You type a question into ChatGPT. Or ask Claude to rephrase your letter. Or ask Gemini what to cook for dinner.
Quick counter-question: Have you ever wondered where that input actually ends up?
Probably not. Most people don't. And that's fine — but it's worth thinking about at least once.
The short answer: it depends
Whether your chats are used to train new AI models depends on two things: the provider, and whether you're using a free or paid account.
The rough rule of thumb: Free = you pay with your data. Paid = usually not, or at least opt-out is available.
That sounds cynical. It is, a little. But it's true.
OpenAI (ChatGPT): The big one
OpenAI uses conversations from free users for training by default — at least if you don't change anything. In the settings there's an option called "Improve the model for everyone" that you can turn off. Somewhat hidden, but it's there.
With a paid ChatGPT Plus subscription, your history isn't used for training by default — unless you explicitly agree to it.
API usage (when developers build the AI into their own apps) is training-free by default anyway.
Anthropic (Claude): A bit more transparent
Anthropic — the company behind Claude — states that free chats can be used for safety training and model improvement. Paid accounts and API requests, according to their privacy policy, are not trained on.
What Anthropic does relatively well: they explain why they use data — mainly to detect abuse and find safety issues. Whether that's better than no data use at all is for everyone to decide.
Google (Gemini): Worth knowing
Google is Google. That means: if you use Gemini in free mode without opting out, human reviewers can read your inputs — for quality assurance, they say. It's in the terms of service. Small print, but it's there.
Gemini Advanced (the paid subscription) is supposedly more restrictive. And using Gemini through Google Workspace — in a corporate setting — comes with yet another, stricter set of rules.
What does "training" even mean?
Quick explanation: AI models learn from massive amounts of text. If your inputs flow into that dataset, something you typed could theoretically influence the model. In practice, your specific sentence doesn't end up verbatim in the AI. But dumping sensitive information into huge data pools is still not a great idea.
What you should never type in — regardless of provider
Here's the simple list:
- Passwords (sounds obvious, but it happens)
- Real names in sensitive contexts — like "write a complaint for Jane Smith, who on March 15th..."
- Trade secrets — business plans, internal figures, confidential strategies
- Health data — yours or anyone else's
- Bank details, credit card numbers (yes, this happens too)
- Anything you wouldn't want posted on a public bulletin board
That sounds paranoid. It isn't. It's pragmatic.
What you can do
First: check the privacy settings. All three major providers have some option to disable training somewhere. It's sometimes a bit buried, but it's there.
Second: keep sensitive stuff local. If you really need to run confidential text through an AI, look into a local model — for example via Ollama. Then nothing leaves your own machine.
Third: just stay aware. AI isn't a diary. It's a public tool running on someone else's servers somewhere.
No reason to panic. But a reason to think.
Unpaid opinion, personal experience. As of spring 2026 — privacy policies change, so it's worth occasionally checking the current terms of the respective providers.