When you chat with an AI, your words usually travel to a company's servers — where they may be stored, sampled by human reviewers for "quality," used to train future versions, or exposed in a leak. For everyday questions, none of that matters. But some information, once you've pasted it in, can't be un-pasted — and a simple habit about what not to send saves you from the regret.
Where your words can go
Stored and sometimes reviewed
Many services keep your conversation history, and some have staff review samples to improve the product. Assume a human could read what you wrote.
Used to train future versions
Your inputs may help train the next model — which can, in rare cases, surface fragments to other people. What you teach it doesn't always stay yours.
Leaks, bugs, and shared devices
Stored data gets breached; chat histories have been exposed by bugs; and your history may sit in your account or on a shared/public computer for the next person.
"Free" often means you're paying in data
If a tool is free, your inputs are more likely to be part of the bargain. Read what the setting says, and assume the most data-hungry default unless told otherwise.
Don't paste these
Keep these out of any AI chat (unless it's a tool you specifically trust, and need, for exactly that):
- Passwords, PINs, security answers, or 2FA codes. Never, to anything.
- Full financial details — card numbers, bank account and routing numbers, full statements.
- Government IDs — Social Security / national ID, passport, driver's license numbers.
- Health details tied to your name — diagnoses, records, results — unless using a service meant and trusted for that.
- Other people's private information. You can't consent on their behalf — a friend's diagnosis, an employee's file, your kid's details.
- Work confidential or proprietary material. Pasting it may breach your employer's policy or an NDA, and has cost people their jobs.
- Anything you'd be devastated to see made public — private photos, secrets, anything sensitive enough that a leak would hurt.
Get the help without the exposure
You almost never have to choose between AI's help and your privacy — a little prep gives you both.
- Redact before you paste. Swap real names, numbers, and addresses for placeholders — "[NAME]", "[ACCOUNT]", "$X". The AI can still explain the clause or draft the letter around the blanks.
- Ask the general version. "How do disputes like this usually work?" gets you most of the value without uploading your whole file.
- Turn off history and training where offered. Many tools let you disable chat history or "improve the model" — do it for sensitive work, and delete sensitive chats after.
- Match the tool to the stakes. For sensitive tasks prefer a service that says it doesn't train on your inputs — but still don't paste outright secrets, even there.
- Clean up on shared devices. Log out and clear the history on any phone or computer that isn't only yours.
- When in doubt, leave it out. You can always give the one private detail to a human professional instead of typing it into a chat box.
Your "before I paste" kit
Tick before sending anything sensitive.
The close
None of this means clamming up — AI is genuinely useful, and most of what you ask it is harmless. It means keeping one small reflex: before you paste, glance at whether there's a password, an ID, a financial number, or someone else's secret in there, and take it out. The help you get from redacting "[NAME] owes me $X" is identical to pasting the real thing — minus the part you can't take back. Postcard, not diary. When you're weighing whether an answer is even safe to rely on, run the trust check.