Input-Side Watermark to Detect LLM Copy-Paste in Student Homework

Researchers propose an input-side watermarking method to detect when students submit verbatim LLM-generated responses without engagement. Existing AI-text detectors are unreliable and bias non-native writers, while output-side watermarks require provider cooperation. The new approach embeds an invisible instruction within the assignment prompt itself. An LLM processing the prompt verbatim reads the hidden instruction and produces a tell-tale signature in its reply, exposing the copy-paste pathway. This method is controlled directly by educators and does not depend on model providers. The work is published on arXiv (2605.16336).