What never to paste into an AI chat (and what is actually fine)

Where a pasted secret actually goes

Nothing on this page is about hackers. A pasted secret creates copies through the ordinary, documented workings of chat products, and the copies are the problem.

Four mechanisms do the work. First, retention: what you paste becomes part of a stored conversation, on a documented schedule. OpenAI, for example, schedules deleted chats for permanent removal within 30 days, and Google's Gemini activity sits under an auto-delete setting that defaults to 18 months. Second, training defaults: the consumer tiers at OpenAI, Anthropic, and Google can use conversations to improve models unless you turn the setting off, while business and API tiers default to off. Third, review: all three providers state that flagged conversations can be read by people, which is routine safety practice and still means a human can end up looking at what you pasted. Fourth, sharing: a conversation shared as a link carries everything in it, including the document you uploaded three messages before the part you wanted to show someone.

None of these is a scandal. All of them together mean one thing: a chat box is a place where text gets kept, sometimes learned from, and occasionally seen. Paste accordingly. The rest of this page is just that sentence applied to specific cases, and is AI safe to use covers the wider safety question around it.

The never list: six items a placeholder replaces for free

These items have no legitimate reason to enter a chat box, because the model does not need them to help you. Every one of them can be replaced with a placeholder without changing the answer you get.

Never paste	Why it is different from other data	What to do instead
Passwords, API keys, access tokens, two-factor codes	A credential is valuable to anyone who ever sees it, and rotating one you leaked is work you can avoid	Type PASSWORD or KEY where it would go; the answer will not change
Full card, account, or routing numbers	Payment fraud needs exactly these strings and nothing else	"my card" and "my account" work fine in any money question
Government identifiers: SSN, passport, national ID, tax numbers	Identity theft is built from these, and they cannot be rotated like a password	The model never needs the real number to explain a form or a process
Other people's personal data: names with addresses, health details, HR records	It is their information, and privacy law in most places treats sharing it as your act, not the chatbot's	Swap real names for Person A and Person B before pasting
Documents under NDA or an employer confidentiality policy	The question is contractual, and "I pasted it into a chatbot" is a disclosure whether or not anything bad happens after	Ask the question with the structure of the document, minus the protected specifics
Anything you would not email to an outside consultant without a contract	This is the test that catches everything the rows above missed	If the answer is no, redact until the answer is yes

The last row is the whole table compressed. An AI chat is an outside party you have not signed anything with, being helpful. Calibrate exactly as you would for that.

The items nobody thinks of

The hard list above is easy to follow because the items look sensitive. The pastes that actually catch careful people are the ones that do not.

• Screenshots. The error message you meant to share arrives with your notification tray, your open tabs, an email subject line from someone else, and a calendar reminder in the corner. Crop to the thing you are asking about.
• Code files with secrets inside. GitGuardian counted 23.8 million new secrets leaked in public GitHub commits during 2024, and a pasted config or .env file with a live key in it is the chat-window version of the same mistake. Search a file for key, secret, token, and password before it goes in.
• Meeting transcripts and recordings. Everyone in that meeting spoke to the people in the room, and none of them consented to a chat upload. Summaries you wrote yourself carry no such problem.
• Spreadsheets with a customer column. The question is about the formula; the paste includes every customer email in the sheet. Delete the identifying columns first, or paste ten rows of structure instead of the sheet.
• Contracts with counterparty names. Understanding a clause needs the clause, and almost never needs who signed it. Summarizing documents with AI works exactly as well on a redacted file.
• Exported calendars and inboxes. One file, hundreds of other people's names, times, and subjects. Ask about the scheduling problem, paste the shape of the week, keep the export.

The pattern across all six: the sensitive part is usually the context around what you need, and cropping to what you need removes it. That crop is thirty seconds.

What is actually fine

The point of a short never-list is that everything outside it is usable, and most everyday pasting carries no issue at all: your own drafts and notes, public information, general questions however personal the topic, your own code without embedded secrets, and any document you have redacted down to what the question needs.

The enabling habit is redaction before pasting, and the reason it works is that models reason over structure rather than identity. Test it once yourself: ask for the letter to your landlord about the broken heating with the name in, then with [LANDLORD] in its place, and compare. "Explain this contract clause" does not improve when the model knows which companies signed. In the rare case where a name genuinely matters, you can add it back into the output yourself, against the alternative of an unredacted paste you cannot take back.

Two settings finish the job. Check your chat app's training toggle once, because the consumer defaults are not in your favor, and know your app's retention numbers. What a provider should publish looks like Whizi's data page: what is stored, for how long, and what deletion removes. Whizi's own answers there: uploads expire after up to 30 days, and your content does not train Whizi-owned models and is not sold as training data. Any tool you use daily should answer the same questions somewhere you can cite. If you are choosing between subscriptions on exactly this axis, is a third-party AI subscription safe walks the same standard across the whole category.

The rule that replaces this whole page once it becomes habit: if you would not email it to an outside consultant without a contract, do not paste it. Everything else, paste freely and swap out the names.