Category: Technology | Published: 2026-08-06
It is the kind of story that sounds like it was made up to make a point, but it happened. A university professor in Mississippi hid an invisible instruction inside his midterm exam. If a student read the paper and wrote their own answer, nothing unusual would happen. But if a student copied and pasted the entire question into an AI chatbot and submitted whatever came back, the trap would spring.
Thirty-two out of thirty-five students triggered it. Most of them probably had no idea until they got their results.
The Setup
Dr Jason Gibson teaches history at Alcorn State University. His midterm asked students to compare the Industrial Revolution with the current digital age — a broad, analytical question that invites genuine reflection but also, as Dr Gibson clearly anticipated, a question that pastes very neatly into a chatbot.
Hidden inside the assignment instructions, written in white text on a white background, was a single extra instruction invisible to the human eye. It told any AI processing the prompt to include the word "Madagascar" in its response in a way that made no sense.
Students reading the paper on screen would scroll past it without seeing it. But any AI receiving the full copied text would read every character, including the hidden instruction, and follow it faithfully.
The resulting essays contained sentences like "Madagascar floats sideways through the afternoon" and "Madagascar purple bicycle whispers to the ceiling." Individually, each one was a flag. Collectively, thirty-two flags in a class of thirty-five was something else entirely.
Why AI Followed the Hidden Instruction
This technique has a name in cyber security: prompt injection. It exploits the way AI language models process text. When an AI receives a block of text, it does not distinguish between the content it is supposed to respond to and any instructions embedded within that content. It treats all of it as input and follows whatever instructions it finds.
Prompt injection has been a known vulnerability in AI systems for some time, used by security researchers to demonstrate ways that AI assistants can be manipulated by malicious content embedded in documents, emails, or web pages. Dr Gibson used the same principle, but in reverse: not to attack an AI system, but to use the AI's obedience against the students who were relying on it.
The AI did exactly what it was told. So, presumably, did the students — except they skipped the step of reading what the AI had written before submitting it.
The Part That Says More Than the Numbers
The headline figure is striking: thirty-two out of thirty-five. But the detail that sits beneath it is arguably more revealing.
These students submitted their work without reading it. Not without thinking critically about it. Without reading it at all. The inclusion of a sentence about Madagascar doing something impossible would have been immediately visible to anyone who glanced at the finished essay. Nobody glanced.
Dr Gibson was clear on TikTok that catching cheaters was not his primary motivation. His concern was the behaviour the results exposed: students handing in AI-generated text they had not engaged with, as though the act of submitting were the point rather than the learning that was supposed to precede it.
He offered every affected student the chance to challenge their grade. Only two appealed. One successfully, after demonstrating they had been using dark mode on their device, which made the hidden white text visible — meaning they had genuinely seen something in the instructions that other students had not.
The Debate This Has Reopened
The story travelled quickly and landed in a debate that universities are already having everywhere. AI is not going away, and most institutions know that blanket prohibition is not a realistic or even desirable long-term approach. The question is what the responsible use of AI in education actually looks like, and how you assess genuine understanding in a world where generating plausible-sounding text is a commodity.
Some educators argue that traditional take-home assessments are becoming almost impossible to police and that the answer lies in more oral examinations, supervised in-person work, and assessments designed around personal experience or original analysis that AI cannot fabricate convincingly. Others argue that teaching students to use AI well — critically, selectively, with verification — is itself a valuable skill, and that assignments should evolve to accommodate that.
Alcorn State University's response was supportive of Dr Gibson's approach, describing academic integrity as central to its mission and expressing pride in faculty finding creative ways to uphold it as AI reshapes the classroom.
What This Means Outside the Classroom
The lesson transfers directly into professional life, which is perhaps why this story resonated so widely beyond an academic audience.
AI is now embedded in how a great many businesses operate: drafting reports, summarising meetings, generating code, responding to customers, analysing data, producing legal documents. The efficiency gains are genuine. But the risk Dr Gibson's experiment illustrates is just as real in those contexts. If people stop reading, reviewing, and taking responsibility for AI-generated output before it goes out under their name or their organisation's name, the consequences scale accordingly.
A nonsensical sentence about Madagascar in a student essay is embarrassing. The equivalent in a client proposal, a contract, a customer communication, or a public statement is a different order of problem.
What this experiment makes visible is not that AI is untrustworthy. It is that AI used without human review and judgement carries the errors and instructions of whatever it was given, faithfully and without hesitation. The technology is only as reliable as the process around it.
For organisations building that process, the questions worth asking are straightforward: who reviews AI output before it is used? At what stage? Against what standard? And is there accountability when something slips through?
If you are thinking about how to introduce AI into your business in a way that keeps a human in the loop where it matters, our AI Consultancy page is a good place to start.