Personal privacy · Practical guide
Conversations with an AI: telling history, memory and data reuse apart

Before describing a personal situation to an AI assistant, separate three questions: where to find the conversation again, what information is used to personalise the answers, and whether your conversations may be used to improve the models. A single switch does not necessarily answer all three.
Identify the service and the account concerned
This guide uses personal ChatGPT and personal Gemini as documented examples. Work or school accounts may be subject to administered settings and different terms. Note the service, the plan and the account displayed before following a procedure.
Cybermalveillance.gouv.fr, the French government cybersecurity assistance platform, encourages you to limit the data you give to assistants and to keep a critical eye on their answers. It also advises adjusting your account’s privacy settings to reduce how long conversations are kept and their reuse for training, and clearing your history regularly. For a writing or organisation question, start with a fictitious example or a reduced extract. The privacy setting comes next; it does not exempt you from choosing what you share. Advice on using AI assistants (in French).
History: finding or removing a conversation
The history is the list of conversations you can find again. Hiding or archiving a conversation does not automatically mean deleting it. According to the OpenAI help page cited, turning off the use of new ChatGPT conversations to improve the models leaves the history available. ChatGPT data controls.
In personal Gemini, the “Keep Activity” setting controls the recording of activity as described by Google. For users aged 18 or over, Google states that this setting is on by default and that activity older than 18 months is deleted automatically, with the option of choosing 3 or 36 months or turning this deletion off. The Activity page lets you view and delete all or part of the activity. Even when the setting is off, conversations remain saved in the account for up to 72 hours, without appearing in the activity. Deleting a shared chat does not delete its public link, which has to be removed separately. For a work or school account, this setting is controlled by the Google Workspace administrator. Managing Gemini activity.
Memory: checking what personalises the answers
OpenAI distinguishes between conversations and the information used by memory. Deleting a chat does not necessarily delete a memory saved separately; asking the assistant to stop mentioning a piece of information does not delete its source. The options vary from one account to another. How ChatGPT memory works.
Our method: identify the information you want to correct, look at where it appears and use the controls for that source. An answer from the assistant saying it has “forgotten” is not, on its own, a complete inventory of the data kept. Check the settings and the items available in the account.
Reuse and temporary chats: reading what they cover
In personal ChatGPT, the model improvement setting concerns the use described by OpenAI for new conversations. It does not turn your chats into purely local files. According to the OpenAI help page on temporary chats, these conversations may be kept for up to thirty days for safety reasons, and a personalisation choice allows a temporary chat to use an existing memory without creating a new one as long as it remains temporary. What temporary chats cover.
Do not transpose these conditions to Gemini or another provider. Check that provider’s own policy on retention and reuse. If you cannot establish which rules apply, do not hand over the sensitive document in the meantime.
Run a check without personal data
Create a harmless test conversation and locate the controls available: history, memory, data use. Note what each action is announced to cover and the visible result, without trying to infer how the service works internally from a single answer.
For a professional file, the decision goes beyond personal settings. The guide on using AI with a client file covers the scope, the contract and the data that may be shared. Here, the expected result is simpler: knowing which setting you changed and which question it does not resolve.
Choosing the setting according to your goal
To stop seeing a conversation in the list, look at the history controls. To correct a piece of information used in the answers, look at the personalisation sources. To act on model improvement, check the data controls of the plan you use. These goals can be combined, but their result must be checked separately.
Prepare a short note before changing anything: service, account, information concerned, goal and check carried out. Do not copy the sensitive content into it. If you work in several accounts or workspaces, check which one is active; a setting observed in your personal space proves nothing about a work space.
A temporary chat can change status
According to the OpenAI help page cited above, saving a temporary chat turns it into an ordinary conversation. The account settings then apply to that saved conversation. Before choosing this option, review what you want to keep and the personalisation and model improvement settings.
Another limit: if an action sends data to a third party, that recipient’s rules apply to the information received. The duration announced for temporary chats is not a promise of deletion by that third party. For simple wording work, a redacted description of the problem is often enough: remove the names, documents, identifiers and case details that are not needed for the question.
Sources and limitations
This guide draws on the public sources below. It explains documented features and a suggested method, without testing your account or device. Settings may vary by version and plan. Legal information concerns France and the EU where specified and is not individual legal advice. Sources may be in French.