Turn off AI training in ChatGPT, Claude, Gemini and 5 more

Turning off AI training takes seven switches across six assistants, and two have none

The count as of 21 October 2026: one switch each in ChatGPT, Claude, Gemini, Perplexity and DeepSeek, and two for Grok (one in the Grok app and on grok.com, another inside X). Microsoft says its updated Copilot app doesn't use chats to train foundation models, so there is nothing to switch off. Meta AI offers only an objection form, in the EU. That squares with our check of nine chatbots, seven of which train on your chats without asking.

Each row below was read from the vendor's own help center, FAQ or policy on 26 September 2026 and checked again on 21 October 2026, for consumer accounts. The last column is what the vendor still keeps after you switch off, because a switch without a retention number is half a fact.

Three rows changed recently. Microsoft's pages for the Copilot app updated on 18 August 2026 say prompts, responses and files "aren't used to train foundation models". They list memory, chat history and ad controls, with no training switch, and the two switches survive only in the older app. Google renamed "Gemini Apps Activity" to "Keep Activity" in a post dated 13 August 2025, so any guide that names the old setting points at a screen with a new label. And anyone using Grok without signing in, outside the EU and UK, "won't have the option to opt out of model training", in SpaceXAI's own words.

One more switch sits outside the chat apps. LinkedIn trains the generative AI models behind features like its writing assistant on your profile details and public posts unless you turn off Data for Generative AI Improvement, which its help center says "is set to 'on' by default". The path is the Me icon, Settings & Privacy, Data privacy, then that setting (it also opens directly at linkedin.com/mypreferences/d/settings/data-for-ai-improvement). LinkedIn says private messages are not included, that turning it off "does not affect training that has already taken place", and that feedback you give can still be used.

How do I turn off ChatGPT training on my data, and do I also need the Privacy Portal?

Turning off "Improve the model for everyone" is enough on its own, and the Privacy Portal request is an alternative, not a second step. OpenAI's help center, updated on 26 September 2026, says so directly: "Either option is sufficient for ChatGPT conversations and Codex tasks. You don't need to do both."

1. On the web, open your account menu and select Settings.
2. Select Data controls.
3. Select Improve the model for everyone, turn it off, and select Done.
4. On iOS or Android, open the sidebar, tap your profile icon, select Data controls and turn the same switch off.

Signed in, the choice follows your account to every device. Signed out, it is saved in that browser only. Codex users have one more setting: "Include environments" in Codex settings decides whether context from your Codex environments can train OpenAI models, and neither the ChatGPT switch nor the Privacy Portal changes it.

What the switch does, in OpenAI's words: "When Improve the model for everyone is off, your new conversations won't be used to train OpenAI models. They can still appear in your chat history." Nothing is deleted, and archiving a chat changes nothing about training. The switch dates from 25 April 2023, when OpenAI added it with a 30 day retention window for abuse review. The API never needed it: OpenAI's developer docs state that "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models", and Business, Enterprise and Edu workspaces are excluded by default.

Does turning off AI training stop everything? No, four things keep running

No: at every vendor the switch stops future training on new chats, and four things carry on regardless. The first one undoes the switch without warning.

A thumbs up reopens training. OpenAI: "If you choose to provide feedback, such as selecting thumbs up or thumbs down on a response, the entire conversation associated with that feedback may be used to train OpenAI models." Anthropic: "If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models." Google says future chats with Keep Activity off will not train its models "unless you choose to send Google feedback", and X says feedback on Grok lets it "use that conversation" to train. Of three top-ranking opt-out guides we read on 26 September 2026 (Cloaked, TrustScan and The Slow AI), only one mentions this, and for X alone.

It is not retroactive. Anthropic: "Your data will still be included in model training runs that are already in progress, or in models that have been trained." Perplexity says previously collected training data "cannot be deleted or removed". A model that learned from a chat does not unlearn it when you flip a toggle. Anthropic adds one more exception: conversations its safety classifiers flag "may still be used to improve our internal trust and safety models".

It is not a private mode. The switch keeps your history and stops training; the private mode drops the history too. A ChatGPT temporary chat stays out of history and memory and "may be retained for up to 30 days for safety purposes". Claude's Incognito chats are excluded from training even with the setting on. Grok's Private Chat, the ghost icon at the top right, is deleted within 30 days.

It does not touch memory. OpenAI: "Memory controls how ChatGPT personalizes responses. Model-training controls determine whether eligible content can be used to improve OpenAI models. These are separate settings." Microsoft says the same of Copilot: you can opt out of training and keep personalization on. How ChatGPT memory works covers the other half.

The rule that follows: the switch for chats you keep, a private mode for chats you would rather not have kept, and a business or API plan for anything your employer would rather the vendor never saw. A vendor that does not train on a password still received it, so the short list in what never to paste into AI applies with every switch off.

Does Whizi train on your chats?

No. Whizi's data and privacy page says: "Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models. Whizi does not sell that content as training data." The limit is the provider: a message you send to Claude reaches Anthropic and a message you send to GPT reaches OpenAI, under that provider's terms, as where Whizi stores data lays out. Whizi has no per vendor switch and no documented temporary chat mode, so the switches in the table above remain the way to control each vendor's own app.