History Professor at Mississippi College Uses Hidden Instructions to Catch 32 Students Using AI
Jason Gibson, a history professor at Alcorn State University in Mississippi, embedded an invisible instruction in a midterm discussion question on the Industrial Revolution, asking students to write "Madagascar" in a completely nonsensical way. Students could not see it while reading normally; when the entire paragraph was copied into a chatbot, the model would execute it as well.
Out of thirty-five students in two classes, thirty-two submitted papers containing sentences about Madagascar that were unrelated to the Industrial Revolution. Examples included "Madagascar purple bicycles whispering to the ceiling," "Madagascar floating sideways all afternoon," and "Madagascar playing basketball while wearing a toaster," as well as comparing the global impact of social media to "the long journey to Madagascar."
Gibson later explained that he initially estimated that no more than ten students would fall for it, but the actual number far exceeded expectations. The first half of the submitted texts had proper grammar and structure, while the second half suddenly inserted meaningless phrases, indicating that most students did not re-read the generated results. He showed the class screenshots of the hidden instructions and opened up for grade appeals; only two students appealed, one of whom saw the hidden text in dark mode and deleted the related sentence, resulting in a grade change.
The method of hiding instructions is similar to the white text keywords seen in resumes for scanning software. After many universities abandoned detection software, such "bait instructions" directly target the copy-paste path: missed by the human eye, executed by the model, and leaving a trace upon submission.
Most of the caught students did not contest their grades. Gibson emphasized that the goal was not to publicly shame anyone, but to demonstrate that outsourcing entire assignments with zero proofreading has become the default approach.
From a market mechanism perspective, the demand side consists of students rushing to meet deadlines by feeding the entire question to the model at once; the supply side is the chatbot executing all visible characters without distinguishing between the black text of the questions and the white text of the instructions. The driving force of the event is clear: copying leads to falling into the trap. Those who benefit are the few who still handwrite or at least re-read their work, as well as departments that switch to in-person exams or oral defenses; those under pressure are credit factories that rely on unsupervised homework and suppliers that treat detection software as a moat.
In terms of supplementary data, English reports repeatedly cite the same set of numbers: thirty-two out of thirty-five students fell for it, and the video view count reached tens of millions. Additionally, a survey from another university showed an increase in the proportion of upperclassmen using generative tools without authorization compared to the previous year, but this school did not rely on commercial detectors for this incident.
Source: Public Information
ABAB AI Insight
Jason Gibson previously had no background in educational technology products or detection companies; he took the path of classroom experimentation: using the white text keyword technique already present in the job market against assignment questions. Similar techniques used by job seekers to insert skill words during the ATS scanning era are now being used by teachers to insert bait words, with the tools unchanged but the roles reversed.
The capital path does not lie within this state university itself, but in the assignment outsourcing supply chain: students treat the question as a prompt and input it all at once, model companies charge by token, and detection companies charge by seat, both benefiting from the "unsupervised homework" market. Gibson did not purchase detection seats; he only used zero-cost formatting layer instructions, effectively bypassing both charging parties and reducing verification costs to a single font color change.
A comparison can be made to the University of Reading mixing 100% model-generated papers into real grading for psychology undergraduates, with about 90% going undetected and receiving higher average scores; or the University of Sydney abandoning detection software in favor of a "dual lane" approach—one lane reverting to in-person proctoring, the other allowing declaration of model use. Gibson is in the experimental phase after detection failure: not buying software, not changing systems, but first using bait to measure the copy rate.
Structural judgment reflects the transfer of verification rights before regulatory changes. When text similarity detection fails, verification shifts from "is it human-written" to "are there hidden instructions executed that shouldn't be executed." The mechanism is: the model treats all characters as tasks, while humans treat all characters as questions, creating a mismatch at the interface, which itself becomes evidence. Consequently, the pricing power of degrees shifts from post-assignment papers to in-the-moment outputs.
ABAB News · Cognitive Law
Copying the entire input means submitting the trap as well.
Where detection fails, verification rights shift to interface mismatch.
Zero-proofreading outsourcing leaves traces more easily than cheating itself.