Privacy with AI tools: what you should and should not enter

What happens to what I enter into an AI chat?

As soon as you press send, your text goes to the provider's servers. There the model calculates the answer, and the chat usually ends up in your history so that you can open it again later. Uploaded files and photos are processed in exactly the same way as typed text.

In addition, your inputs can serve as training data, meaning they help improve future models. Whether that happens depends on the provider, the plan and your settings. With several big assistants, training is switched on in the personal plan until you turn it off yourself.

Some providers also have humans review a portion of chats, for example for quality control or to detect misuse. Google explicitly asks Gemini users not to enter anything confidential that they would not want a reviewer to see. Google keeps reviewed chats for up to three years, separately from your account.

What should never go into an AI chat?

A simple rule of thumb: anything that could harm you or others if it fell into the wrong hands stays out. That applies even if you have turned off training, because your chats are still stored.

• Passwords, PINs, one-time codes (TANs) and login details, not even "just to check quickly".
• Bank details, credit card numbers, tax ID numbers and ID card or insurance numbers.
• Health data with names, such as test results, diagnoses or doctors' letters, whether yours or other people's.
• Customer data and third-party data: names, addresses, phone numbers, contracts or entire customer lists.
• Trade secrets such as internal figures, unreleased products or source code containing access keys.
• Photos, chat histories or documents belonging to other people without their consent.

If you still need help with such a topic, first remove everything that makes a person identifiable. "Mr Smith from Bristol, test results from 3 May" becomes "a patient in his mid-50s, recent test results". The AI almost never needs real names to give a good answer. How to still provide enough context is explained in our guide Writing prompts for beginners.

Where do I turn off training in ChatGPT, Claude, Gemini, Copilot and Mistral?

The following paths come from the providers' help pages (as of November 2026). Menu names change occasionally, and the interface in your language may be labelled slightly differently. Microsoft has also introduced a new Copilot app, where the paths may differ.

Important: turning it off only applies to future chats. Conversations that are already stored remain in your history until you delete them yourself. And even without training, providers keep chats for a while, for example to monitor misuse. What lies behind the term is explained in our glossary under training opt-out.

Personal or business: what applies to freelancers and companies?

Free versions and personal subscriptions such as Plus or Pro are meant for individuals. Business and enterprise plans are aimed at companies and usually have stricter data rules: training on company data is generally switched off by default there, administrators manage the settings centrally, and there are data processing agreements. You can read when a paid plan is worth it in general in the guide Free or subscription.

Even so, do not rely blindly on the label. With Mistral Vibe, for example, training is switched off by default in the Enterprise plan, whereas in the Team plan the admin has to turn it off themselves. If your employer provides an AI tool, use exactly that tool for work and not your personal account.

If you are self-employed and enter customer data into an AI tool, you are processing personal data within the meaning of the GDPR. The basic idea: you need a legal basis for this, must inform your customers transparently and usually have to conclude a data processing agreement (DPA) with the provider. Personal plans are generally not designed for this.

The simplest approach is not to enter any personal customer data at all and to work with anonymised examples. This section is not legal advice. For specific questions, contact a data protection professional.

What alternatives are there for sensitive data?

The most data-minimising option is local AI. The model runs directly on your computer or phone, and your inputs do not leave the device. On the other hand, local models are smaller, slower and less reliable on facts than the big cloud assistants. How to set this up without any prior knowledge is shown in Use AI offline on your phone.

There are also local solutions for individual tasks: Whisper (local) transcribes recordings on your own computer; you can find out more in the tools for meetings and transcription. Obsidian + lokale KI connects your notes with a model that runs on your device.

A second option is providers based in the EU, such as Mistral from France or DeepL from Germany. That can simplify questions around data transfers to third countries. Being based in the EU is not a free pass, though: Mistral Vibe uses the inputs of personal users for training by default until you object.

Checklist: what do I check before I enter something?

1. Would I also show this text to a stranger?
2. Does it contain names, addresses, account numbers or health information? Then remove them or replace them with placeholders.
3. Does the content concern only me, or also customers, colleagues or family?
4. Is training switched off, or am I using a temporary or incognito chat?
5. Is it work-related? Then use the tool my employer has approved.
6. Am I uploading a file? Then also check comments, hidden worksheets and metadata.
7. Would a local AI perhaps be enough for this task?

By the way, you should treat the answers just as critically as your inputs. How to do that is shown in the guide Checking AI answers.