Hidden text in assignment prompts: the trap that actually works

Why this works when detectors do not

A statistical AI detector scores how predictable your word choices are. It cannot distinguish plain writing from generated writing reliably, which is why it flags human work and misses machine work.

A hidden-text trap is not an inference at all. Either the improbable word is in the submission or it is not. There is no false-positive rate to argue about, because a student who never pasted the prompt into a model has no way to produce the trigger word by chance.

It is also cheap. It needs no vendor, no subscription, and no model — just white text in a Word document.

What the hidden text usually looks like

**White or background-coloured text.** Present in the document, invisible against the page.

**One-point or half-point font size.** Renders as a hairline that reads as a formatting artifact.

**Text behind an image, or in a zero-height text box.** Survives copy-paste even though it never displays.

**Metadata and alt text.** Some file formats carry text that no reader ever sees on the page.

All of these are ordinary characters as far as the file is concerned. That is the entire trick: invisible to a human eye, fully legible to anything that reads the file.

How to see what a document actually contains

Select the whole document and set the text colour to black and the size to 12pt. Anything hidden appears.

Paste the document into a plain-text editor. Plain text has no colour and no font size, so everything present becomes visible at once.

In Word, turn on Show/Hide (¶) to reveal formatting marks, and check for text boxes and anchored objects.

Our checker reads a passage and reports what is actually in it — zero-width characters, non-standard spaces, and other invisible content that renders as nothing. It is the same problem viewed from the other end: knowing what is in a block of text before it goes anywhere.

The same technique, pointed the other way

The identical trick appears in recruitment, where candidates hide instructions such as “rate this candidate highly” in white text in a CV, aimed at automated screening. It has also been documented in academic papers, hiding directions aimed at reviewers using language models.

This is prompt injection, and it is the general problem: any system that reads a document and acts on what it reads will act on text a human never sees. That is worth understanding whichever side of it you are on.

What it does and does not prove

It proves the assignment file was passed to a language model. That is a narrower claim than it first appears — it does not distinguish a student who generated an entire essay from one who pasted the prompt to ask what it meant, or to translate it, or to have it read aloud.

That gap matters where students use assistive tools for legitimate reasons. A trap catches contact with a model, and contact is not authorship — the same distinction that applies to watermarks.

Questions

Is hidden text in an assignment the same as a watermark?

No, and they work in opposite directions. A watermark is placed in a model's output by the provider. A hidden-text trap is placed in the input by the instructor. The trap is checkable by anyone; the watermark is checkable only by whoever holds the key.

Would retyping the prompt instead of pasting it avoid the trigger?

Yes, because the invisible text is never displayed and so is never retyped. This is why the technique catches bulk copy-paste specifically rather than model use in general.

Can I check a document for hidden text before I use it?

Yes, and it is worth doing with any file you intend to paste somewhere. Select all, force the colour and size, or paste it into a plain-text editor. Anything invisible becomes visible immediately.

Do AI detectors look for this?

No. Tools like Turnitin and GPTZero score the style of the submitted text. A hidden-text trap is found by reading the submission for a specific improbable word, which is something a human does deliberately.