What not to paste into an AI tool

An AI tool is somebody else's service. What goes into it leaves your business, and from that moment it sits somewhere you do not control and cannot necessarily say for how long.

That is not an argument against using one. It means there is a line, and the line is not where most people guess it is.

The things that never go in

• Card details, in any form.
• Passwords, access keys and tokens for other services.
• Medical records or any health information.
• Official identity numbers.
• Biometric data, including face images used for identification.
• Criminal or legal information about an identifiable person.

Those are the same categories we do not collect in the systems we build, absent a specific approval and a risk review. If something does not belong in a database we designed, it certainly does not belong in a chat box.

The part that gets forgotten: other people's information

A customer list is not your information. It is theirs, entrusted to you for a particular purpose. Pasting it into an outside tool to "tidy up the table" moves other people's data to a third party they never agreed to.

The same goes for a message a customer sent you, a scan of a document they uploaded, or a transcript of a call. That it is in your possession does not make it yours to do as you like with.

And what does

Most day-to-day work is fine. Marketing copy, the wording of an email, a summary of a public document, ideas, rephrasing something you already publish. Code too, as long as there are no keys or secrets inside it.

If you need to work on real data, work on a disguised version: same structure, invented names and numbers. That covers almost everything anyone is actually trying to achieve, and it takes the question off the table.

And when the data has to be real

Then it is not something to paste, it is something to build. A system that processes real data does it server-side, with a named provider, under an agreement that says what is retained and for how long. On this site, for instance, the automatic content translation only ever reads content that is already published, and touches no customer data at all.

The difference between the two is not technological. It is who is answerable if something gets out.

A system that processes real data in a defined way