Hidden Tokens / Academy / Lesson 0
What an AI sees of your files
Not as it is: whatever you paste or upload, including the names in a page header, the phone number in a comment and the dates of birth in a spreadsheet, reaches the AI provider's servers, and telling the AI to ignore the names does not unsend them. It becomes defensible when those values are replaced on your own computer before anything leaves, and when you know what that replacement does not cover: a pseudonymised file is not an anonymous one.
What travels with a file
When you upload a document to an AI, it does not receive what you see on screen. It receives the file, and a Word or Excel file carries much more than its visible text: headers and footers, comments, tracked changes, hidden rows and sheets, document properties. This lesson makes you find those places yourself, on an invented case file, in three minutes.
Step by step
-
Download the sample folder
Download the Academy sample folder (.zip) and unzip it. It holds four files from a fictitious London law firm: a letter, a payroll spreadsheet, a services agreement and an identity card scan. Everything in it is invented. No installation is needed for this lesson.
-
Open the letter and look at the page header
Open
01-letter-to-client.docxin Word or LibreOffice. The header at the top of the page names the client, Rowena Kettleby. A header is part of the file: an AI that reads the document reads it too, even if you never scroll up to it. -
Show the comments
In Word, choose Review › Show Comments. A colleague's comment gives the client's mobile number. Comments, tracked changes and hidden text travel with the file; some viewers simply do not display them.
-
Look at the document properties
Choose File › Info. Here the properties say "Invented data — Hidden Tokens Academy"; in a real file they often hold the author's name, the firm and sometimes the client's name in the title.
-
Open the payroll and the contract
In
02-payroll-october-2026.xlsx, each row has a date of birth, a National Insurance number and an IBAN. In03-services-agreement.pdf, the parties, their company numbers, the signatories and a fee of £48,500 sit in the body of the text. -
Count what an upload would send
Uploading these four files to an AI would send about a dozen people's names, their dates of birth, bank details and identity numbers, plus a phone number hidden in a comment. That is the starting point of every other lesson: the AI should receive codes instead, and the real values should never leave your computer.
What Hidden Tokens does, and what it does not
Hidden Tokens replaces each person, company, address, identifier and bank detail with a stable code, on your own computer, before the AI reads the file: the same person keeps the same code in every document of the folder, so the AI can still reason across files. The key that turns codes back into names stays on your machine, and the AI's result is decoded there.
What it does not do, said plainly:
- It does not make the data anonymous. The coded copy is pseudonymised: with the key, it can be reversed.
- It only protects what goes through it. Text typed into an AI in a browser without the extension, in a private window, in a phone app or in the ChatGPT desktop app is not coded.
- An amount written in the body of a contract stays readable in the default setting. That is deliberate (the AI needs figures to work), but it is your call when the amount itself is the secret.
- It does not certify your compliance. It is a tool that supports it.
What can go wrong
- You see no header or comment. Quick previews (the Mac's Quick Look, a browser, a phone) often hide them. Open the file in Word or LibreOffice.
- Word shows the comment as a small mark only. Use Review › Show Comments, or the comments pane.
- You are tempted to try this on a real client file. Don't, not even "just to see". The sample folder has the same traps, and nothing in it can hurt anyone.
Check yourself
Three questions, answered on this page. Nothing is sent anywhere.
✓ Lesson done: every step ticked and every answer right. Your progress stays in this browser only.
Frequently asked questions
Is ChatGPT Enterprise or Claude for Work enough to protect client data?
Business plans generally commit not to train on your data and give you retention controls, which matters. But the files still reach the provider's servers, in clear. For professional secrecy or a strict client contract, the question is often whether the data leaves your organisation at all; pseudonymising before sending answers that question, and can be combined with a business plan.
Is pseudonymised data anonymous?
No. Pseudonymised data has its identifiers replaced by codes, but whoever holds the key can reverse it, so under the GDPR it remains personal data for you. Clues such as a rare job title, a town and a year of birth can also point to a person without any name. Anonymous data, by contrast, cannot be linked back to anyone by anyone.
Can I just tell the AI to ignore the names?
No. An instruction in the prompt does not change what was sent: the provider has already received the file, names included. Protection has to happen on your computer, before the file leaves.
Written for app 0.14.4 and connector 2.42.0 on . Not yet replayed on screen on Mac and Windows: if a label differs from what you see, trust your screen and tell us.