What data fitness apps collect
Most users underestimate how much they give away every day.
Data category | Examples | Sensitivity |
|---|
Vital data | Heart rate, HRV, blood oxygen, breathing rate | Very high |
Activity | Steps, calories, workouts, routes | High |
Sleep | Sleep stages, sleep duration, wake times | Very high |
Body data | Weight, body fat, muscle mass | High |
Location | GPS tracks, frequent places, movement patterns | Very high |
Cycle data | Menstruation, fertility, symptoms | Extremely high |
Each of these says a lot on its own. Combined, they add up to a complete health profile, often more detailed than your medical record. That's why the GDPR counts health data among the special categories under Art. 9, which are subject to stricter rules.
Who is interested in your fitness data
The app providers obviously need your data to run their services. But they also use it for product improvement, personalised advertising and sometimes for selling aggregated analyses to third parties. What exactly happens to the data depends on what kind of AI is actually behind the app, from a simple rule set to a trained model.
For advertisers, fitness data is gold. Low sleep score? Ads for supplements. Location at the gym? Ads for local fitness offers. The link between your data and what you're being sold is no coincidence.
Insurers are showing growing interest. Health insurers offer discounts for activity data, life insurers use it to calculate risk. In some countries this is already reality, in Germany it's still limited.
Data brokers make up the invisible market. They bundle data and resell it, supposedly without any link to the individual person. In practice, though, data sets can often be traced back to a person.
Fitness data in ChatGPT, Claude and Gemini
If you use ChatGPT, Claude or Gemini to analyse your training, you're sharing sensitive health data with AI companies. On consumer plans, all three major providers use your chats for model training unless you switch it off. That applies to everyone who uses LLMs for AI training planning.
How much athletes actually trust these chatbots is something I looked at in a study on AI chatbots in fitness.
Training and retention at a glance
Provider | Training by default | Retention | Business plans |
|---|
ChatGPT (Free/Go/Plus/Pro) | On until you opt out | Deleted chats up to 30 days | No training |
Claude (Free/Pro/Max) | Your choice since autumn 2025, the switch was preset to "on" | 5 years with training on, 30 days with training off | No training |
Gemini (free/Google AI Pro) | "Keep Activity" on | 72 hours with "Keep Activity" off, chats reviewed by humans up to 3 years | No training |
Sources: OpenAI Data Controls FAQ, Anthropic: retention, Anthropic: changes to the consumer terms, Gemini Apps Privacy Hub
Which data goes into training
With training switched on, your prompts (meaning your training and health data), the AI's answers, uploaded files such as CSVs with HRV data or training logs, and your feedback all go into model training. With Gemini this also includes uploaded screenshots, for example from your Garmin or Strava app, and the transcripts when you talk to Gemini Live.
One point many people miss: thumbs-up or thumbs-down feedback is a separate channel. OpenAI can use the related conversation for training even after you opt out, and Anthropic keeps feedback conversations for up to 5 years. In sensitive chats, it's better to leave the thumbs buttons alone.
If you analyse your HRV values, sleep patterns or body composition with an LLM while training is on, this data can end up in future model versions. The providers say they filter or pseudonymise it, but you can't undo training that has already happened.
What changed in 2026
Three developments from this year matter for athletes.
In January, OpenAI introduced ChatGPT Health. It's a separate area in ChatGPT where you can connect Apple Health, MyFitnessPal or Peloton. Conversations there are stored separately from your other chats and, according to OpenAI, aren't used for training. It started with a small test group and is gradually rolling out to everyone on the web and on iOS. Connecting medical records and some of the apps is only available in the US, and Apple Health requires an iPhone. In the same month, Anthropic launched health connectors for Claude, initially as a beta in the US for Pro and Max. There too, the connected data is not supposed to go into training. None of these promises cover normal chats where you paste your data in yourself. There, your training setting is what counts.
ChatGPT now shows ads in Europe too. In Germany, for example, they started on 24 August 2026 (in German), and only on the Free and Go plans. The ads are based on the topic of the conversation. If you ask about training plans, nutrition or weight loss there, expect matching ads. Personalised ads weren't available in the EEA at launch. Plus, Pro and the business plans stay ad-free. The plan doesn't change anything about training, though: on Plus and Pro you also have to switch it off yourself.
In October, a study by IMDEA Networks showed that six of nine AI chatbots examined in the browser pass information from conversations to tracking and advertising services, such as chat titles, links to chats or even verbatim input. A chat title like "knee pain after interval training" can end up with an ad service this way. According to netzpolitik.org (in German), ChatGPT, Claude and Gemini are affected as well. The German consumer organisation Stiftung Warentest therefore recommends rejecting cookie banners and using a tracker blocker. Switching off training doesn't help against this problem, it's a separate issue.
Switching off training in ChatGPT, Claude and Gemini
ChatGPT (OpenAI)
On Free, Go, Plus and Pro, open ChatGPT, click your profile picture and go to Settings → Data controls. There, switch off "Improve the model for everyone". The easiest way is in the browser at chatgpt.com. The setting applies to your whole account, including the app. Alternatively, you can submit the request "Do not train on my content" in the OpenAI privacy portal. According to OpenAI, either one is enough, and it applies to ChatGPT and Codex. After that, OpenAI no longer uses your new conversations for training.
You can also use a Temporary Chat. These chats don't show up in your history and aren't used for training. OpenAI still keeps them for up to 30 days for safety reviews. Under Data controls you'll also find "Delete all chats". Deleted chats disappear from your history immediately, but they also stay on the servers for up to 30 days.
For Business, Enterprise and the API, training is off by default. In the desktop app you should also leave Full access switched off (Settings → General → Permissions). Otherwise ChatGPT can edit any file on your computer without asking, including training exports, lab results or doctor's letters stored there.
Claude (Anthropic)
On Free, Pro and Max, open Claude, click your profile icon and go to Settings → Privacy. There, switch off "Help improve our AI models". You can go straight to the page at claude.ai/settings/data-privacy-controls.
Important with Claude
The difference in retention is big: 30 days when training is off, up to 5 years when it's on. The 5 years apply to new and resumed chats. If you switch training off later, Anthropic no longer uses your old chats for future training. Models that have already been trained stay as they are, though. Incognito chats (ghost icon at the top right) never go into training, even if the switch is on. They're still stored for 30 days.
If the automated safety systems flag a chat as violating the usage policy, Anthropic keeps inputs and outputs for up to 2 years and the classification scores for up to 7 years. That's stated in the retention overview.
For Claude for Work, Enterprise and the API, training is off by default.
Gemini (Google)
On free Gemini and Google AI Pro, open myactivity.google.com/product/gemini. In the browser you can also get there from gemini.google.com via "Settings & help" → "Activity" at the bottom left. The page is called "Gemini Apps Activity". There, set "Keep Activity" to "Off". When you switch it off, you can also delete your existing history.
With the setting off, your future chats no longer go into training or to human reviewers. Google still stores them for 72 hours. There's one exception: chats that have already been reviewed by humans are kept for up to three years, separately from your account. They remain even if you delete your activity.
Gemini Live has its own setting for audio, video and screen sharing. It's on the same page right below the switch and, according to Google, off by default, but it's worth a quick look anyway. For sensitive conversations, use Temporary Chats. They don't go into training and are also deleted after 72 hours.
For Google Workspace, training is off by default. If your account is managed through Workspace, for example by your employer, only the administrator can change "Keep Activity".
The settings at a glance
Provider | Where | Setting | Privacy mode |
|---|
ChatGPT | Settings → Data controls | "Improve the model for everyone" off | Temporary Chat (stored 30 days) |
Claude | Settings → Privacy | "Help improve our AI models" off | Incognito chat (stored 30 days) |
Gemini | myactivity.google.com/product/gemini | "Keep Activity" off | Temporary Chat (stored 72 hours) |
Checklist for your data
Checklist
ChatGPT: switch off "Improve the model for everyone" under Data controls.
Claude: switch off "Help improve our AI models" under Privacy.
Gemini: switch off "Keep Activity" and delete the old history along with it.
Use only Temporary or Incognito chats for health data.
Don't give thumbs-up or thumbs-down ratings in chats with health data.
Remove names, dates of birth and insurance numbers from doctor's letters, lab results or training exports before uploading them.
Use chatbots in the browser only with cookies rejected and a tracker blocker.
Check shared chat links and delete them when you no longer need them.
Set up privacy zones around your home in Strava and Garmin.
Set location access for fitness apps to "While using the app".
Go through connected third-party apps and remove old connections.
Turn off ad tracking on your smartphone.
Does a paid plan change anything?
Not on consumer plans. Many people think: "I'm paying for Pro or Plus, so my data isn't used." That's not true. ChatGPT Plus, Claude Pro and Google AI Pro treat your data for training exactly like the free plans. You have to opt out yourself. The only difference with ChatGPT: Plus and Pro are ad-free.
Plan type | Training by default | Opt-out possible? |
|---|
Consumer (Free/Go/Plus/Pro/Max/Google AI Pro) | On | Yes |
Business (Team/Enterprise/Workspace) | Off | Not needed |
Only business and enterprise plans have training switched off by default. On consumer plans you're paying for features, not for privacy.
What stays stored after you switch it off
Even with training switched off, all providers keep your data for a while, for legal and security reasons. OpenAI and Anthropic remove deleted chats from their servers within 30 days. Google keeps chats for 72 hours when activity is off, and chats that have already been reviewed for up to three years. If a chat is flagged for abuse or there's a legal obligation to retain it, all providers keep it longer.
On top of that, chats aren't end-to-end encrypted. Employees and contracted service providers can have access for safety reviews, as Google explicitly states in its Privacy Hub. In a data breach, your conversations could be affected.
Treat LLM chats like semi-public spaces. Don't share anything you wouldn't mention in an office meeting.
Local models, API and privacy modes
If you want maximum privacy but still want to use AI for fitness, you have three options.
Local LLMs via Ollama, LM Studio or GPT4All run entirely on your computer, and no data leaves your device. The quality is good enough for simple analyses but below the big cloud models. On the red-flag questions in the AI fitness benchmark, it is the small models you can run at home that come last: Ministral 3 14B with 30, Ministral 3 8B with 36 and Ternary Bonsai 2 9B with 38 out of 100. The larger Gemma 4 variants that run locally score 66, level with Claude Opus 5 and Gemini 3.1 Pro (as of the benchmark from 4 November 2026). For an HRV trend analysis or evaluating a training log, Gemma is perfectly usable.
Access via the API from OpenAI, Anthropic or Google is the best compromise between quality and privacy. According to OpenAI, API data isn't used for training by default. OpenAI does keep it for up to 30 days for abuse monitoring, though. No storage at all ("Zero Data Retention") is only available on request and mainly for business customers. For most athletes, the effort of an API key and a separate interface is too much.
The privacy modes of the consumer apps, meaning Temporary Chat in ChatGPT and Gemini and Incognito chat in Claude, aren't used for training and don't end up in your history. They're not entirely storage-free, though: OpenAI and Anthropic keep these chats for up to 30 days, Google for 72 hours. This is the simplest route for sensitive analyses if you don't want to set up an API.
Your rights under the GDPR
If you live in the EU, you have strong rights, with fitness apps just as with LLM providers.
With the right of access (Art. 15) you can find out at any time what data is stored about you, where it comes from, who it has been passed on to and how long it will be kept. An email to the data protection officer or the form in the app is enough. The provider has to respond within one month.
The right to erasure (Art. 17) lets you demand that your data be deleted, with restrictions where there are legal retention obligations. With fitness apps, you delete the account and check whether all the data really disappears. In ChatGPT you'll find "Delete all chats" under Settings → Data controls, in Claude you delete your account under Settings → Account, and in Gemini you delete the history via myactivity.google.com.
The right to object (Art. 21) applies when a provider bases the processing on its "legitimate interest", as is often the case with AI training. You can then object without giving a reason. The switches in the settings are the fastest way to do that. If you want it formally documented, also write to the provider's privacy address.
Fitness apps compared on privacy
Provider | Data storage | Third-party access | Export available |
|---|
Apple Health | On the iPhone, optional iCloud sync | Only with consent | Yes |
Garmin | USA, UK, Australia (EU contracting party: Garmin Würzburg) | No sale, product improvement only with consent | Yes, full data export in the account |
Oura | Company in Finland, servers worldwide incl. USA | No sale, no measurement data to advertisers | Yes |
Strava | USA (EU contracting party: Strava Ireland) | Profile and activities public by default, ad networks (can be turned off) | Yes |
Whoop | USA | No sale, health data used for research only in anonymised form, not for third-party ads | Yes, CSV in the app |
Fitbit/Google Health | Google account (Fitbit accounts shut down since May 2026) | No health data for Google ads, noyb complaint 2023 | Yes |
MyFitnessPal | USA | Personalised advertising, marketing services such as Meta | Premium only, according to guides |
With Apple Health, the data stays on your iPhone as long as you don't sync it via iCloud. Oura states in its privacy policy that it doesn't sell data and doesn't pass measurement data to advertisers. Garmin also says in its privacy statement that it doesn't sell personal data, and your activities in Garmin Connect are private by default. With Strava you have to adjust the privacy settings actively: according to Strava, profile and activities of users aged 18 and over are visible to "Everyone" by default, including people without an account and search engines. Strava says it doesn't use health data from connected devices, such as heart rate or HRV, for advertising. Whoop says it doesn't sell data, uses health data for research only in anonymised form and not for third-party advertising. MyFitnessPal, on the other hand, relies on personalised advertising and uses services such as Meta for marketing analytics; according to several guides, the data export is only available with a Premium subscription. In 2023 the privacy organisation noyb filed three complaints against Fitbit, because new users have to agree to data transfers to the US to be able to use the app at all. I haven't found a decision by the authorities to date. Since May 2026 the Fitbit app has been called Google Health, and if you want to keep using Fitbit, you need a Google account. In the privacy FAQ for Google Health, Google commits to not using health and fitness data for Google ads.
If you still don't know what happens to your data after five minutes of searching, that's an answer too. How these apps differ technically, meaning whether they use rule-based algorithms or real machine learning, also affects how much data they need in the first place. The table reflects what the providers themselves write in their privacy policies. Whether they stick to it and comply with the GDPR can ultimately only be determined by a supervisory authority or a court.
Settings in fitness apps
With fitness apps there are four settings that make the biggest difference.
The first is location. Allow GPS only while using the app, not permanently. Set up privacy zones in Strava or Garmin so that your starting point, usually your home, isn't publicly visible.
The second is social features. Does your profile need to be public? Do activities need to be shared automatically? In most cases the answer is no.
The third is connected apps. Check which third parties have access to your health data and remove old connections you haven't used in months.
The fourth is ad tracking. On iOS you'll find it under Settings → Privacy & Security → Tracking, on Android under Settings → Google → Ads → Delete advertising ID.
With LLMs the priority is clear: switch off training, use Temporary or Incognito chats for sensitive analyses and delete old chats regularly. Don't upload raw data if a description will do. And anonymise where possible. "A 35-year-old man with runner's knee" is all the model needs, your name isn't necessary.
How to spot problematic apps
If there's no understandable privacy policy and you still don't know what happens to your data after five minutes, leave it.
Caution is also advisable with free apps that have an unclear business model. "If the product is free, you are the product" isn't always true, but often enough to take a closer look.
Also pay attention to permissions. A calorie counter app doesn't need access to your microphone.
And if there's no export function, the provider wants to lock you in. Your data belongs to you, and that should apply to every provider.
My take on ChatGPT, Claude and Gemini
You should definitely set the privacy settings before you trust a chatbot with your training data, blood values or injuries. It takes a few minutes and is the most important step.
What you can't be sure of is that nothing will be stored or analysed afterwards. The opt-out is a promise by the provider that you can't verify yourself. On top of that, the rules keep changing. In 2025 Anthropic switched to a model where the training switch was preselected. Since 24 April 2026, GitHub has used input to its coding assistant Copilot for training by default unless you object (heise). What applies to your health data today may look different tomorrow. A look at the settings every few months is worth it.
Some things aren't in your hands at all. In the copyright dispute with the New York Times, OpenAI had to keep even deleted chats for months in 2025 under a court order (Engadget). Anyone who discussed their lab results or an injury during that time and then deleted the chat couldn't rely on it really being gone. This obligation has no longer applied since the end of September 2025, but the data preserved up to then remains. And the trackers from the IMDEA study above run regardless of any training setting.
Local models only solve part of this. If you analyse your HRV exports or training logs with a model on your own computer, the data doesn't leave your device. In return, you're responsible for security yourself. Ollama, for example, has no password protection. By default it's only reachable on your own computer. But if you share it on your network for your phone or laptop and accidentally expose it to the internet, anyone can use it. In early 2026, security researchers found around 12,000 publicly reachable Ollama servers. So only download models from known sources, keep the software up to date and don't expose it to the internet.
My recommendation: set the settings, share health data such as blood values, diagnoses or cycle data sparingly and without your name, and assume that everything you type into a chat may be stored somewhere.
Pros and cons of AI chatbots for fitness analysis
Pros
Instant, personalised analyses
More flexible than specialised apps
Opt-out and privacy modes available
Cons
Data on third-party servers, not end-to-end encrypted
Training on by default on consumer plans, storage even after deletion
Policies can change at any time
Conclusion
Your fitness data is sensitive, whether it's in an app or in a chat with an AI. ChatGPT, Claude and Gemini use chats on consumer plans for training unless you opt out. Switching it off takes less than a minute per provider, and the checklist above covers the rest.
The GDPR gives you strong rights, so use them. And if a provider won't tell you what it does with your data, that's an answer too. For the full overview of AI in fitness, from the technologies to training planning and data analysis, see my guide to AI in fitness.
Sources