Home General The Invisible Word That Trapped Dozens of AI-Cheating Students

The Invisible Word That Trapped Dozens of AI-Cheating Students

0
8
Conceptual illustration of an exam paper with hidden text revealed by an AI, symbolizing cheating detection

An American professor has found a radical way to check whether his students were using an artificial intelligence assistant to write their papers: hiding, within the exam prompt itself, an invisible instruction aimed at AI. The result of this experiment, widely shared across the Atlantic and later in the French-language press, has reignited the debate over the true extent of AI-assisted cheating in higher education.

An absurd word slipped in using white text

Jason Gibson, a history professor at Alcorn State University in Mississippi, inserted into his midterm exam a instruction written in white text on a white background, making it invisible to the naked eye for a student reading the prompt on screen. This instruction asked any artificial intelligence processing the text to work the word “Madagascar” into its answer, with no apparent justification.

A student who writes their own answer never sees this instruction. But a student who pastes the exam prompt into a generative AI tool to get a ready-made answer unknowingly passes it along to the algorithm, which carries it out faithfully. The result: nonsensical sentences turned up in numerous papers, such as “Madagascar floats sideways in the afternoon,” or equally incongruous references slipped into the middle of otherwise perfectly structured paragraphs.

32 out of 35 students caught in the trap

Across two classes totaling 35 students, 32 papers contained the telltale mention, proof that their authors had delegated the entire writing process to an AI without even proofreading the result. Many of the papers were also written in a more polished style than usual, but the presence of the incongruous word was enough to give away their origin. The professor explained that he wanted to concretely illustrate a phenomenon that teachers had long suspected without always being able to prove it.

This anecdote illustrates a well-known cybersecurity technique, prompt injection, usually used to test how AI systems hold up against instructions hidden within content they analyze. Here, the method was repurposed for educational rather than malicious ends.

A problem bigger than a single campus

Gibson's case is not an isolated one. In the United States, several institutions have recently tightened their practices in response to the massive use of generative AI in academic work, with some abandoning the honor system in favor of strictly supervised exams. In France too, the issue is stirring up higher education, where teachers and students are still searching for the right balance between legitimate use of AI tools as learning aids and outright delegation of intellectual work.

For institutions, the stakes go beyond simply catching cheaters: the real issue is rethinking assessment methods at a time when these tools are available to everyone within seconds. Several education experts are already calling for oral exams or supervised testing conditions, seen as harder to get around than a take-home assignment submitted online.

Frequently Asked Questions

What is prompt injection applied to an exam?

It involves hiding, within a text meant to be read by an AI, an additional instruction that is invisible to a human reader but is carried out by the algorithm if it processes the document. In this case, the instruction asked for an absurd word to be inserted into the generated answer.

How were the students caught?

By pasting the exam prompt into a generative AI tool to get an answer written for them, they unknowingly passed along the hidden instruction, which the AI carried out by slipping the trap word into the text it produced.

Could this method become a common practice?

It remains a one-off tactic and isn't foolproof: it only works if the student copies the entire prompt without altering it. But several instructors are already drawing inspiration from it as a complement to other detection methods.

Sources

L'Avenir

The Register

Planet.fr

Author Profile

Ethan Cole

This article was written with the help of artificial intelligence. Editorial policy

Previous articleNonna Maxxing: The 2026 Lifestyle Trend Bringing Back Italian Grandmothers’ Pace of Life
Next articleMandatory E-Invoicing: All French Businesses Affected This Fall

Editorial policy and use of artificial intelligence