How to do it
The opt-out (the choice to pull yourself out) and the temporary chat cover two different risks: the first blocks the use for training, the second also avoids saving the conversation. Do both.
Turn off training. From browser or app the path is equivalent, only the names change:
- ChatGPT: profile, Settings, data controls (Data Controls), turn off "Improve the model for everyone."
- Claude: name at the bottom left, Settings, "Privacy," turn off "Help improve Claude."
- Gemini: Google account, "Gemini Apps Activity," pause it.
Check that it's really off. Reopen the same screen after a few minutes: the switch must stay in the off position. It happens that an update to the terms resets it.
For sensitive topics, use the temporary chat. It's the mode (icon at the top, often a label like "Temporary chat") in which the conversation isn't kept in history, creates no memory, and isn't used for training. Use it every time you paste in other people's data.
Reduce upstream what you write. Before pasting, remove what identifies people: replace names and numbers with placeholders ("Client A," "amount X"). The model understands the problem just the same.
If you can't find a button because the interface has changed, ask the AI in the box: "Where is the setting now to not use my chats for training?".
The operational syntax for anonymizing a text before giving it to the AI:
I'm pasting you a text. Before answering, handle it like this:
- assume I've already replaced real names with "Client A", "Supplier B", dates and amounts with placeholders;
- don't ask me for the real data and don't try to guess it;
- work only on the structure and the generic content.
Text: [paste the anonymized version here]
A concrete example
Sara handles the payroll for a firm. She wants help writing an internal memo without exposing salaries. She turns off training on ChatGPT, then opens a temporary chat. In the text she replaces the amounts with "band 1, band 2" and the names with the roles. The AI structures the memo for her; she re-enters the real data by hand in the final document, on her own computer. No real data ever passed through the system.
When it does NOT work (and how to fix it)
If you turned off training but also want to delete the past
The opt-out doesn't remove the already-saved chats or those already used in completed training runs. Fix: manually delete the history you don't need. For material already in a model there's no extraction possible: the only lever is reducing what goes in from now on.
If the temporary chat makes you lose the context
The temporary one remembers nothing between one session and the next, so you repeat the instructions every time. Fix: keep a base text with the recurring instructions to paste at the start of every temporary chat.
A tip from someone who actually uses it
The most solid protection isn't a setting, it's a rule of your own: data that identifies a person never goes into the box, period. You anonymize first, work on the structure, re-enter the real data only in the final document that stays on your device. That way you're safe even if one day the company changes its terms.
Frequently asked questions
Is turning off the switch once enough?
Almost. The opt-out stays, but a major update to the terms of service can ask you to choose again and sometimes restores the default value. Recheck the setting every few months and after "new terms" notifications.
Is the temporary chat really safer than the opt-out?
They cover different things. The opt-out says "don't use my chats to train," but the conversation stays saved. The temporary one adds "don't save it at all." For confidential data you need both: belt and braces.
If I use a corporate account am I automatically protected?
That's the belief to correct. A business or enterprise account as a rule excludes training by contract, but it doesn't protect you from the colleague who opens your chat in a shared account, nor from uploading data you don't have the right to handle. The tool blocks training; the judgment on what to paste stays yours.