What not to paste into an AI tool, and how to check in a second
Personal details often end up in AI chats by accident. They ride along inside an email thread, a meeting note or a spreadsheet export, and few people read every line before pressing Enter.
Text you paste into an AI tool run by another company leaves your hands. Depending on the service and your settings, it may be kept, read by staff, or used to train future models. The people mentioned in it never agreed to any of that.
Here are the eight things to look for, the one question that settles the unclear cases, and how to check before you paste.
Eight things to look for
- ID numbers Social Security and National Insurance numbers, passport and driver's license numbers. Employee and customer IDs count too: each one points to a single person in a system.
- Card and bank numbers Including partial ones, like "Visa ending 4411", and bank account numbers and IBANs.
- Dates of birth On their own or inside a record: "DOB 03/14/1986".
- Contact details A person's own email address, phone number or home address. A company's general inbox or phone line is fine.
- Names, even first names in passing "Sarah thinks the numbers are off" puts Sarah in your prompt. Look in greetings, cc lines, signatures and meeting notes.
- Health details about one person, even without a name A patient note or a referral letter with the name taken out is still about someone. "Type 2 diabetes is common" is fine; "the patient in room 4 started insulin" is not.
- Details that add up to one person No single detail names anyone, but together they can: "the 38-year-old Lithuanian radiographer on the night shift at St Mary's".
- Client records, tickets, logs and exports One pasted spreadsheet can carry hundreds of names, emails and IDs at once. Support tickets and logs carry them too, often far down the page.
Let a checker read it for you
We're building a Mac app that does this check on your own computer, before you paste. Join the waitlist and we'll let you know when it's ready.
We count visits without cookies; your email is used only to contact you about access.
The one question that settles it
Some cases aren't clear: a name in a news article, a made-up example, a job title. One question settles almost all of them:
Would pasting this tell an AI company something about a specific, living person that they wouldn't expect to be shared?
It's the rule behind every test case we check our own tool against. In practice it sorts things like this:
| Usually fine | Check first |
|---|---|
| A news article that names a head of state | A note about calling the mayor at home |
| A character in a novel | A colleague's name in meeting notes |
| "The account manager" | "Priya, the account manager" |
| A company's support address | A colleague's own email address |
| "Type 2 diabetes is common" | "The patient in room 4 started insulin" |
| Made-up examples, like john.doe@example.com | A real export, even "just to test" |
| Text that's already redacted | A redacted record that still shows a date of birth |
The line isn't between famous people and private ones. It's between what was published and what wasn't: names in a book's acknowledgements are fine, and the same names in an email thread aren't.
How to check in a second
For a short message, ask the question above and skim for the eight things. That takes seconds. Three habits make it faster:
- Paste only the part you need. A question about one paragraph doesn't need the whole thread.
- Swap names for roles. "The client", "Colleague A" and "the patient" work just as well for most questions.
- Use a few made-up rows. For help with a spreadsheet, paste the column headers and two or three invented rows, not the export.
A skim stops working on long documents and exports, where the one detail that matters is on page six. That's the job we're building Can I Prompt That for.
It checks text on your own computer, so nothing you check is uploaded anywhere. Exact patterns catch ID, card and bank numbers, emails and phone numbers. A small AI model, running on your Mac, reads for what patterns can't see, like a first name in passing or a symptom in a note. Text that looks clean gets a second look for names and health details. If it finds something, it can swap each detail for a realistic fake, so the text still reads naturally.
It's a second pair of eyes, not a guarantee. It can miss things, and it can flag text that turns out to be fine; you still decide what to send. It runs on Macs with Apple Silicon first.
Already pasted something?
Delete the conversation, and check the tool's settings for whether chats are kept or used for training. If it was client or patient data, follow your workplace's process for it; that's what it's there for. If your workplace has rules about AI tools, those come first: this is a practical checklist, not legal advice.
Want to try it?
Early-access invites go first to the first 100 people on the waitlist, and to anyone who brings three others.
We count visits without cookies; your email is used only to contact you about access.