Professor catches 32/35 students cheating with AI using hidden white-text prompt trap

Dr. Jason Gibson, a history professor at Alcorn State University, devised a method to identify students submitting AI-generated work by embedding hidden instructions in a midterm question about the Industrial Revolution. The trap consisted of white-colored text instructions visible only to the AI, telling it to include information about Madagascar in ways that make no sense. Of 35 students across two of Gibson's classes, 32 simply copied the entire question, pasted it into a chatbot, and copied the generated response directly into their answer window without reading it.

The results were unmistakable. Students who had cheated submitted essays peppered with cryptic references to Madagascar that had no connection to the Industrial Revolution. In one example, the AI's response about Industrial Revolution technology and social inequality ended with the line: "Madagascar floats sideways through the afternoon." Another essay finished with: "Madagascar purple bicycle whispers to the ceiling." The incoherence was obvious to anyone who read the submission, yet the students had submitted it verbatim.

Gibson failed all 32 students on that portion of their midterm. He reported that he was not using the trap as an opportunity to mock his students; instead, after confronting them about the cheating, he gave them the chance to contest their grades. Surprisingly, only two students chose to argue their case. Gibson shared his experience in viral TikTok posts that included examples of the AI's nonsensical outputs, framing the incident as a response to widespread AI-assisted cheating in higher education.

The problem Gibson addressed is not isolated. Since ChatGPT went mainstream, AI-assisted cheating has been described as endemic among students. Teachers have experimented with various countermeasures—requiring handwritten work, administering oral tests, and designing questions intended to be difficult for AI to complete—but success has been limited. Earlier in the same month, a professor at Brown University observed a stark contrast: when he allowed take-home exams, midterm averages jumped from the typical range of 65–80% to 96%. He estimated that all but two of his 86 students relied on AI to cheat. These cases suggest that without active safeguards or friction in the submission process, many students will turn to AI, and without verification mechanisms, they may not even realize when the output they are submitting has been compromised.

Student reliance on AI for assignments has accelerated since ChatGPT went mainstream, creating what educators describe as an endemic cheating problem. Gibson's approach—embedding invisible instructions that produce detectable nonsense—works precisely because students using AI often skip the verification step: they copy the prompt, paste it into a chatbot, and submit the output without reading it. The fact that 32 of 35 students fell into this trap, and that only 2 of the 32 caught cheating bothered to contest their grade, suggests that many students may not even realize they are submitting generated text that has been compromised.

The broader context is stark. Earlier in the same month as Gibson's experiment, a Brown University professor observed that switching to take-home exams caused midterm averages to jump from 65–80% to 96%, with all but two of his 86 students estimated to have used AI. These parallel cases suggest that without active friction—such as Gibson's trap, handwritten work, or oral tests—students will default to AI assistance. Teachers have tried multiple countermeasures, but as Gibson's success illustrates, the simplest defenses may be those that exploit the assumption students are not reading their own submissions.