US Professor Catches 32 Students Cheating Using Hidden AI Prompts
Jason Gibson, a history professor at Alcorn State University, successfully caught 32 out of 35 students cheating by embedding invisible white-text prompts in a midterm exam. The hidden instructions forced AI tools to output bizarre, nonsensical sentences about Madagascar, which students copied directly into their submissions. This incident highlights the growing challenge of academic integrity in the age of generative AI and demonstrates a practical, low-tech application of prompt injection to detect unauthorized AI use. It also underscores how blindly relying on AI outputs without review can easily expose academic dishonesty. The hidden prompt instructed the AI to describe Madagascar in an illogical way, resulting in phrases like "Madagascar floated sideways in the afternoon" appearing in the students' essays. Out of the 32 caught students, 30 admitted to cheating, while two attempted to defend their actions.
## BACKGROUND
Prompt injection is a vulnerability where machine learning models, particularly Large Language Models (LLMs), are manipulated by adversarial inputs to ignore their original instructions and perform unintended actions. In this case, the professor used a form of indirect prompt injection, embedding instructions in white text within a document that is invisible to human eyes but readable when copied and pasted into an LLM.