Explained · Free to read · 5 min read
AI and privacy: What happens to what I type into ChatGPT and other chatbots?
Whatever you type into an AI chatbot is sent to the provider's servers, stored there, and, depending on your settings, may be used to train future models or be read by human reviewers. The big services let you switch off training and delete chats, but copies often remain for a while. The safest rule is simple: do not type anything you would not hand to a company you do not know.
What happens when you type
An AI chatbot such as ChatGPT, Gemini or Claude does not run on your phone or computer. It runs in the provider's data centres. Everything you type or upload is sent there, processed, and the answer is sent back.
Germany's Federal Office for Information Security (BSI) points out that the operator can access your inputs and the generated answers. Whether and how that data is reused or shared with third parties is set by each provider in its terms of use and privacy policy. It is also possible that unauthorised people gain access to it.
One common misunderstanding: the language model does not learn while you are chatting with it. Your conversations can, however, be stored and later used to train a new version of the model. Training means feeding the model large amounts of text so it can learn patterns from it.
Training: what your conversations may be used for
On the free and personal plans of the large providers, using chats for training is often switched on by default or requested during sign-up. According to Germany's consumer advice centres (Verbraucherzentrale), you have to actively object if you do not want your data used this way.
For ChatGPT Free, Plus and Pro, OpenAI says data sharing is on by default. You can turn it off under Settings, Data controls, with the switch "Improve the model for everyone". After that, new conversations are no longer used for training. For its business plans and its programming interface (API), OpenAI says it does not train on customer data by default.
With Anthropic's Claude, you decide whether your chats may be used for training. If you agree, Anthropic says your data may be kept in de-identified form for up to five years in its training pipelines. If you later change your mind, training runs already in progress and models already trained are not undone (as of October 2026).
Humans who may read along
Many people assume only a machine reads their chats. That is not always true. In its Gemini privacy notice, Google states that a subset of chats is reviewed by human reviewers, including trained staff at service providers. Google explicitly asks users not to enter confidential information they would not want a reviewer to see.
According to Google, reviewed Gemini chats are kept for up to three years, even if you delete your activity. By default, Gemini activity is kept for 18 months. You can change this to 3 or 36 months, or keep it indefinitely, and you can turn activity saving off (as of October 2026).
Other providers also keep conversations longer if automated systems flag them as breaking usage rules. Anthropic gives a figure of up to two years for this.
Deleting, temporary chats and what is left behind
The major services offer a kind of guest mode. ChatGPT calls it temporary chat, Claude calls it incognito chat. These conversations do not appear in your history and, according to the providers, are not used for training. They are not entirely without trace, though: OpenAI may keep a copy for up to 30 days for safety purposes. With Gemini, temporary chats and chats made with activity saving off are kept with your account for 72 hours.
When you delete a normal chat, it disappears from your list straight away. OpenAI and Anthropic say it is removed from their servers within 30 days. Exceptions apply if the law requires longer retention or the provider needs the data for security reasons.
These periods and feature names change from time to time. The figures here are as of October 2026. If in doubt, check the help pages of the service you use.
Your rights in the EU
In the European Union, the General Data Protection Regulation (GDPR) applies. As Germany's consumer advice centres note, it covers every company offering its services in the EU, including providers based in the United States.
Among other things, the GDPR gives you a right of access: you can ask which personal data a provider processes about you, for what purpose and who receives it. You can also ask for data to be deleted or corrected, and you can object to processing. In its Gemini notice, Google explicitly mentions the right to have inaccurate personal data in answers corrected.
Our assessment: these rights matter, but they do not solve everything. Anthropic states that data already included in a trained model cannot be taken out again. Being careful about what you type remains the most effective protection.
What you are better off not typing
The BSI and the consumer advice centres agree: share as little personal and sensitive data with AI services as possible. That includes passwords, bank details, ID numbers, detailed health information, full names and addresses of yourself or others, and confidential documents from work.
Most questions work just as well without those details. Instead of "My neighbour Mr Miller at 4 Linden Street owes me money", you can write "An acquaintance owes me money, how do I bring it up?". The answer is usually just as useful.
If you use AI at work, there is one more rule: company data belongs only in tools your employer has approved for that purpose. Business versions of these services follow different rules from personal accounts.
Sources
- BSI: Generative AI and its risks (German)
- Verbraucherzentrale: AI chatbots at a glance (German)
- OpenAI Help Center: Keep history on but disable model training (German)
- OpenAI Help Center: Temporary chat in ChatGPT
- Google: Gemini Apps Privacy Hub
Show all 6 sources
Created with AI assistance and editorially approved before publication.