In depth
Your privacy and AI assistants
Where your words go, who trains what with them, what never to paste, your rights, and the settings to change tonight.
Reviewed on August 22, 2026
Where do your words go when you type?
Everything you type, paste or dictate to an assistant is sent to the provider's servers, most often in the United States, sometimes in Europe: encrypted on the way, but readable on arrival, because the model needs plain text to answer. The conversation is then kept in your history, tied to your account, until you delete it or the provider erases it at the end of its retention period. It can serve three purposes: giving you the service (always), improving the product or the models (depending on the service and your settings, see below), and, for a sample, human review for quality or safety. So an assistant is neither a private diary nor an end-to-end encrypted messenger: it is an online service that keeps records.
Read the discovery version
When you write to an AI, your sentence doesn't stay in your phone: it travels over the Internet to a huge building full of computers, often far away, sometimes in another country. There, the machine reads your question, builds the answer and sends it back. But your sentence doesn't vanish by itself: it is filed in a drawer with your name on it, so you can find the conversation again tomorrow. And a few people who work there may sometimes open the drawer, to check that the AI is answering well.
Who learns from your conversations?
At the time of writing (August 2026), most consumer assistants use your conversations to improve their models, by default on free and personal plans, with a setting to refuse. ChatGPT: "Improve the model for everyone", under Settings > Data Controls, on by default for Free, Plus and Pro; Temporary Chats do not feed . Claude: since autumn 2025, Free, Pro and Max accounts choose a training setting; if you accept, conversations may be kept for up to five years, otherwise thirty days. Gemini: as long as "Keep Activity" is on, your exchanges may be used for training and a sample is reviewed by humans; temporary chats are excluded. Le Chat (Mistral): the free plan is used to improve the models unless you turn it off in your account settings; Team and Enterprise are excluded by default. Copilot (Microsoft): "Model training on text" and "on speech", on by default for signed-in consumers, under Profile > Privacy. Meta AI: your exchanges with the assistant are used to train the models, in Europe too since 2025, whereas your person-to-person WhatsApp messages stay end-to-end encrypted and out of reach. In every case, business and education plans and access are excluded by default.
Read the discovery version
Remember: an AI learns by reading enormous amounts of text. Some companies add to its lessons... the conversations of the people who use it! So your question from yesterday may help make the AI stronger tomorrow. It is not a secret: it is written in the settings, and there is almost always a little button to say "no thanks, don't learn from my sentences". You just need to know it exists, and ask a grown-up to find it with you.
What you should never paste
Four families of things never to paste into a consumer assistant. Secrets: passwords, keys, card numbers, bank details, one-time codes, even "just to reword an email". Health data and other sensitive data (opinions, religion, orientation, criminal record): yours, and even more so a relative's who never asked. Minors: a child's first name, school, photos, difficulties; most services require users to be 13, sometimes 18, and an adult should not describe an identifiable child there. Your employer's or your clients' data: source code, contracts, customer files, meeting notes; without a business plan approved by your IT department, it is a leak, not a time saver. The test that saves you: before sending, anonymise (names, amounts, addresses replaced by Xs); if the answer is still useful, you never needed the real thing.
Read the discovery version
Imagine talking to someone very kind and very clever, but who lives far away, writes everything down, and works in a building where other people can read their notebooks. You would tell them about your school day, of course. But the door code, Mum's password, a photo of your ID card, the doctor's letter? No! Simple rule: what you wouldn't shout in the playground, you don't type into the AI. And many assistants are not made for children: ask an adult first.
Your rights: GDPR, regulators and the AI Act
The GDPR applies to AI assistants as to any online service that processes your data. You can request access to your data (Article 15), its erasure (17), object to processing based on legitimate interest such as model (21), and get it back in a readable format (20); providers expose these requests in their settings or through a privacy form, and must answer within a month. If no answer comes, you turn to your country's authority: the CNIL in France, the AEPD in Spain, the ICO in the United Kingdom. Since 2 August 2026, Article 50 of the EU AI Act adds a transparency duty: a system that converses with you must tell you clearly that it is an AI, and generated or manipulated content (images, audio, video) must be flagged as such.
Read the discovery version
In Europe, a big rule called the GDPR says your information belongs to you. You can ask a company: "show me what you know about me", "erase it", or "stop using it". And if it turns a deaf ear, there is a referee, a data protection authority in each country, you can complain to. Another brand-new rule also says an AI must always warn you that it is an AI talking: no pretending to be a real person.
The settings to change tonight
Five habits, from the simplest to the most committed. 1) Turn off : ChatGPT, Settings > Data Controls > "Improve the model for everyone"; Claude, Settings > Privacy, training setting; Gemini, myactivity.google.com > Gemini > "Keep Activity", plus auto-delete at 3 months; Copilot, Profile > Privacy > "Model training"; Le Chat, account settings > data sharing. 2) Use the ephemeral mode for sensitive topics: Temporary Chat on ChatGPT and Gemini, incognito on Claude and Le Chat; it feeds neither history nor memory, even though it is still kept a few weeks for safety. 3) Tidy up: delete old conversations, turn off "memory" where it exists, review connected apps. 4) At work, insist on the business plan (ChatGPT Team or Enterprise, Claude for Work, Copilot with a work account, Le Chat Enterprise): it excludes training by contract and keeps data inside the organisation's perimeter. 5) For the truly confidential, a local model: Ollama, LM Studio or GPT4All run an open model (Llama, Mistral, Gemma, Qwen) on your own computer; nothing leaves the machine. One hour on a Sunday covers the first four.
Read the discovery version
Here are the easy moves, like locking the door at night. One: ask an adult to find the "don't use my conversations to learn" button in the settings and switch it off. Two: for a question that is nobody's business, use the "temporary chat": it keeps almost nothing, like a drawing in the sand the sea washes away. Three: never give your real name, your address or your school. Four: if you are not sure, ask before sending, never after.