If you use a free ChatGPT, Gemini, Claude or Copilot account, here is something worth knowing: unless you say otherwise, your conversations may be used to help train future versions of the model. Not in the vague, aggregated sense — the actual chats you type, including the documents, photos and screenshots you upload. It is how these companies get the real-world conversations they need to improve their tools. And on most personal accounts, the setting that allows it is switched on from the moment you sign up.
This is not a scandal story. Every company below states this in its privacy policy, and every one of them gives you a way to opt out. The problem is that the defaults quietly favor their product over your privacy, and the switches live in menus most people never open. Here is what each one does, and where the switch is.
ChatGPT: one toggle under Data Controls
On a personal Free, Plus or Pro account, training is on by default. Open Settings from your profile picture, go to Data Controls, and switch off “Improve the model for everyone.” The setting applies to your whole account, so you do not need to repeat it on your phone. When it is off, OpenAI says your new conversations won’t be used to train its models.
Three things are worth understanding before you stop reading. First, it only applies going forward — conversations already folded into a trained model cannot be pulled back out. The earlier you opt out, the less of your history is ever eligible. Second, opting out of training is not the same as deleting your history; your chats can still sit in your sidebar while training is off. Third, there is one exception OpenAI states itself: if you give feedback on a response with the thumbs-up or thumbs-down buttons, that entire conversation may still be used for training, even with the setting off. If you have opted out, leave those buttons alone on chats you care about.
For a genuinely sensitive question, use a Temporary Chat instead. Temporary chats are excluded from training even if the main toggle is on, and they are not saved to your history.
Claude: the policy change you may have missed
Anthropic used to have the cleanest stance in the business — consumer chats were not used for training, full stop. That changed in 2025. In August of that year, Anthropic introduced a consumer training choice, shown as “You can help improve Claude,” with a decision deadline of September 28, 2025. Accounts that stayed in now have conversations retained for up to five years and eligible for model training, versus the previous 30-day retention.
Check where your account stands: on claude.ai, go to Settings → Privacy and look for the “Improve Claude for everyone” toggle. Turning it off stops future chats and coding sessions from being used for training and drops retention of new conversations back to 30 days. As everywhere else, the change is not retroactive.
Claude also has an Incognito mode (the ghost icon) for one-off sensitive questions — those conversations are never used for training, regardless of your main setting. One honest carve-out: chats flagged by Anthropic’s safety systems may still be reviewed to improve abuse detection, even with the training toggle off.
Gemini: turn off “Keep Activity” — and know about the human reviewers
This is the one people underestimate, because Gemini’s data handling runs through your Google Account rather than a simple toggle inside the app. While “Keep Activity” (previously called Gemini Apps Activity) is switched on, your conversations — and the files, photos and screenshots you upload — can be used to improve Google’s products, including the AI models behind Gemini. A subset of conversations is also read by human reviewers who judge whether responses were low quality or inaccurate, in order to suggest improvements. Google disconnects reviewed chats from your account before reviewers see them, but reviewed conversations can be retained for up to three years on a separate track.
To opt out: open your Google Account, go to Data & Privacy, find Gemini Apps Activity, and choose Turn off. From that point, new chats are not used for training or human review. Google still keeps conversations for up to 72 hours to operate the service, and its own help pages suggest treating anything you type as something a reviewer could read while the setting is on. Paid Workspace accounts play by different rules: content under Workspace terms is not used to train models.
Copilot: two toggles under Privacy
Microsoft gives signed-in Copilot users two separate switches: “Training on conversation activity” and “Training on voice conversations.” On copilot.com, it is your profile icon → your profile name → Privacy; in the Windows or Mac apps, profile icon → Settings → Privacy; on mobile, menu → profile icon → Account → Privacy. Opting out excludes future conversations from training.
A few groups are excluded from training automatically: people who are not signed in, users under 18, anyone on a work or school account, and users in a handful of excluded countries. And note the honest fine print on Microsoft’s own support page: opting out stops model training, but your conversations can still be used for other product improvements, advertising, digital safety and security purposes. Training is also independent of personalization — Copilot can keep remembering your preferences while training stays off.
What opting out will not do
Worth saying plainly, because the settings create a false sense of finished business:
- Nothing here is retroactive. Every provider is explicit about this: the change covers new conversations only. Data already absorbed into a trained model stays there.
- Training is not history. You can opt out of training and keep your saved chats, or delete your history and still be opted into training. They are independent choices.
- Safety review happens anyway. Abuse detection and trust-and-safety systems operate outside these toggles.
- Policies change. Anthropic’s 2025 reversal is the clearest reminder: a company that did not train on your chats one year may ask to the next. A yearly settings check takes two minutes.
- The real protection is what you type. Toggles are a backstop, not a guarantee. Passwords, financial details, client documents and anything you would not want read aloud should not go into a consumer chatbot at all.
None of the companies here are doing anything hidden — the settings exist, they work, and finding each one takes about a minute. The defaults just quietly favor their product over your privacy. Now you know where the switches are.
