yellow or red day will not turn green because of a withdrawal, and restrictions already in force will not be lifted — switching off data collection is not a way to remove a restriction.

The consequences differ from one consent to another. Withdrawing consent for the check-in switches the questionnaire off and deletes the history of your answers, and affects nothing else. Withdrawing consent for health data stops the key functions: without heart rate, variability and the questionnaire the platform does not calculate load and does not build a plan, and the data itself is deleted within no more than 30 days, together with your whole conversation with the assistant. You cannot currently opt out of sending data to the model alone and keep the plan: it is one consent.

Notifications and the Strava caption are two separate switches in your profile, and both can be turned off at any time. The notification is a single text without details because it passes through Apple's and Google's servers; the Strava caption is on from the start, and your profile shows samples for every outcome, undershoot and overshoot included, so you can see exactly what goes out: training figures, nothing about health or readiness.

Viewing your data, exporting it, your account settings and deleting your account are never blocked — including by the consent-update screen. Otherwise the only way to refuse consent would be to stay in the product. How to delete your account and what stays afterwards is on a separate page.

What is kept longer than the rest

Your confirmation that you saw a doctor and were cleared to return to training goes into a separate log: the date and time of the confirmation, the date of the visit you gave, your IP address and browser details, the version, the language and the exact text you were looking at in that moment. Without this there is no way to prove what exactly you confirmed.

The log survives a withdrawal of consent and is kept for as long as your account exists, and for up to three years after it is deleted. This is stated directly in the Policy rather than done quietly. It is not used in running the service after a withdrawal: not in calculations, not in restrictions, not in the interface.

In its questionnaires the platform does not ask for medical documents, medication names or the medical reason for a restriction. It asks about one diagnosis only: osteoporosis or osteopenia. If you write about such things to the assistant yourself, they stay in the conversation. Only the user can record the confirmation, from their own account: support and administrators have no such action.

Where to find the details

The full list of data, the purposes of processing, retention periods, your rights and how to withdraw consent are in the Privacy Policy. It also says who the operator is and where to write with questions about your data.

The medical disclaimer — “the service is not a medical service and does not replace a doctor” — lives in the Terms of Use and in the consent texts. We deliberately do not repeat it on every screen in the interface: repeated at every step, it stops being read. But the route to help is always there in the interface — and how that works is described in “Safety and where medicine begins”.

Terms in this article

What to read next