You typed something into a chatbot you would not say out loud, then wondered who reads it later. Here is the verdict with no hedging: most consumer chatbots train on your conversations by default, and the off switch sits a few menus deep. To opt out of AI training you open each tool's privacy settings and flip one control. ChatGPT calls it "Improve the model for everyone." Gemini calls it "Keep Activity." Claude calls it "Model Improvement." Copilot calls it "Training on conversation activity." Flip it and you protect future chats only. Anything already absorbed into a model stays absorbed.
Below is the exact path for each major platform, the real difference between opting out, deleting history, and using temporary chats, plus a 10-minute checklist. Settings reflect June 2026. Privacy controls move fast, so re-verify against the linked source pages before you rely on them.
Signing out feels like the private move. It is the opposite. Use ChatGPT logged out and OpenAI still uses those chats for training and safety, and there is no opt-out toggle when you have no account to attach it to. The privacy switch only exists once you log in.
That single fact reorders the whole problem. The default matters more than the off switch. On personal accounts, ChatGPT, Gemini, and Copilot feed new conversations into model development unless you intervene. Claude makes you opt in instead. Knowing which way the default points tells you whether you even need to act.
Short answer: opt out per platform, because the consumer default usually trains
There is no single industry switch. Each company puts the control in a different menu under a different name, and the default flips by plan. Free and paid personal tiers usually default to training on. Business, enterprise, and API tiers usually default to training off, the inverse of what most people assume.
Two facts drive every decision below. Opting out is forward-looking. It removes future conversations from training, never past ones. Deleting your chat history is a separate action, because on some platforms deletion does not reach copies already routed to human reviewers. Opt-out, deletion, and temporary chats are three different tools for three different problems.
ChatGPT: the exact data-controls toggle and what it covers
Open Settings, then Data Controls, then switch off "Improve the model for everyone." On personal Free, Plus, and Pro accounts this toggle is on by default, so new conversations feed model training until you act. The change applies to new chats going forward and syncs across your devices, because it ties to your account rather than one browser.
Two details people miss. Turning off training does not delete your chat history. History and training are independent, so you can keep conversations searchable while excluding them from training. And Temporary Chats stay out of training by default, so the fastest way to shield one sensitive session is to start it as a temporary chat rather than change your account-wide setting.
Want the account-wide guarantee? Flip "Improve the model for everyone" off once. Want to shield one conversation? Open a Temporary Chat. The first holds until you change it. The second resets every session.
Gemini, Claude, and Copilot: where each opt-out lives
Google Gemini
Google renamed the old "Gemini Apps Activity" toggle to "Keep Activity" during a rollout that began in August 2025, and you still reach it from Gemini settings or at myactivity.google.com. While it is on, Google uses your activity to provide, develop, and improve its services, including training generative AI models. Turn "Keep Activity" off and future chats stop being saved to your activity and stop feeding training, unless you separately send feedback.
The off switch leaves a tail. Google keeps conversations up to 72 hours to deliver responses and maintain safety, and chats picked for human review can be retained up to three years, disconnected from your Google Account. Deleting your activity does not remove those reviewed copies. In Gemini, opting out and deleting history both leave that human-review tail behind.
Anthropic Claude
Claude points its default the other way. On consumer plans (Free, Pro, Max), Anthropic uses your chats for model improvement only if you switch on the "Model Improvement" setting in Privacy Settings, if a conversation is flagged for safety review, or if you join a program like the Trusted Tester Program. Do not relax just yet. After the September 2025 consumer-terms update, existing users hit a one-time prompt asking them to choose, and if you turn the setting on, related conversations and feedback can be retained up to five years. Incognito chats are never used for training, even when the setting is on.
Confirm your own toggle rather than assuming. Open Settings, then Privacy, and check whether Model Improvement reads on or off for your account. Commercial products, including Claude for Work and the Anthropic API, run under separate terms and are not used to train models by default.
Microsoft Copilot
Consumer Copilot splits the controls into "Training on conversation activity" and "Training on voice conversations." On copilot.com, select your profile icon, then your profile name, then Privacy. On Windows and macOS, go profile icon, Settings, Privacy. In the mobile app, open the menu, select your profile icon, then Account, then Privacy. Switch these off and your future conversations leave training.
Training is on by default for consumer Copilot in most regions, with exceptions like the European Economic Area and the UK, where it is off. Copilot also separates training from personalization, so you can opt out of AI training and still keep personalization on. Copilot then remembers details from recent conversations to tailor replies while Microsoft stops using those conversations to train its models. You lose nothing about the experience by declining training.
| Platform | Setting name and path | Consumer default | Opt-out covers |
|---|---|---|---|
| ChatGPT | Settings > Data Controls > Improve the model for everyone | On (Free, Plus, Pro) | Future chats; history kept separately |
| Gemini | Gemini Apps Activity > Keep Activity | On | Future chats; 72h + up to 3y human-review tail remains |
| Claude | Settings > Privacy > Model Improvement | Opt-in (off unless enabled) | Training off unless flagged or enabled |
| Copilot | Profile > Privacy > Training on conversation activity | On (off in EEA/UK) | Future chats; personalization can stay on |
Opt-out vs deletion vs temporary chats: what each actually does
People conflate these three actions constantly, and the confusion costs them privacy. Opting out stops future conversations from entering training pipelines. Deleting history pulls conversations from your visible account. Temporary or incognito chats stop a single session from being saved or used in the first place. Doing one does not do the others.
- Opt out of training: forward-looking. New chats leave model development. Past contributions stay where they are.
- Delete history: clears conversations from your account view, but may not reach copies already sent to human reviewers. Gemini retains reviewed chats up to three years, disconnected from your account.
- Temporary or Incognito chat: the cleanest per-session option. ChatGPT Temporary Chats stay out of training by default, and Claude Incognito chats are never used for training.
- Send feedback (thumbs up or down): can trigger use of that specific exchange even when your training setting is off, and on Claude it can store the related conversation up to five years. Skip it if privacy is the goal.
The rule of thumb is short. Opt out for ongoing protection, use temporary chats for anything sensitive, and delete history only to manage what you can see, never to undo training.
The hard limit: opting out does not unlearn past data
Opting out is not a time machine. If your conversations trained a model before you flipped the switch, that training already happened, and toggling the setting today does not pull your data out of a model already built on it. Machine unlearning, the technical work of surgically removing one person's contribution from a trained model, is still an open research problem, not a button in a settings menu.
Here is the line nobody at a chatbot company will print on the settings page: opting out caps future exposure but cannot reverse past exposure. Timing is the whole game, and architecture beats any after-the-fact switch for data you actually cannot afford to leak.
This is also why deletion and opt-out give weaker guarantees than people expect. Retention windows for safety and human review, like Gemini's up-to-three-years for reviewed chats, sit outside the training toggle entirely. Read each platform's retention language, not just its training language.
Business and zero-data-retention accounts: what changes
The defaults flip on business-grade plans. OpenAI does not train on inputs or outputs from ChatGPT Business, ChatGPT Enterprise, or the API by default, the opposite of personal accounts. Anthropic's commercial products, including Claude for Work and the Anthropic API, are also excluded from training by default and run under separate commercial terms. If you handle client or regulated data, the tier you pick does more than any single toggle.
Retention still applies even when training does not. OpenAI keeps API inputs and outputs up to 30 days for abuse monitoring before deletion, and zero-data-retention, where eligible inputs and outputs are not stored after a request completes, is offered only to qualifying enterprise customers on supported endpoints. Scope varies by provider and contract, so confirm the specifics in writing rather than trusting a marketing page.
If a chatbot is part of a paid work tool you already own, check whether your organization's tier already excludes training. Many teams pay for protection they never switch on, then run sensitive prompts through a personal account that trains anyway.
Where an external memory layer fits
Per-platform toggles control what a model does with one conversation. They do not change where your long-term context lives. If your preferences, history, and personal facts sit inside one vendor's account, your privacy posture is only as strong as that vendor's defaults and retention policy, both of which can change without your say.
This is the gap MemX (memx.app) is built for. MemX is an external, model-agnostic AI memory layer. Your context lives in one place you control and travels across assistants, instead of staying locked inside whichever chatbot you used last. It is private by architecture, with per-user isolation, encryption at rest, and on-device options. To be precise about the claims, MemX is not end-to-end encrypted and is not zero-knowledge. The point is structural. Keep memory outside the model, and a vendor's training default stops being the only thing standing between you and your own data.
A 10-minute privacy checklist (status evolves, re-verify)
Run this once across every assistant you use. It takes about ten minutes and closes the most common gaps. Because these settings change, treat the steps as a starting point and confirm each against the linked source pages.
- ChatGPT: Settings > Data Controls > turn off "Improve the model for everyone." Use Temporary Chats for sensitive sessions, and stay logged in, because the toggle does not exist when you are signed out.
- Gemini: open Gemini Apps Activity, turn off "Keep Activity." The 72-hour and up-to-three-year human-review windows still apply.
- Claude: Settings > Privacy > confirm "Model Improvement" is off. Use Incognito chats for anything sensitive.
- Copilot: profile > Privacy > turn off "Training on conversation activity" and "Training on voice conversations." Personalization can stay on.
- Audit your feedback habit: thumbs up or down can trigger use of that exchange. Skip it on private chats.
- Match the account to the data: use a Business, enterprise, or API tier for client or regulated work, where training is off by default.
- Separate the actions: opting out is not deletion, and neither unlearns past data. Plan around that limit.
- Re-check quarterly: defaults and menu names shift after product updates. A setting you turned off last year may have been renamed or reset.
01Does ChatGPT train on my conversations by default?
On personal Free, Plus, and Pro accounts, yes. "Improve the model for everyone" is on by default. Turn it off under Settings > Data Controls to exclude future chats. Temporary Chats and Business plans are excluded by default.
02How do I stop Google Gemini from using my chats to train AI?
Open Gemini Apps Activity and turn off "Keep Activity," the renamed control. Future chats stop feeding training. Google still keeps conversations up to 72 hours, and chats picked for human review can be retained up to three years, disconnected from your account.
03Does Claude use my chats for training?
On consumer plans, only if you turn on "Model Improvement" in Privacy Settings, or if a chat is flagged for safety review. Incognito chats are never used. Commercial products and the Anthropic API are excluded by default.
04If I opt out, is my old data removed from the AI model?
No. Opting out only stops future conversations from being used. Data already used in a completed training run stays in the model. Removing one person's data from a trained model is an unsolved technical problem, not a settings toggle.
05Is deleting my chat history the same as opting out of training?
No. Deleting history removes chats from your account view; opting out stops future chats from feeding training. They are separate actions, and on Gemini, deleting activity does not remove copies already sent to human reviewers.
One last reminder. The menu names and defaults above reflect June 2026. Providers rename toggles and reset defaults after major releases, so the safest habit is to re-verify against the official source pages each quarter and after any large product update.
