Your data: what is sent, what stays, the settings

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

By the end of this lesson, you will be able to

explain what is sent to the provider when you write to an AI;
tell apart two questions: “is it kept?” and “is it used to train the models?”;
find and choose your privacy setting;
open a conversation that does not stay in your history;
rewrite a request so that it contains nothing identifiable.

Why it matters to you

When you write to an AI, your text leaves your computer. Ten minutes of settings today save you from asking yourself the question with every request.

Three things to remember, even if you read nothing else:

Your “never” list, whatever the setting: never write to an AI your passwords and codes, your social security, bank card, or IBAN numbers, health information, private information about other people, or your employer’s confidential documents without permission.

A setting decides whether your conversations are used to train future models. Even with it turned off, if you give feedback on an answer (thumbs up or down), the whole conversation concerned may be used for training.

Deleting a conversation does not make everything disappear. On Claude, it disappears from your history right away, but generally stays on the servers for up to 30 days. What memory has already kept from it may remain, and on Claude, it is turned on by default: check it too (module 2, lesson 5).

Want to start right away? Go to step 1 to find your setting. Section 4 explains the details when you need them.

Before you start


You need	Your account from lesson 1, and your assistant’s ID sheet
Cost	€0
Duration	30 to 40 minutes
Important	Providers’ rules change. The information below is dated: check it again every 6 months

Understand it in 3 minutes

What is sent: your text, the files and images you add, your voice if you dictate. Everything is sent to the company’s computers (its servers), which calculate the answer.

Two questions not to mix up:

Is it kept, and for how long?
Is it used to train future models?

Four words people often confuse:

history: the list of your conversations, which you can see and delete;
retention: the copy kept on the company’s servers, for a length of time that depends on the service and on your settings;
memory: a setting that makes the assistant remember things about you from one conversation to the next; on Claude, it is turned on by default (module 2, lesson 5);
training: using conversations to improve future models.

The “30 days” you will come across in this lesson always refer to a copy on the servers, never to memory.

The details on Claude, special cases included (feedback, safety, memory)

According to Anthropic’s privacy center (Free, Pro, and Max plans), read on 24/10/2026:

Situation	What happens
Model improvement setting turned off	Your conversations are not used to train future models, except in the “feedback” and “safety” cases below, where they may be. What has already been used in a training run stays there
Setting turned on	They may be used for training, and kept for up to 5 years in a form that no longer identifies you directly
You delete a conversation	It disappears from your history right away, then from the servers within 30 days, as a general rule (the law or a security issue may require keeping it longer). What memory has already kept from it may remain: check your memory settings too and delete the relevant memories there (module 2, lesson 5)
You use an incognito chat	The conversation is not used for training, even if the setting is turned on
You give feedback (thumbs up or down)	The whole conversation concerned is kept for up to 5 years, separated from your identifier (for example your email address), and may be used to analyze the service and train the models
A conversation is flagged for a safety issue	Its content may be kept for up to 2 years and analyzed, and its safety scores for up to 7 years
Memory is turned on (this is the default setting)	Claude remembers elements of your conversations for the following ones. You manage it in the settings (module 2, lesson 5). An incognito chat does not create any

The details on ChatGPT (setting, feedback, temporary chats, memory)

According to OpenAI’s help pages (read on 24/10/2026):

a setting named “Improve the model for everyone” decides whether your new conversations are used for training; it is turned on by default on the Free, Plus, and Pro plans, and voice has its own setting.
Even with the setting turned off, if you give feedback on an answer (thumbs up or down), the whole conversation concerned may be used for training.
Temporary chats do not appear in the history and are not used for training as long as you do not save them; OpenAI may keep a copy for up to 30 days for security reasons.
ChatGPT’s memory, when turned on, may remember preferences and details from your conversations for the following ones (module 2, lesson 5).

Do it together

1Step 1: Find your setting on Claude

Why are we doing this? This setting decides whether your next conversations may be used to train Claude’s future models. It is not memory, nor the history, nor deleting your conversations. The goal: know where it is, and choose with full knowledge. No position is required.

Log in on claude.ai.
Click your name, then Settings, then Privacy. Shortcut on a computer, once logged in: the address claude.ai/settings/data-privacy-controls opens this page directly.
Find the “Help Improve our AI models” switch. This path and this label come from Anthropic’s official help, updated on 03/08/2026; your screen may look a little different.

How to choose. Turned on, you help improve the models, and your conversations are kept longer. Turned off, they are not used for training, except in two cases where they may be: if you give feedback on an answer, or if a conversation is flagged for a safety issue. Both choices are legitimate. If you hesitate, turn it off: you can change your mind later.

2Step 2: On ChatGPT, if you use it

In the settings, look for the data control section (“Data controls”), then the model improvement setting. According to OpenAI’s help pages (read on 24/10/2026), this setting is turned on by default on the Free, Plus, and Pro plans: until you turn it off, your new conversations may be used for training. Voice has its own setting.

3Step 3: A conversation that does not stay

For a one-off question that you do not want to find in your history or see used for training: the incognito chat on Claude, or the temporary chat on ChatGPT. This is not an immediate erasure: Claude and OpenAI may both keep a copy for up to 30 days, for security. So write nothing from your “never” list in it. To open one on Claude: start a new chat (outside a project), then click the ghost icon in the upper right corner. On ChatGPT: open a new chat, then select Temporary before your first message (labels taken from the official help pages of 28/10/2026, not seen on screen).

4Step 4: Make a request hard to identify

Here is an overly chatty request (the example is made up):

My neighbor Mrs. Lemaire, at 14 rue des Tilleuls, has owed me €200 since March. She is on sick leave right now. Help me write to her to ask for the money.

The same one, with nothing identifiable, and just as effective:

A neighbor has owed me €200 for 3 months. Help me write a polite but firm message to remind her, offering to be repaid in two installments.

What disappeared: the name, the address, and the health information. None of it was useful for writing a good message.

Your turn

Choose your setting on your assistant, and write your choice down in the “My privacy” section of your ID sheet, with the date.
Write your “never” list in the sheet: at least 3 things you will not share with an AI, adapted to your life (your work, your family, your papers).
Make a real request hard to identify: take a request you would like to make this week, remove everything that lets someone be recognized, then send it.

Check your work

The exercise worked if:

you know where your setting is, and you made a choice (it doesn’t matter which);
you can explain the difference between deleting a conversation and turning off training;
your request no longer contains a name, an address, an employer, or a rare detail that would let the person concerned be recognized, and the answer is still useful.

Common pitfalls:

believing a deleted conversation never existed: on Claude, it generally stays up to 30 days on the servers, and on both Claude and ChatGPT, what memory kept from it can stay until you erase it;
believing a paid plan keeps nothing: check the rules of your plan;
giving information about other people “for context”;
pasting a work document without knowing whether your employer allows it.

Keep it and pick it up later

Your ID sheet now contains your setting, the date of your choice, and your “never” list.
Set yourself a reminder in 6 months: “check my AI’s privacy settings again”.

Going further

optional

Claude’s privacy center, in English: privacy.claude.com (articles “Is my data used for model training?” and “How long do you store my data?”, read on 24/10/2026).
OpenAI’s help center: help.openai.com (articles “Data controls in ChatGPT” and “Temporary chat in ChatGPT”, read on 24/10/2026).

New words

5 words explained

Word	What it means
Server	A company computer that receives your request and calculates the answer
History	The list of your past conversations, kept in your account
Training a model	Using texts, sometimes including conversations, to build the next version of the AI
Incognito chat, temporary chat	A conversation that is not kept in your history nor used for training
Making hard to identify	Removing from a text everything that lets someone be identified: name, address, employer, rare details. Anonymization in the strict sense (CNIL, the French data protection authority) goes further: any identification must become impossible

Still to verify: the exact on-screen labels of the Claude and ChatGPT settings (taken from the official English help pages, not from a live test), the location of the incognito chat and of the temporary chat.