Most AI assistants can use your conversations to improve their models unless you tell them not to. The switches exist, but each company puts them in a different place and describes them differently. This guide covers ChatGPT, Claude and Gemini, and explains what the settings do and what they leave untouched.
Settings and menu names change often, so check the current wording in your own app if something looks different.
ChatGPT
OpenAI describes the control in its Data Controls help page. Open Settings, choose Data controls, and turn off “Improve the model for everyone”. The setting applies to your entire account. Your chats still appear in your history, but they are not used to train ChatGPT.
For one-off sensitive questions, use Temporary Chat. OpenAI says temporary chats are deleted after 30 days. Team, Enterprise and Edu plans have additional data controls of their own.
Claude
On consumer plans (Free, Pro and Max), Claude uses your chats for training only if you allow it. The control is in Settings, under Privacy, with the toggle labeled “Help improve Claude”. Check that it is off if you do not want your chats used.
Claude also has an Incognito chat. Reports on Anthropic’s policy say incognito conversations are not used for training even when the training toggle is on. Reports on the 2025 policy change also say that allowing training lengthens how long chats can be kept, so check the current retention periods in Anthropic’s privacy policy.
Gemini
Google’s control is called Gemini Apps Activity, and newer versions of the app label it Keep Activity. You find it in the Gemini app under your profile picture. According to Google’s help page, human reviewers read, annotate and process some Gemini conversations to improve Google’s AI, and turning the activity off stops future conversations from being reviewed or used to improve its machine-learning technologies.
Off does not mean gone. Google says conversations are still saved with your account for up to 72 hours, and Notebookcheck’s report adds that the same window applies to temporary chats. Chats that reviewers have already seen are not deleted when you delete your activity, and Notebookcheck says Google keeps them for up to three years. The order matters: switch the setting off first, because deleting history afterwards does not recall chats that have already been reviewed.
What these switches do not do
Turning off training is not the same as making a chat private. The companies can still process your messages to answer you, protect their services and meet legal obligations, and your chats may stay in your account until you delete them.
The settings also only apply to future conversations. Anything you shared before you changed the toggle may already have been used, so decide early, not after the fact.
Advertising is a separate question. OpenAI has said it will test visual ads in ChatGPT, and the data controls help page notes that there are separate controls related to ads, so look for them if that matters to you.
Habits that matter more than any toggle
The most reliable protection is what you type. Keep identity numbers, bank details, passwords and medical records out of chatbots you do not fully trust, and use temporary or incognito chats for anything sensitive. Delete conversations you no longer need, and review the settings again after major app updates, because defaults and menu names change.
A Proton survey of 4,014 chatbot users found that 44% would share more if they were guaranteed their chats would not be used for training. The settings above give you a version of that guarantee, within the limits described here.
If you want a different approach
Settings let you limit what a mainstream assistant does with your chats. A different route is an assistant built around privacy from the start, and Proton’s Lumo is one of them. Proton says it uses zero-access encryption, keeps no logs and does not train on conversations, and its code is open source, so the claims can be checked by outsiders.
That is a different kind of protection from a toggle: it is part of how the product works, not something you must remember to switch off. Your prompts still reach Proton’s servers, so you are placing trust in a different company, but one that publishes its design and its code for inspection. The free plan is enough to judge the answers yourself before you consider paying for Lumo Plus.
If you want to try it, you can start through this Proton link.
