The ‘Oishidesu’ Hidden in an Assignment: Can White Text Expose AI Use?

Invisible to a student glancing at the page, potentially legible to a machine reading all its text: one absurd word can turn an assignment into a test of how the tool was used.

On October 23, a Korean forum reposted an X status by @st_gregory5. The post described placing a white-colored instruction roughly meaning 'insert Oishidesu at random' inside assignment directions, apparently to see whether AI-generated answers echoed it. There is no independent confirmation here that a class actually received the file or that any student was caught.

The trick exploits a gap between what people see on a page and what a text-reading system may ingest. Beneath the joke is a real issue: an AI can mistake instructions embedded in reference material for instructions addressed to itself.

The repost lends itself to a playful 'professor strikes back' reading. Yet an actual grading policy would need to address prior disclosure, accessibility and whether one strange word could fairly prove misconduct. Those are editorial questions, not invented quotations from commenters.

White text can blur the line between data to be read and instructions to be followed. The broader lesson isn't simply a trick for catching AI users; it is the need for clear classroom rules and AI systems that treat untrusted document content as data.

The verified public trail is the dated Korean repost and its exact X link describing the proposed hidden-text trick. No school, outcome or number of students caught has been established. OWASP separately describes hidden instructions in external documents as a form of indirect prompt injection.

The available evidence shows that the idea was posted, not that a trial succeeded. Any later claim of results would need its own documentation and clearly defined scope.