You paste an email into ChatGPT so it can help you reply. Inside: your coworker's full name, her phone number, and a line about her sick leave. You don't think twice. You just want a polite answer, fast.
That's the habit worth changing. An AI assistant is genuinely useful, but what you type doesn't stay on your screen: it goes to a company's servers. This guide explains, in plain terms, what happens, what you should never paste, how to clean a text before sending it, and which settings to look for. No legal or technical background needed.
What happens to your messages, in general
When you write to ChatGPT, Claude, Gemini or Le Chat, your message is sent to the company running the service, processed by the model, and the conversation is usually kept in your history. Nothing shocking so far: it's like an email stored by your provider.
The difference is how those conversations may be used. Depending on the tool, your account type (free, paid, business) and your settings, your chats may be used to improve the models, and some conversations may be reviewed by people as part of that. The rules differ between services and change over time. So I won't give you retention periods or exact button names: they could be wrong tomorrow.
The only reliable method: read your tool's privacy policy and open its settings. Meanwhile, follow one simple rule: never write anything to an AI that you wouldn't want a stranger to read.
If you want to know how a model "learns" from text, the glossary explains training in two sentences.
The never-paste list
Keep this list in mind, or better, save it somewhere:
- Passwords and codes: passwords, verification codes, PINs, access keys. There's no good reason to give them to an AI.
- Bank and ID numbers: card numbers, bank account numbers, social security number, passport, ID card or driver's license numbers.
- Health data: diagnoses, treatments, medical reports, yours or anyone else's.
- Other people's data: names, contact details, personal situations of coworkers, clients, friends or family. They didn't agree to anything.
- Your employer's confidential information: contracts, internal figures, unannounced projects, private source code, customer data.
- Your exact address, and anything that makes you easy to locate (home address, daily routine, your kids' school).
You can still ask for help on all of these topics. Just do it without the real data. "Explain what this term on a prescription means" works perfectly well without your name or insurance number.
How to anonymize a text before pasting it
Anonymizing means removing what tells someone who you're talking about, while keeping what the AI needs to help you. Three moves cover most cases.
1. Replace names with roles
"Sophie Martin" becomes "my coworker", "client A" or "[Name]". The assistant doesn't need the real name to draft a reply or summarize a situation.
2. Round or hide numbers
"$47,382 in revenue" becomes "about $50,000" or "[amount]". The order of magnitude is almost always enough to reason with.
3. Remove identifiers
Phone numbers, emails, addresses, file numbers, company names, very precise dates: delete them or swap in a bracket. Check the email signature too, that's where contact details tend to hide.
Before and after
Before:
Hi Sophie Martin, following your sick leave for your herniated disc, can you call me back at 555-0142? The Dupont file (no. 2024-883) at Buildplus is due March 14, budget $47,382.
After:
Hi [Name], following your absence, can you call me back? A client file is due mid-March, budget around $50,000. Help me rewrite this message so it sounds warm and professional.
The AI has everything it needs to help you write. You put the real details back yourself, in your email app, right before sending.
Settings to look for in your assistant
Names and locations vary by tool and change with updates. Look in the settings (often under something like "Privacy", "Data controls" or "Personalization") and in the official help pages. Here's what usually exists:
- History: the list of your conversations. You can usually delete one conversation, or all of them.
- Use for training: a setting that says whether your chats may be used to improve the models. Depending on the tool and plan, you may be able to turn it off.
- Memory: some assistants remember things about you across conversations. Check what's been saved and delete anything that shouldn't be there.
- Temporary chats: on some tools, a mode similar to private browsing that doesn't keep the chat in your history. Check the help pages for what it does exactly.
- Data export and deletion: an option to download your data or delete your account.
Also think about your custom instructions: what you write there is read in every conversation. Put your job or skill level in them, never anything sensitive. The article on ChatGPT custom instructions covers what to include.
If you live in the European Union, the GDPR gives you rights over your data, such as the right to access it or ask for it to be deleted. Services usually explain how to use these rights in their privacy policy.
At work: your employer's rules come first
At work, it's not only your call. Your company may have an AI policy: approved tools, business accounts, data you must never share. Some ban these tools, others provide an internal one.
So the rule is simple: follow your employer's policy, and if you don't know it, ask before pasting work data into any tool. A two-minute message to your manager or IT team saves you a real problem. The article on using ChatGPT at work gives concrete uses that stay safe.
Photos, files and screenshots carry data too
Many assistants accept images and documents. Handy, but a file often says more than you think:
- a screenshot can show notifications, names in a sidebar, an open tab;
- a photo of a document may include an address, a signature or a number in small print in a corner;
- a spreadsheet can hide extra tabs or columns full of customer data;
- a photo can contain technical details such as where it was taken, depending on your phone.
Before sending, crop the image, blur whatever's extra, or copy only the useful part of the document into a new file.
Mini exercise: 10 minutes to check and clean up
- Open your assistant's settings (ChatGPT, Claude, Gemini, Le Chat or another).
- Find three things: where history is managed, whether there's a setting about using your data for training, and where memory lives if the tool has it. Write down what you found.
- Look at the memory: delete anything you wouldn't have wanted to share.
- Pick a real message you'd like help rewording (an email, a note to a neighbor, a request to a company).
- Run it through the checklist: names replaced with roles? Numbers rounded? Phones, emails, addresses and reference numbers removed? No health, banking or password info? Signature gone?
- Send the cleaned version to the assistant with your rewording request.
What to notice: is the answer worse without the real data? In the vast majority of cases, no. You'll see that the AI almost never needs personal information to help you well.
Going further
The free chapter Light the Lantern helps you understand what an LLM is, choose your assistant, set privacy, custom instructions and memory, then write a first real request. Free, in your browser, no account.
Brand new to all this? Read how to use ChatGPT for beginners. To practice writing effective prompts without sensitive data, the prompt lab lets you build a request piece by piece, entirely in your browser, with nothing sent.
To put this guide into practice, play Safe to paste?, a free game where you spot the personal details to remove from a text before pasting it into an AI chat, then leave with a redacted version and your own paste checklist.
Frequently asked questions
Does ChatGPT read my conversations?
Your messages are sent to the company's servers and kept according to its rules. Depending on the tool, your plan and your settings, they may be used to improve the models, and some chats may be reviewed as part of that. Check your tool's privacy policy and settings to see what applies to you.
If I delete a conversation, is it gone everywhere?
It disappears from your history, but how long it actually takes to be erased on the server side depends on the service and its obligations. The exact answer is in your tool's official help pages. The safest approach is still not to paste sensitive information in the first place.
Is a temporary chat mode enough to protect my data?
When the tool offers it, it limits what stays in your history, but it doesn't turn the assistant into a vault. Your message is still sent and processed. Keep the same rules: no passwords, no health or banking data, no other people's data.
Can I use AI on a confidential work document?
Only if your employer allows it, with the tool they've approved. If you don't know the rule, ask first. In the meantime, you can often get help on an anonymized version, or just on the structure and tone, without the confidential content.