Professor Uses Hidden Prompt to Flag AI Cheating
Alcorn State University professor Jason Gibson reported in late July that a hidden instruction led 32 of 35 students across two classes to submit nonsensical Madagascar references in part of a midterm. Gibson said the responses showed unreviewed AI use; two students appealed their grades and one appeal succeeded, according to later reporting.
Alcorn State University history and African American studies professor Jason Gibson reported that 32 of 35 students across two classes were initially flagged after their midterm responses included nonsensical references to Madagascar. Gibson said he had embedded a white-text instruction inside an Industrial Revolution assignment prompt.
The hidden text directed a model to place “Madagascar” in the response in a way that made no sense. It was not readily visible against the assignment background, but it remained in the text when the full prompt was copied into a chatbot.
Gibson shared examples in TikTok videos, including “Madagascar floats sideways through the afternoon” and “Madagascar purple bicycle whispers to the ceiling.” The phrases were unrelated to the visible assignment and indicated that students had submitted generated responses without reviewing them.
Later reporting added an important grading detail. Gibson said two students appealed after he explained the trap, and one appeal succeeded. NBC/TODAY reported that the successful appeal involved the assignment's display in dark mode. The public reports do not provide the full assignment, grading policy or students' accounts, so the article does not independently assess the final grades.
What the incident establishes
The reported result is evidence of a specific copy-and-paste workflow, not a model-specific detection method. The trap identifies submissions that reproduce a planted instruction; it does not identify which chatbot produced a response, and it cannot establish AI use when the planted instruction is absent.
The originating TikTok videos could not be retrieved directly during this audit because the platform blocked access. The Register, NBC/TODAY, TechSpot and Futurism each retrieved or embedded Gibson's account and consistently reported the core sequence. With no retrieved official or first-party exact-event page, the public source list retains no designated primary.
A prompt-injection lesson
The episode also illustrates a familiar prompt-injection property: instructions can persist in content that appears innocuous or invisible in a user interface. When a person copies an entire document into an LLM, the model processes the document text rather than the person's visual interpretation of it. Here, the hidden instruction competed with the visible academic prompt and appeared in generated answers.
For educators, this technique is better understood as a narrow provenance check than a general AI detector. It can reveal direct, unreviewed copying when a planted instruction appears in a submission. It cannot reliably distinguish permitted AI-assisted drafting, independently written work, paraphrased model output or outputs from systems that ignore embedded instructions.
The same technical risk appears in organizational LLM workflows. Hidden text, web pages, retrieved documents, PDFs and uploaded files can contain instructions intended to alter a model's output. Useful safeguards include separating trusted instructions from retrieved content, inspecting source text before it enters a prompt, constraining tool permissions and requiring human review before high-stakes submission or publication.
Key Points
- 1Gibson reported that 32 of 35 students were initially flagged after submitting nonsensical Madagascar references; two appealed and one appeal succeeded.
- 2The white-text trap demonstrates that LLMs process copied source text, including instructions users may not notice in the interface.
- 3The technique is a narrow provenance check for unreviewed copy-and-paste behavior, not a general or model-specific AI detector.
- 4Comparable prompt-injection risks make content inspection and human review relevant beyond classrooms, particularly in retrieval-augmented workflows.
Scoring Rationale
This is a concrete, widely discussed example of unreviewed AI use and prompt-injection behavior in higher education. It is not a new model, policy or general detection control, but it offers a useful demonstration of copy-and-paste risk for practitioners.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems
