FR Newsletter

AI for Life

Protect your personal data when you use an AI assistant

An AI assistant feels like a private notebook. It is not quite one: what you type is stored by a company, may be read by reviewers in some cases, and may be used to train future models depending on your settings. The good news is that you keep a lot of control. Here is what to check, in plain terms, for ChatGPT, Claude and Gemini.

What happens to what you type

By default, a consumer AI assistant saves your conversations in your account so you can find them again. Depending on your settings, the provider can also use them to improve its models. Google also explains that human reviewers read some Gemini chats, and asks users not to enter "confidential information that you wouldn't want a reviewer to see".

So there are three separate questions: is the chat saved in my history, can it be used for training, and how long does the provider keep it anyway? Each assistant answers them with different settings.

Screenshot coming after our hands-on test
Screenshot of the privacy settings in ChatGPT, Claude and Gemini: coming after our hands-on test.

1. Decide what you never paste

The simplest protection is the text that never leaves your computer. Keep out of any chat:

  • Identity and access data: passwords, bank card or account numbers, ID or passport numbers, social security numbers.
  • Health, legal or financial details about you that you would not share with a stranger.
  • Other people's data: a colleague's email, a client file, a family member's medical letter.
  • Work secrets: contracts, internal figures, anything your employer treats as confidential.

When you need help with such a document, remove or replace the sensitive parts first:

Here is a letter I need to rewrite. I replaced the names with [PERSON A] and [PERSON B] and removed the account numbers. Keep these placeholders as they are in your answer.

2. Check the training setting

Each provider lets consumers choose whether their conversations help train its models. The names below are those used on the official pages at the time of writing:

  • ChatGPT: in Settings, then Data controls, the option is called Improve the model for everyone. When it is off, OpenAI says your new conversations won't be used to train its models, and they can still appear in your history.
  • Claude: the choice is in your Privacy settings. Anthropic says it uses your chats to improve its models if you allow it there, and also when a conversation is flagged for safety review, whatever your setting.
  • Gemini: the setting is called Keep Activity. Google says that when it is off, your future chats are not used to improve its AI models unless you send feedback, and are not reviewed to improve its services.

One trap applies to all three: feedback. OpenAI states that a thumbs up or down may lead to the whole conversation being used for training even after you opted out. Anthropic keeps the whole conversation linked to a feedback for up to 5 years. Think before you click.

3. Use a temporary chat for sensitive questions

All three assistants offer a mode for one-off conversations:

  • ChatGPT temporary chat: not in your history, no memory created or updated, not used to improve OpenAI models, and may be kept for up to 30 days for safety purposes.
  • Claude incognito chat: Anthropic says incognito chats are not used to improve Claude, even if model improvement is on in your settings.
  • Gemini temporary chat: not used to train Google's AI models, and kept with your account for 72 hours.

"Temporary" does not mean "nothing is kept". It means the conversation stays out of training and out of your long-term history. The rule from step 1 still applies.

4. Clean up your history

Turning off training does not erase anything. OpenAI says explicitly that it does not delete or hide your saved chats: delete a chat, or your account, separately. Google lets you delete Gemini activity at any time, and lets you choose an auto-delete period. One limit is worth knowing: Google states that chats already reviewed by humans are not deleted when you delete your activity and are kept for up to three years.

A good habit: once a month, open your history and delete the conversations that contain anything personal.

5. Export your data before you leave

You have the right to get a copy of your data. OpenAI states that signed-in users on its personal plans can request a copy from the ChatGPT settings or through its Privacy Portal. Google says you can export your Gemini information. For Claude, look for the relevant article in the Anthropic Privacy Center. Exporting is also a good way to see what a service actually keeps about you.

6. At work, follow your organisation's rules

A work account can follow different rules. OpenAI says that by default it does not use content from its Business, Enterprise and Edu workspaces for training, and that your organisation's own retention and access policies apply. The CNIL, the French data protection authority, recommends that organisations set internal rules listing authorised and forbidden uses, and in particular forbid giving certain confidential or personal data to these tools. If such a charter exists where you work, it wins over any personal habit.

What to watch

  • Settings change. Names, menus and retention periods evolve often. Check the official pages linked below before relying on this guide.
  • Plans differ. Free, paid and business accounts do not always have the same options.
  • Connected apps count too. If you link your email, files or calendar to an assistant, read what each provider says about that data.

Tools in this guide

Questions

If I turn off training, are my old chats deleted?

No. OpenAI states that turning off Improve the model for everyone does not delete or hide saved chats. Deleting a chat or your account is a separate action, in every assistant.

Is a temporary or incognito chat completely private?

No. It is kept out of training and out of your history, but the provider still keeps it for a short time, for example up to 30 days for safety at OpenAI and 72 hours at Google, according to their pages.

Can I paste a colleague's or a client's email into an assistant?

Avoid it. It is someone else's personal data, and your employer may have rules. The CNIL advises organisations to define which personal or confidential data must never be given to these tools.

Sources

  1. OpenAI Help Center, Data controls in ChatGPT
  2. Anthropic Privacy Center, Is my data used for model training?
  3. Google, Gemini Apps Privacy Hub
  4. CNIL, Questions and answers on the use of a generative AI system (in French)