What ChatGPT Remembers About You and How to Turn It Off

A conversation with ChatGPT works in a way that part of what you say lives longer than one conversation. There is model training on your chats, and there is a separate memory that collects facts about you on its own and pulls them into new conversations. These are different mechanisms, and they are turned off in different ways.

Here are three settings that close the main leak points in a few minutes, with no loss in answer quality.

Model training on your chats

In the settings there is a Data Controls section, and in it an item called "Improve the model for everyone". By default on the Free, Plus, and Pro plans it is on: new conversations may be used to train OpenAI models. You turn it off with one switch: Settings, Data Controls, Improve the model for everyone, Done.

Turning this item off doesn't make answers worse. It's purely a question of whether your data goes into the training set, and the quality of answers doesn't depend on it.

Memory between conversations

Separately from model training there is Memory, a personalization feature that remembers facts from conversations and pulls them into later ones: where you work, your clients' names, what's bothering you, what your habits are. You can see what has been collected and delete the extras in Settings, Personalization, Memory.

But there's a trap here: deleting the conversation itself from your history does not delete the memory built from it. These are two different storage systems, and they have to be cleaned separately.

Temporary Chat for one-off questions

For a conversation that should not end up in your history, in memory, or in training, there is Temporary Chat. It's the closest thing to a private mode available on consumer plans: the conversation isn't saved in history and isn't used to train the model.

It's handy to turn on before a question that involves other people's personal data, for example when checking a contract or a correspondence with a third party.

What not to tell an AI at all

The settings cover storage and training, but they don't replace common sense when entering data:

• - passwords and access keys in full, even in a temporary chat;
• - document numbers (passport, card) in plain form;
• - other people's personal data without anonymizing it, unless it's a one-off consultation in Temporary Chat mode.

For documents with other people's data there is a simple working trick: ask the AI itself to first replace names, numbers, and addresses with placeholders, and only then ask your question about the anonymized text.

Bottom line

The three settings cover different risks. Improve the model for everyone is responsible for training on your data. Memory is responsible for what ChatGPT remembers between conversations. Temporary Chat is responsible for a one-off conversation with no trace in history or memory. You need to sort them out once, and after that they are just a background that works by itself.

I write about the practical side of working with AI in the channel «Растём с ИИ» ("Growing with AI"). A related security topic: how to check code written by AI if you're not a developer.

If you need to work out how to safely build AI into your own work processes, you can start with a free consultation.