Professor Jason Gibson hid white text inside a midterm prompt. Chatbots read it; students didn’t. Thirty-two of his 35 students failed part of the exam.
Gibson teaches history at Alcorn State University in Mississippi. The midterm asked students to compare the Industrial Revolution with today’s digital age.
Hidden in white font was an instruction telling any AI to slip random Madagascar references into its answer. Students who pasted the prompt into a chatbot never saw it.
The essays came back polished but bizarre. One said, “Madagascar floats sideways through the afternoon.” The assigned topic was the Industrial Revolution.
Gibson posted the story on TikTok on July 20, 2026. The videos reportedly drew millions of views. He let students appeal, and only two tried.
Only one appeal succeeded. That student said she noticed the hidden text because her device was in dark mode.
Gibson said no student had requested an accommodation to use AI. He also said he takes no pleasure in failing anyone and does not plan to make a habit of setting traps.
Why the Trap Worked
The trap was not a fake exam question. It was a hidden instruction tucked inside a real one.
Chatbots read every character you paste, including text you can’t see. White text on a white page looks blank to you, not to the AI.
The chatbot treated the hidden line as part of the assignment and obeyed it. Security researchers call this trick a prompt injection.
The second failure was human. Nobody proofread.
A quick read would have caught the line about Madagascar wearing a toaster to a basketball game.
A related weakness is hallucination. When a prompt contains a false premise, a chatbot can answer as if it were true.
That happens because AI tools predict likely-sounding text. They don’t check whether an instruction or premise makes sense.
Both cases end the same way: confident, fluent text that no human checked.
This Isn’t an Isolated Case
Brown University economics professor Roberto Serrano spotted a different warning sign: a huge grade gap.
This spring, Serrano gave a take-home midterm for the first time in nearly two decades. He offered it after a December campus shooting left students anxious about classrooms.
The class average was 96%. Historically, it runs 65% to 80%. Forty of 86 students earned perfect scores.
Serrano and his graders ran the questions through ChatGPT. The answers looked much like his students’, including an oddly convoluted proof.
He then made the final exam in person and closed-book. The comparison was stark.
| Exam | Format | Class Average |
|---|---|---|
| Midterm | Take-home, unmonitored | 96% |
| Final | In person, closed-book | 48.6% |
Eighteen students dropped the course. Nine more skipped the final. Nineteen ultimately failed.
Serrano voided the midterm and made the final worth 80% of the grade.
These are a professor’s suspicions, not findings against named students. Some observers add that exam conditions and time pressure can also lower scores.
Brown’s academic code committee reportedly asked him to file a separate complaint for each suspected student.
What Students Should Actually Take Away
The line between AI help and cheating is set by your instructor, not by a national rule.
Policies vary, but many sort AI use roughly like this:
- Often allowed: brainstorming, explaining concepts, and making practice questions, when the syllabus permits AI.
- Gray area: asking AI to rewrite your sentences or summarize graded readings.
- Usually cheating: submitting AI-written work as your own when AI is banned or undisclosed.
Rules can also change by assignment. One professor may ban AI on exams and require it on projects.
Three habits protect you:
- Read the AI policy in every syllabus.
- Ask in writing when the rules are unclear.
- Save drafts and version history as you write.
Also, read what you submit. Gibson’s own advice was simple: if you use AI, at least proofread the result.
How Schools Are Responding
Many professors are moving graded work back into the room.
- Blue-book exams: handwritten, in-class writing with no devices.
- Oral exams: Cornell’s Chris Schaffer uses an “oral defense,” and NYU is expanding oral assessments.
- In-class writing and presentations: shorter tasks completed under supervision.
- Process-based assignments: drafts, outlines, and version history count toward the grade.
Detectors are losing ground. Yale, Vanderbilt, Johns Hopkins and Indiana University have policies that bar or discourage using them as sole evidence.
At least a dozen universities, including Northwestern, Georgetown and NYU, have also disabled Turnitin’s AI detection tool.
The shift has critics. Some educators say timed, handwritten tests can disadvantage multilingual writers and students with disabilities.
Online, self-paced courses also make in-person exams impractical for many students.
FAQ
Can AI detectors be wrong?
Yes. Studies have found AI detectors produce false positives on human writing.
Research also shows bias against non-native English writers. Some accused students have filed lawsuits against their universities.
That’s why many schools now bar detector scores as the sole proof of cheating.
Is using Grammarly cheating?
It depends on the feature and your course policy. Basic spelling and grammar checks are generally treated differently from tools that rewrite or generate text.
Generative features can trigger AI detectors. Notre Dame, for example, counts AI editing tools as generative AI when an instructor bans it.
When in doubt, turn generative features off or ask your instructor first.
What happens if you’re falsely accused?
Process varies by school, but these steps are common:
- Stay calm and don’t admit to anything you didn’t do.
- Ask, in writing, what evidence the instructor has.
- Gather drafts, version history, notes, and sources.
- Request a meeting with the instructor.
- File a formal appeal if the meeting fails.
Your school’s academic integrity office or student advocate can explain your rights.