Documents you shouldn't upload to an AI

Uploading a file takes a second: you drag it in and it’s done. What doesn’t take a second is the consequence, because an uploaded document can’t be un-uploaded.

This isn’t about fear. It’s about knowing what you’re deciding when you drag something in.

The question to ask first

It isn’t “is this confidential?” — almost everything is, a bit. It’s this one:

Whose information is this, and did I ask?

A client contract isn’t yours: it belongs to both of you. Your mother’s medical report isn’t yours. A CV someone sent you isn’t yours. Your team’s payslips aren’t yours.

What’s yours, you decide. What’s someone else’s, you’re deciding for them.

What I’d never upload

Someone else’s health data. Even family. It’s one of the most protected categories there is, and it isn’t your call.

Work documents under a confidentiality clause. Tender documents under NDA, contracts with secrecy clauses, client information. The problem here isn’t technical, it’s contractual: you may be breaching something you signed.

Third-party personal data in bulk. A customer list, an employee file, a database of names and emails. One data point is one thing; a whole file is another, and that carries legal obligations good faith doesn’t resolve.

Credentials and secrets. Passwords, API keys, tokens. It sounds obvious and it’s among the most uploaded, hidden inside a config file or a screenshot.

Identity documents. ID cards, passports, deeds. Their value to whoever wants them is high and permanent.

What you can, with your head on

Your own contract, your own invoice, a report someone gave you to decide on, your own payslips. There you’re deciding for yourself, and the benefit of having it explained usually outweighs the risk.

Two things that work:

Redact before uploading. Often the document reads the same without the name, the tax number and the address. If removing four fields stops it being sensitive, remove them.

Upload the part, not the whole. If the doubt is in clause seven, upload clause seven. You don’t need the entire contract to answer one question.

What genuinely differs between services

This is where it’s worth looking and almost nobody does. The questions that matter:

• Is it stored? Does the file stay anywhere after answering you?
• Is it trained on? That’s different from storing, and usually sits in a setting that’s on by default.
• Who else sees it? Many services hand your document to a third party to process it.
• Can you delete it? And if you do, does it disappear from backups?

The answers are in each provider’s privacy policy, and they’re surprisingly different between services that look identical from outside. We go into it in what happens to what you tell an AI.

How we do it

We say this because it’s checkable, not as a sales line: in The Judge what you upload isn’t stored. The document travels with the question, the jury’s AIs read it, and it’s forgotten — there’s no file store because there’s nowhere to keep them. History is off unless you turn it on, and without it neither your question nor the verdict is saved anywhere.

That doesn’t exempt you from the question above. If the document belongs to a client and you signed an NDA, us not storing it doesn’t change what you signed.

The ten-second rule

Before dragging a file in, two questions:

1. If this became public, who would it hurt? If the answer is “someone else”, don’t upload it without asking them.
2. Do I need to upload all of this for what I want to ask? Almost never.

That settles 90% of cases without reading a single privacy policy.