Privacy and your data
What you type into a chatbot is stored, sometimes read by people, and sometimes used for training. Like writing on a postcard. Anyone who handles it can read it. Learn what to leave out, and which settings to check.
First, what do you think?
You delete a chat. Is everything you typed gone now? Have a guess.
What happens to what you type?
The whole story in three sentences. What you type into a chatbot goes to the company behind it. Most of these companies store it. Sometimes people read it, and sometimes it's used to train the next model.
Think of a postcard. You write it, and it travels through many hands. Anyone who handles it can read it. So you don't write your PIN code on it.
- It's stored. The Dutch data protection authority warns that most companies behind chatbots store everything you type.
- People may read it. Google writes that some Gemini chats are read by human reviewers. Its advice: don't type anything confidential.
- It may train the model. On many plans, your chats are used for training. Unless you turn it off.
And it goes wrong in practice. The same authority got reports of data breaches. A GP practice employee typed patient data into a chatbot. A telecom employee typed in a file of customer addresses.
What the predictor does
The predictor doesn't keep secrets. That's the rule.
It doesn't know what's private. It only sees text in its context window. Your name, your address, your diagnosis: to the predictor, it's all tokens.
Training makes it worse. A model learns from its training text. If your chats go into that text, they become part of what it learned from.
That's why the settings matter. Each tool has its own. On some, training is on unless you turn it off. On others, it only happens if you allow it.
Type into a chatbot as if you're writing a postcard. Leave out what others shouldn't read, and check the training setting.
Try it yourself
Eight things you might type. Fine as it is, or not like this?
Everything runs in your browser. Nothing you type is sent anywhere.
How do you protect your data?
Four steps. Use them before you paste anything.
Do the postcard test
Before you type: would you write this on a postcard? If not, don't type it like this.
Leave it out, or replace it
Remove names, addresses, numbers and passwords. Write "my neighbour" instead of a name. The predictor rarely needs the real details.
Check the training setting
Look at the tool card, and at the tool's settings. Turn training off if you don't want it. Some tools have a chat mode that's never used for training.
At work: only approved tools
Customer data, patient data, company secrets. Use only the tools your employer allows, under their rules.
In practice
Three people write a postcard.
Case 1The angry message
Noor (15) wants help replying to a mean message in a group chat. She wants to paste the whole chat, with everyone's names and phone numbers.
Paste it all in.
Those are other people's details. They never agreed to share them.
Describe it in her own words, without names. "Someone in my class wrote something mean. How do I reply calmly?"
She gets help, and her classmates' details stay out.
Case 2The customer list
Joost (41) has to sort a list of customer complaints. He wants to paste the file, with names and addresses.
Use their own free chatbot, because it's quick.
The data protection authority had exactly such a report. A file of customer addresses went into a chatbot.
Ask which tools his employer allows. Remove names and addresses first.
He works faster without causing a data breach.
Case 3The lab results
Wim (70) has blood test results with his name, date of birth and patient number. He wants them explained.
Upload the full page.
Health data is among the most private there is. It may be stored, read or used for training.
Type only the values he wants explained, without his details. Then ask his doctor about what it means for him.
He understands the words, and his medical file stays his own.
When not to
- Not with passwords or bank detailsNobody needs them in a chat.Never type them in.
- Not with other people's dataThey didn't agree to it.Describe the situation without names.
- Not with work data in a private toolThat's how data breaches happen.Use only approved tools.
By situation
Leave out your full name, your school and your classmates. Ask your teacher which tools you may use.
Customer, patient and company data stay in approved tools. Remove details you don't need.
Check the training setting of the tool you use. Leave out names, addresses and numbers. Do the postcard test.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| A chat is private | No | It's stored, and some chats are read by people. |
| Deleting a chat removes it everywhere | Not always | Each maker has its own rules. Check the settings. |
| Training is always off | No | On some plans it's on unless you turn it off. |
| It's fine at work if it saves time | No | Data breaches happened exactly that way. |
Related topics
Read first
- Checking the answerthe other half of using AI carefully.
Next
- Deepfakes and scamswhen others misuse AI.
Related
- The context windoweverything you type is text the predictor sees.
Sources
- governmentAutoriteit Persoonsgegevens. Caution: use of AI chatbot may lead to data breaches.The Dutch data protection authority got reports of data breaches: a GP practice employee typed patient data into a chatbot, a telecom employee a file of customer addresses. Most companies behind chatbots store everything you type. (Checked through a search engine.)
- documentationGoogle. Gemini Apps Privacy Hub.Some chats are read by human reviewers, so don't type anything confidential. With Keep Activity on, chats can train Google's models; turn it off and they're kept 72 hours.
- documentationAnthropic. Is my data used for model training?.Chats on Free, Pro and Max are used for training if you allow it. You change that with the Model Improvement setting. Incognito chats are never used.
- documentationOpenAI. Data controls in ChatGPT.On Free, Plus and Pro, your chats are used for training unless you turn off 'Improve the model for everyone'. (OpenAI blocks automatic reading; checked through a search engine.)
- documentationMistral AI. Can I opt out of my input or output data being used for training?.On the free plan, your chats are used for training unless you turn it off in the privacy settings. (Checked through a search engine.)
- documentationMicrosoft. Microsoft Copilot privacy controls.Signed in, you can choose whether your conversations are used for model training. Microsoft says the newer app may not show this setting yet. (Checked through a search engine.)
Test
You get 7 questions about the whole page. Get 5 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.