Data privacy in AI chat apps

1 · What it is

When you chat with an AI assistant, your words go to the company that runs it. Data privacy here means knowing what that company may do with them, and which settings you control. This page sticks to four providers: OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini) and Microsoft (Copilot). Their rules differ, and they change, so check the provider’s own page first.

The big question for most people is training. Training is how a company teaches an AI model, using huge amounts of text. On consumer plans, meaning the free and personal paid ones, your own chats can sometimes become part of that text.

A second question is retention. Retention means how long a company keeps your chats before deleting them. A company can keep a chat for a while even when it does not train on it.

ChatGPT. In ChatGPT, the switch is called Improve the model for everyone. To turn it off, open Settings, then Data controls, pick that switch, turn it off and press Done. OpenAI leaves temporary chats out of model improvement, but may keep them for up to 30 days for safety. One catch: if you give thumbs up or down feedback in ChatGPT, the whole conversation may be used for training.

Claude. Anthropic asks Claude Free, Pro and Max users to choose whether their chats can be used for training. They can change that choice in Privacy Settings at any time. Anthropic says it keeps data for five years if you allow training. If you do not allow training, it keeps its 30-day period. Anthropic says a chat you delete is left out of later training. Incognito chats in Claude stay out of model improvement, even when your training setting is on.

Gemini. For Gemini, the Keep Activity setting controls whether future chats are used to train Google’s AI models. Google keeps Gemini temporary chats out of its AI model training. Even with Keep Activity off, Gemini still saves new chats for 72 hours so it can reply, handle feedback and keep people safe. Gemini deletes your activity after 18 months by default. You can pick a shorter or longer period (three months or three years) or switch automatic deletion off.

Copilot. Microsoft’s consumer Copilot lets signed-in users decide if their chats may help train the model. Microsoft says a new Copilot app became available on August 18, 2026. This setting is described only for the older app. On the web or the new app, Microsoft points you to a newer privacy page instead.

What “human review” means. Some companies have real people read a small share of chats to help improve their services. Google says human reviewers read a subset of Gemini chats to help improve its services. Google says it removes the link between a chat and your account before outside reviewers see it. Deleting your activity does not remove chats a reviewer has already seen. Google may hold those for as long as three years. Turning off Keep Activity stops future Gemini chats being reviewed this way.

Deleting your account. If you are signed in to ChatGPT on Free, Go, Plus or Pro, you can remove your account in settings. Anthropic says that if you delete your Claude account, your data is left out of future training.

Work accounts follow different rules. Microsoft says prompts and responses in its work Copilot are not used to train foundation models, meaning the big general models that apps like Copilot are built on. Anthropic’s training change does not apply to Claude for Work, which includes Team and Enterprise, or to its API. OpenAI says that by default it does not train on chats from ChatGPT Business, Enterprise and Edu workspaces. In ChatGPT workspaces, your organisation’s own privacy and retention rules also apply.

Finally, watch what you paste in. Google asks people not to enter confidential information they would not want a reviewer to see. Microsoft says you should not give Copilot confidential or sensitive personal data you would not want it to use. So leave out passwords, bank details and ID numbers.

For example, say you want help writing a complaint to your phone company. You could paste the whole bill, account number and all, or just say the bill went up without your agreement. The second version gives the assistant nothing it does not need.