To prevent 'AI cheating', test papers to include 'AI traps' - attention drawn to reverse research approach
Researchers are exploring ways to prevent AI-powered cheating on exams by tricking AI into giving incorrect answers or refusing to respond. A team led by Professor Max Song of Soongsil University's AI Safety Research Center conducted experiments using two AI models, Alibaba's Qwen3-VL-2B and Anthropic's Claude Fable 5.2 and Claude Haiku 5.5.
The team found that adding fake instructions or distorting graph information could mislead some AI models into giving incorrect answers, while adding a declaration stating that the exam is in progress and AI assistance is not allowed could also be effective in some cases. Similar research is being conducted at Arizona State University, where a system has been developed to automatically insert AI-disrupting information into exam papers.
Summarised by urgent.news, not translated. Machine-written; may contain errors — check the original.
This story
This is one outlet's version. Read the fullest account.