Adversarial AI Could Shield Educational Exercises from Cheating

A new study on arXiv (2608.01112) proposes using adversarial machine learning to protect educational tests from AI-driven cheating. It introduces a method involving multimodal multiple-choice questions that have slight visual tweaks designed to mislead AI into selecting certain wrong answers. If students try to copy AI responses, they would show these altered patterns, which would help identify them as non-genuine. The research tests this method in real-world black-box situations with three cutting-edge multimodal language models: Anthropic’s Claude, Google’s Gemini, and OpenAI’s ChatGPT. This work is a reaction to the rise of generative AI, which enables students to avoid engaging in critical thinking and independent thought. The paper was recently published on arXiv.