AI assistant privacy settings you should check on every device

Why this matters to you

Your 2-minute quick win

If you use ChatGPT, open your account settings and go to Gestion des données, then Améliorer le modèle pour tous. Turn off the data collection boxes for model improvement. You have done it right when those boxes are no longer selected.

Most AI assistants are built to learn from use. That means your prompts, replies, voice clips, or public posts may help train later models unless you change the default.

The CNIL explains that more and more companies reuse users’ personal data to train AI models. Its guidance is practical for ordinary people because it shows that opt-out controls already exist on major services.

This matters on every device because people switch between web, mobile apps, and built-in assistant panels. If you only change one app, another device may still keep the old setting.

It also matters because chat tools often keep a conversation trace, even when you did not create an account or did not type directly identifying details. A saved history can still hold enough context to point back to you.

Some assistants collect more than typed text. CNIL lists options for text training and voice training in Microsoft Copilot, which shows that different data types may need separate checks.

Public activity can count too. With X and Grok, CNIL says public data and interactions with Grok in X may be collected for training unless you change the setting in X.

What you are protecting

You are protecting the content of your conversations first. Prompts often contain names, schedules, health details, work issues, travel plans, or drafts you did not mean to share beyond the session.

You are also protecting your conversation history. CNIL notes that turning off Gemini app activity can delete past conversations and stops later conversations being saved, which shows how closely privacy and history are linked.

Your voice is another data type to protect. In Copilot, the options Formation du modèle sur du texte and Formation du modèle sur la voix are separate, so disabling one does not automatically disable the other.

Your public profile data matters as well. On LinkedIn, the setting Données pour l’amélioration de l’IA générative controls whether your data is used to train content creation models.

Your public social posts matter too. CNIL says Meta forms can cover conversations with Meta AI and public posts on social networks in public mode.

You are also protecting sensitive data that people often type without thinking. CNIL gives examples such as health information, political opinions, and trade union or religious information, and says special care is needed because such data is highly sensitive.

Finally, you are protecting control. A privacy setting is not only about secrecy. It is about deciding whether a service may keep, reuse, or learn from what you share.

What it costs you if you skip this

If you skip these checks, your prompts may be reused to improve a model. That may not expose them publicly, but it does widen the audience and purpose far beyond the help you wanted in that moment.

You can also lose track of where your data went. One service may use a clear toggle, while another may require a rights request form. If you never look, you cannot tell which applies.

Old conversations can stay available longer than you expect. CNIL says chatbot data should be kept only for the time needed for the purpose, yet many users assume short chats vanish automatically when they often do not.

You may reveal sensitive data in a free text box. CNIL warns that chatbots often let users write freely, so people can hand over health or other sensitive details that the operator did not expect but still receives.

Skipping the check can also affect what appears across devices. A web setting may not help if you keep using the assistant through another app with a different control path.

There is also a practical cost. If you later want to object, erase, or correct data, some services ask for screenshots or examples. It is much easier to avoid oversharing now than rebuild the history later.

In rare but serious cases, people may trust a chatbot too much in important processes. CNIL says a conversation without human intervention cannot by itself lead to important decisions such as refusing online credit, raising prices, or blocking a job application.

Step by step

1. Make a list of every AI assistant you use on web and mobile. Include tools you use directly, such as ChatGPT, Claude, Gemini, Copilot, Grok, DeepSeek, Mistral Le Chat, Meta AI, LinkedIn writing tools, and any assistant you access through X, WhatsApp, or social apps. Done correctly, you have a short list you can check one by one instead of guessing later.

2. In ChatGPT, open account settings, then Gestion des données, then Améliorer le modèle pour tous. Clear the boxes related to user data collection for model improvement. Done correctly, those boxes are unchecked.

3. In Claude, open account Paramètres, then Confidentialité. Turn off Aider à améliorer Claude. Done correctly, that option is no longer enabled.

4. In Gemini on the website, go to the Activité tab and choose Désactiver et supprimer l’activité under Activité dans les applications Gemini. In the mobile app, tap the profile icon, then Activité dans les applications Gemini. Done correctly, app activity is disabled and future conversation history is no longer saved.

5. In Microsoft Copilot, open account settings, then Confidentialité. Turn off Formation du modèle sur du texte and Formation du modèle sur la voix. Done correctly, both switches are off, not just one.

6. In Grok, open account Paramètres, then Contrôle de données. Clear Améliorer le modèle. Done correctly, the box is unchecked.

7. In X, open account Paramètres, then Confidentialité et Sécurité, then Grok et collaborateurs tiers. Turn off Autoriser l’utilisation de mes données publiques ainsi que de mes interactions, saisies et résultats obtenus avec Grok et xAI à des fins d’entraînement et d’ajustement. Done correctly, that option is disabled.

8. In DeepSeek on the website, open account Paramètres, then Profil, and clear Améliorer le modèle pour tous. In the mobile app, open account Paramètres, then Contrôles des données, and clear the same option. Done correctly, the box is unchecked on the device you use.

9. In Mistral Le Chat on the website, open account Paramètres, then the Préférences tab in the Le Chat section, and clear Autoriser l’utilisation de vos interactions pour entraîner nos modèles. In the mobile app, open Paramètres, then Contrôle des données et du compte, and clear Activer le partage de données. Done correctly, data sharing is off.

10. In LinkedIn, click your profile photo, open Préférences, then Confidentialité des données, then Données pour l’amélioration de l’IA générative. Turn off Utiliser mes données pour entraîner des modèles d’IA de création de contenu. Done correctly, the toggle is off in both web and app if you use both.

11. If you use Meta AI, go to Meta’s rights centre and select the service you use. Choose Comment m’opposer au traitement de mes informations, then use the form Je veux m’opposer à l’utilisation de mes informations par Meta AI for account data, or Je souhaite m’opposer à l’utilisation de mes informations par des tiers pour Meta AI for data from other sources. Done correctly, you submit the matching form for your case.

12. If you use Meta AI in WhatsApp, use the dedicated form and select Comment formuler une opposition au traitement de mes informations ?, then note that the request concerns l’IA par Meta. Done correctly, your request is clearly tied to WhatsApp interactions with Meta AI.

13. When no simple toggle exists, use the provider’s privacy or rights form. CNIL notes that some forms ask for screenshots or examples of the personal data to correct, erase, or object to. Done correctly, your request includes the example the service asks for, which makes it easier to process.

14. Before every prompt, pause and remove sensitive details you do not need to share. If a chatbot warns you not to enter sensitive data, treat that as a real boundary, not a suggestion. Done correctly, your prompt still gets the job done without names, health details, or personal identifiers.

Illustrative example

Emma uses three AI tools every week. She asks ChatGPT to clean up emails, uses LinkedIn writing help on her phone, and tests Gemini in a browser for travel planning.

One evening, she realises that her prompts often contain meeting times, client first names, and draft text from private messages. None of that feels dramatic, but all of it is still personal data.

She starts with ChatGPT because it is open on her laptop already. She opens her account settings, goes to Gestion des données, then Améliorer le modèle pour tous, and clears the boxes for model improvement.

Next, she opens LinkedIn on her phone. She taps her profile photo, opens Préférences, then Confidentialité des données, then Données pour l’amélioration de l’IA générative, and turns off Utiliser mes données pour entraîner des modèles d’IA de création de contenu.

Last, she signs in to Gemini on the web. In the Activité tab, she selects Désactiver et supprimer l’activité for Activité dans les applications Gemini. She notices the warning that this can remove past conversations and stops future history from being saved.

After that, Emma changes how she writes prompts. She stops pasting full message threads and uses placeholders instead of names. A draft that once said ‘Call Marc at 14:00 about his delayed order’ becomes ‘Write a polite delay update for a customer call this afternoon’.

The outcome is simple but useful. Her main assistants are no longer set to reuse her data for training in those places, less history is stored, and her future prompts hold less personal detail.

The near-miss

Now replay the same week with one skipped step. Emma changes ChatGPT and Gemini, but forgets LinkedIn because she thinks of it as a social app, not an AI tool.

Later, she uses LinkedIn’s writing features from the mobile app to polish a post based on a private work draft. She assumes the earlier changes cover all assistants on all devices.

They do not. Because she never opened Préférences, then Confidentialité des données, then Données pour l’amélioration de l’IA générative on LinkedIn, the separate training setting stays on there.

The concrete consequence is not a dramatic breach on screen. It is quieter. A service she still uses keeps permission to use her data for AI content model training because she treated one device and one app as if they covered the rest.

How to check it worked

Reopen each setting after you change it. Do not trust memory. You want to see the same box still unchecked or the same toggle still off after leaving and returning to the menu.

Check both web and mobile where the source says both exist. CNIL gives separate paths for several services, including Gemini, Copilot, DeepSeek, Mistral Le Chat, and LinkedIn.

Look for side effects that confirm the change. With Gemini, CNIL notes that disabling app activity can delete past conversations and disables conversation history for future chats, so missing history can be a sign the setting took effect.

Keep a small note of what you changed. Write the service name, date, and setting name. This helps if an update moves the option later.

Also test your own behaviour. Open a new chat and see whether you still feel tempted to paste names, phone numbers, health details, or full message threads. A safer setting helps, but it does not replace careful prompting.

Test yourself

Question: In Copilot, which two settings should you turn off for training? Answer: Turn off Formation du modèle sur du texte and Formation du modèle sur la voix.

Question: What happens when you disable Activité dans les applications Gemini? Answer: CNIL says it can delete past conversations and future conversations are no longer saved in history.

Question: If a service has no clear toggle, what should you do next? Answer: Use the provider’s privacy or rights form and give the examples or screenshots the form asks for.

Common mistakes

The first mistake is checking only one device. People often fix the browser setting on a laptop and forget the phone app they use during the day.

The second mistake is turning off one data type and missing another. Copilot is the clearest example because text and voice have separate training controls.

Another mistake is assuming public content does not count. CNIL says public posts and interactions may be used for training on some services, including Meta and X.

Many people also confuse chat history with training. They are related, but not always identical. Gemini shows this clearly because disabling app activity also affects saved conversation history.

A common error is pasting sensitive data just because a box looks private. CNIL warns that free text chatboxes can receive health or other sensitive information unexpectedly, and operators should warn users not to share it.

Some users stop after a failed search in settings. Yet CNIL shows that several services also offer forms for opposition, including Meta and LinkedIn, and that route matters when a toggle is missing or your data appears as a third party.

The last mistake is treating AI as a fully neutral decision-maker. CNIL says a chatbot conversation alone should not produce important decisions without meaningful human intervention and safeguards.

Level up

Your next step is to build a simple personal rule: use privacy settings first, then reduce what you share anyway. The sources show why both matter. A toggle can limit reuse for training, but careful prompting still protects you from putting sensitive details into a chat in the first place.

If you want one concrete habit, keep a short prompt template ready on every device. Remove names, replace exact dates with general timing, and describe the problem instead of pasting the raw document. That habit still works even when a service changes its menus.

For more detail on your rights to object to training reuse, see the CNIL guide at cnil.fr. For broader chatbot privacy principles, including retention and sensitive data warnings, see the related CNIL advice at cnil.fr.