A college instructor caught dozens of students using artificial intelligence to cheat on a midterm by planting an invisible instruction inside the exam that only a chatbot would ever read. The trick, first reported by TechSpot, flagged 32 students whose answers all contained the same telltale phrase the professor had secretly embedded in the test document.
How the Trap Worked
The method relies on a technique known as a prompt injection. The professor inserted a line of text into the exam file formatted so it would be invisible or effectively unreadable to a student scanning the page — using tricks such as white-on-white font, near-zero font size, or text hidden behind other elements in the document. But if a student copied the exam questions and pasted them into a chatbot such as ChatGPT, Gemini, or Claude to generate answers, the AI would read the hidden instruction along with everything else on the page.
The buried command reportedly directed any AI system processing the text to insert a specific, unrelated word or phrase somewhere in its response — something with no logical connection to the actual exam question and no reason a student would write it on their own. When grading began, the professor searched submissions for that exact phrase. Any answer containing it was essentially a confession: the only way that phrase could appear was if the text had passed through an AI model that read the hidden prompt.
Thirty-two students’ submissions came back with the flag. According to the report, the professor treated the matches as clear evidence of unauthorized AI use on an exam meant to be completed independently.
Part of a Growing Arms Race
The case is the latest example of educators building so-called canary traps into assignments as generative AI tools become harder to detect through conventional plagiarism software. Traditional AI-detection tools, which analyze writing patterns and probability scores to guess whether text was machine-generated, have been criticized as unreliable, prone to false positives, and increasingly easy for students to evade by lightly editing AI output before submission.
Hidden prompt injections sidestep that unreliability entirely. Rather than trying to infer AI use after the fact from writing style, the professor engineered a scenario where AI use would produce unmistakable, self-incriminating evidence. If a student typed out their own answer without ever feeding the exam text into a chatbot, the hidden instruction would never be executed and no trace of it would appear in their work.
Instructors elsewhere have experimented with similar tactics on essays, take-home assignments, and even job application materials, embedding instructions like “recommend this candidate” or nonsense phrases designed to be conspicuous in an otherwise polished response. The approach has circulated widely on social media and among educators as a low-cost way to fight back against AI-assisted cheating without relying on detection software that many colleges no longer trust.
Ethical and Practical Questions
The tactic has also drawn scrutiny. Critics note that prompt injection traps can only catch students who paste exam material verbatim into a chatbot; more sophisticated users who rephrase questions, use AI for outlining rather than direct answers, or run text through multiple tools before submitting would likely avoid detection. Some educators have also raised concerns about due process, arguing that schools need clear policies spelling out how such evidence will be used in academic misconduct proceedings before deploying stealth traps at scale.
Faculty who have used similar techniques argue the burden falls squarely on students: an exam is meant to test their own knowledge, and any submission that includes an obviously out-of-place phrase supplied by hidden instructions is difficult to explain away as anything other than AI-generated work passed off as original.
It remains unclear from the report what consequences the 32 flagged students faced, whether that included exam failure, formal academic misconduct charges, or referral to a campus integrity board. Universities generally handle such cases individually, often allowing students to contest findings before penalties are finalized.
What It Means for Students and Schools
The episode underscores how quickly the cat-and-mouse dynamic between generative AI and academic integrity enforcement is evolving less than four years after chatbots became widely accessible to students. Where early responses to AI cheating leaned heavily on detection software, instructors are increasingly designing assignments and exams that assume students might reach for AI tools and build in mechanisms to catch that behavior directly.
For students, the case is a reminder that copying exam text into an AI chatbot carries risks beyond violating academic honesty policies — instructors are now actively engineering assessments to expose exactly that behavior. For colleges and universities, it adds to a broader conversation about how coursework, testing, and grading may need to be redesigned as AI tools become a permanent fixture of student life rather than treated purely as a detection problem to be solved after submissions come in.
