Learn

Model training

Updated

In short

  • Depending on the service and the type of account, your conversations may be used to train the next versions of an AI model.
  • A model does not file your messages away, but it can retain passages, especially rare information, and give them back.
  • The settings stop your new conversations being used for training. They do not erase what a model has already learnt.
Sam
“Your data may be used to improve our models.” I accepted without reading. What does it mean?

The sentence describes training. Here is what it covers.

Training does not happen while you type

  1. Today. You write, the AI answers. The model does not change.

  2. Then. The service keeps the conversation, under its own rules.

  3. Later. The provider prepares a new version: the model is shown huge amounts of text and draws patterns from it. That is training.

  4. A few months on. Today’s exchanges may shape the model other people use.

Are your conversations part of it? It depends on the account

  • OpenAI

    Free, Plus, Pro: may be used for training. OpenAI says it reduces personal information in its training sets.

    Business, Enterprise, Edu, API: not by default.

    Source

  • Anthropic

    Claude Free, Pro, Max: your choice. If you accept, kept for up to five years, in a form separated from your identity (source).

    Business plans: not affected (source).

  • Google

    The Gemini guide asks you not to enter confidential information you would not want a reviewer to see or Google to use to improve its services.

    Source (in French)

What a model remembers

Among many ordinary sentences that blend into the mass, one rare piece of data stands out: that is what the model tends to memorise. ordinary sentences: lost in the mass rare data: remembered first

A model does not file your messages away. But according to the CNIL, France’s data protection authority, models memorise part of the data they were trained on, and it becomes possible to extract personal data from them. Duplicated, rare or unusual data is memorised first (CNIL).

Lost in the mass

An ordinary sentence.

Rare data

An uncommon name next to a diagnosis. A unique number.

Authorities take it seriously: for the European Data Protection Board, a model is anonymous only if extracting personal data from it is very unlikely, and this is assessed case by case (EDPB).

Deleting is not unlearning

A conversation

Can be deleted.

What the model learnt from it

Much less so. With current techniques, it does not seem possible, in general, to find everything a model has memorised about a person; retraining it can cost a disproportionate amount. Research is progressing (CNIL).

What was never sent never has to be removed.

Opting out of training: useful, but not everything

At OpenAI, the “Improve the model for everyone” setting lets you opt out; it applies to your new conversations (OpenAI).

  • The thumbs. If you rate an answer thumbs up or down, the whole conversation may be used for training, even after you opt out (OpenAI).
  • Storage. Your conversations are still kept under the service’s rules.
  • Review. At Google, even with the activity setting turned off (« Conserver l’activité » on the French page), chats are still used to answer you and for safety, including with help from human reviewers (Google, in French).

Your turn: true or false?

The AI learns from my messages while I talk to it.

False. Training happens later, when the provider prepares a new version.

If I delete my conversation, the model forgets it.

False. The conversation can be deleted; what a model has already learnt from it, much less so.

I opted out of training: nobody can read my conversations any more.

False. They are still stored, and human review can continue for other reasons.

The right habit: set up your account once, then ask yourself, every time you send, whether the information is needed for the question.

Key points

  • Training means adjusting a model on a large volume of text, sometimes including users’ conversations.
  • A model can memorise passages, especially rare or repeated ones, and give them back later.
  • Once learnt, an item is hard, sometimes impossible, to remove from the model.
  • Opting out of training applies to future conversations. Storage and human review follow their own rules.

Sources

  1. CNIL, IA : Analyser le statut d’un modèle d’IA au regard du RGPD (22 July 2025) (in French), consulted .
  2. CNIL, IA : Respecter et faciliter l’exercice des droits des personnes concernées (7 February 2025) (in French), consulted .
  3. European Data Protection Board (EDPB), EDPB opinion on AI models: GDPR principles support responsible AI (opinion 28/2024, 18 December 2024), consulted .
  4. OpenAI, How your data is used to improve model performance, consulted .
  5. Anthropic, Updates to Consumer Terms and Privacy Policy, consulted .
  6. Anthropic, How long do you store my data?, consulted .
  7. Google, Guide sur la confidentialité dans les applications Gemini (in French), consulted .

This page is for information only. It is not legal advice and does not replace a professional’s view of your situation. Rules change: each source is dated with the day it was consulted.