Professor Hides White Font in Midterm, Catches Students Using AI in the Stupidest Way Possible
TL;DR
Thirty-two of 35 students across two classes failed part of their midterm after generating their entire answers with AI. The professor had hidden an instruction in white font inside the exam prompt — invisible to readers, but picked up by any language model the text was pasted into. Every student who let AI write the answer dutifully included the hidden instruction and exposed themselves. The trick costs nothing and works anywhere text gets passed along digitally.
Nauti's Take
A cheap, effective step forward for anyone grading written work: no costly detector software, no false accusations against honest people, just one camouflaged line in the file. The risk cuts both ways — the same prompt-injection technique steers models reading résumés, proposals or contracts.
A free tool for teachers, and an open question for everyone else about which untrusted documents they feed a model.
Sources