Do ChatGPT, Claude, and Gemini train on your data?
On personal plans, ChatGPT, Gemini, and Perplexity use your conversations to train on your data by default, and Anthropic lists training on Claude’s personal plans as opt-out, a choice users make when they sign up. Since August 18, 2026, Microsoft’s new Copilot app does not use personal conversations to train its foundation models. On business plans, none of the five trains on your content by default.
Does your AI assistant train on your conversations?
Usually yes on personal plans, and no on business plans, with one recent exception and several fine-print caveats. This page is the training part of what AI does with your data, based on each company’s help pages as of November 10, 2026; Google’s privacy hub for Gemini is the most detailed of them.
“Training” means using your conversations to improve future models. It is separate from storing your chats so you can see your history, and from reviewing them for abuse or safety, which happens whatever your training setting.
At a glance
| Assistant | Personal plans | How to turn it off | Business plans |
|---|---|---|---|
| ChatGPT | On by default (Free, Plus, Pro) | Settings, Data controls, “Improve the model for everyone” | Not by default |
| Claude | Opt-out; you choose at sign-up | Settings, Privacy, “Help improve our AI models” | Not by default |
| Gemini | On by default from age 18 (“Keep Activity”) | Settings & help, Activity, Turn off | Not used where Gemini is a core Workspace service |
| Microsoft Copilot | New app: not used to train foundation models | No training setting in the new app | Not used |
| Perplexity | On by default (Free, Pro, Max) | Account settings, Preferences, “AI data retention” | Never on Enterprise |
On ChatGPT, the setting applies to your whole account, on the web and on mobile, once you are signed in. Anthropic introduced its choice on August 28, 2025, and existing users had to decide by October 8, 2025.
What turning it off does not change
- OpenAI: feedback still counts. A thumbs up or down can make the entire conversation available for training, and voice and Codex have their own settings.
- Anthropic: training runs that have already started, and models already trained, keep your data. Conversations flagged by safety systems can still be used, and rated conversations are stored for up to five years.
- Google: future chats are still kept for 72 hours and still used for safety, with human reviewers. Turning off Keep Activity also disconnects Gmail, Drive, and other Workspace apps from Gemini.
- Perplexity: only future data is covered; data already collected for training cannot be deleted.
Temporary and incognito chats
| Assistant | Mode | Kept for | Used for training? |
|---|---|---|---|
| ChatGPT | Temporary chat | Up to 30 days | No, while it stays temporary |
| Claude | Incognito | 30 days | Never |
| Gemini | Temporary chat (personal accounts) | 72 hours | Never |
| Perplexity | Incognito | Expires within 24 hours | Not stated |
These modes keep a conversation out of your history, not out of the company’s systems. In Claude Team and Enterprise, incognito chats appear in the organization’s data exports.
Who reads your conversations
OpenAI says authorized staff and trusted contractors may access conversations for abuse investigations, support, legal reasons, or model improvement unless you have opted out. Google says some Gemini chats are reviewed by people, including contractors, after being disconnected from your account. Anthropic says shared data is de-identified before review by a small number of people. Microsoft says it has turned off human abuse review for its business Copilot services.
Copilot changed in August 2026
Microsoft launched a new Copilot app on August 18, 2026. In it, prompts, responses, and uploaded files from personal accounts are not used to train foundation models, and the training switch has disappeared from the privacy controls. The older app trained on conversations by default, with an opt-out. Microsoft says you must sign in to the new app before December 31, 2026 to keep your old conversations.
Five rules for work
- Do client work on a business plan, where training is off by default.
- On a personal account, turn training off before your first work conversation.
- Do not rate answers in a conversation that contains anything sensitive.
- Use a temporary or incognito chat for a one-off sensitive question, knowing it is still kept for a while.
- Delete what you no longer need; OpenAI and Anthropic say deleted chats are erased from their systems within 30 days.
Before you paste client data
Training is only one risk. Our page on putting client information into AI covers the rest.
If you use a work account
On a business plan, the company does not train on your chats, but your employer may read them. See what your employer can see.