Dr. Jason Gibson of Alcorn State University used hidden white text in exam assignments to detect mass usage of generative models by students.
What Happened
In the exam instructions, the professor inserted the word "Madagascar" in white font on a white background. This text was invisible to humans but was picked up by neural networks when the assignment was copied into chatbots. As a result, students submitted work containing absurd phrases such as "Madagascar floats sideways through the afternoon," allowing for the unambiguous identification of AI use without even checking the actual answers.
Context
The method exploits the mechanism of blind copying, where users copy an entire block of text, including invisible elements that LLMs interpret as part of the context or instructions. Technically, this can be classified as a form of indirect prompt injection or the use of canary tokens in a text stream to validate data provenance. This case highlights the gap between human perception of text and how tokenizers and model context windows process it.
Why It Matters for the Industry
For the industry, this case demonstrates the simplicity and effectiveness of prompt injection methods for verifying AI usage quality and detecting automated plagiarism. It opens possibilities for creating verification services and adversarial testing tools to check the resilience of RAG systems against hidden instructions. In the long term, this may lead to the standardization of sanitization protocols for input context to remove invisible characters in production-ready AI pipelines.
Why It Matters for Users
For readers and users, this is a clear example of how a simple check for attentiveness and the use of hidden data can expose blind copying of chatbot responses. Such methods could become an accessible way to verify content authenticity in educational and corporate environments.
What Is Not Yet Known / Limitations
There is a difference in how the method is evaluated: while technical specialists view it as a primitive vulnerability requiring improved evaluation methods (evals), product developers see it as a potential business opportunity for verification services. Additionally, future developments in text cleaning methods may render such simple traps ineffective.
Sources
Author
Look at AI, Editorial Staff