What not to tell ChatGPT (and what to do instead)
The short answer: don't tell ChatGPT, or any other AI chatbot, anything you wouldn't be comfortable writing on a postcard to a company you've never met. That means no passwords or codes, no card or bank details, no ID numbers, no medical records, no other people's private details, and nothing you've promised to keep confidential. You can still get almost all the help you want by replacing the real details with placeholders.
The rest of this page explains the list, the two settings worth changing before you start, and what to do if you've already shared something.
Why it matters
When you type into ChatGPT, your words are sent to OpenAI and processed on its servers. By default, conversations on personal accounts can be used to help train OpenAI's models unless you turn that off, and your chats stay in your history until you delete them. OpenAI explains this in its Data controls help page. Other AI tools have their own rules, and all of them change from time to time.
So the safe habit is simple: assume anything you type could be stored, and could be seen by someone other than you.
The "never tell" list
- Passwords, PINs and one-time codes. No genuine service needs you to give a chatbot your password. If a message or website tells you to, treat it as a scam.
- Bank, card and payment details. Card numbers, security codes, account logins and banking app screenshots.
- ID numbers. Passport, national ID, driving licence, tax and social security numbers.
- Medical records and test results with your name on them. You can ask general health questions without them, and a doctor or pharmacist should confirm anything important.
- Other people's private details. Names with phone numbers or addresses, a friend's health news, screenshots of family group chats. They didn't agree to share them.
- Work secrets. Customer lists, contracts, unreleased figures, anything covered by a confidentiality agreement or your employer's rules.
- Your full home address and daily routine together. Fine to say "a small town"; unnecessary to say which house and when it's empty.
- Photos of documents. A photo of a letter, bill or ID card can carry far more detail than you notice.
- Anything you'd be upset to see again. If it would hurt to read it in someone else's hands, leave it out.
What to do instead: use placeholders
You rarely need the real details to get good help. Swap them for placeholders, get the answer, then add the real details yourself:
My bank sent a letter saying my account [number] will be closed unless I call [phone number]. Help me work out whether this is genuine and what I should check.
The advice you get back is just as useful, and nothing private left your device. For letters and bills, type out only the part you need help with instead of uploading a photo.
Two settings to change first
Both are in ChatGPT's Settings, under Data controls, according to OpenAI's help page linked above:
- Improve the model for everyone. Turn this off if you don't want your new conversations used to train OpenAI's models. Your existing chats stay in your history; turning it off doesn't delete anything.
- Temporary chat. For one-off questions, start a temporary chat. OpenAI says these don't appear in your history, don't create memories and aren't used to improve its models, though they may be kept for up to 30 days for safety reasons.
One detail that catches people out: OpenAI says that if you give feedback on an answer, such as a thumbs up or down, the whole conversation may be used for training even if you've opted out. If the chat contained anything personal, skip the feedback buttons.
If you've already shared something
Don't panic, and act on the thing that matters most:
- A password or code: change the password on the real service straight away, and turn on two-step verification.
- Card or bank details: contact your bank using the number on your card or its official app.
- Anything else: delete the chat from your history. Then change the training setting above so it doesn't happen again.
Frequently asked questions
Is ChatGPT safe to use at all?
For everyday questions, yes, if you keep private details out and check important answers. Chatbots can sound confident and still be wrong, so check names, numbers, dates and anything about health, money or the law with an official source. Our five-minute check shows how.
Can people at OpenAI read my chats?
OpenAI's help pages explain how conversations may be stored and used, including for safety. The safe assumption is that a person could see what you type, which is why the never-tell list matters more than any setting.
Does deleting a chat remove it completely?
Deleting a chat removes it from your history. If you need to know exactly how long deleted data is kept, read OpenAI's current privacy information, because retention rules can change.
Is it different for work accounts?
Often, yes. OpenAI says it doesn't use content from ChatGPT Business, Enterprise or Edu workspaces to train its models by default. Your employer's own rules still apply, so check them before you paste anything from work. Our guide to using ChatGPT safely at work covers this in more detail.
What about Claude, Grok and Gemini?
The same never-tell list applies to every AI chatbot. Each company has its own settings and terms, so look for its privacy or data controls page before you start.
If you use AI for work, the AI Safety & Privacy Checklist for Work ($7) puts the never-paste list, a settings record and a one-page team policy on a few printable pages.
Not ready to buy? Start with the free guide, Your first week with AI: ten prompts and a privacy checklist, plus a five-day email course.