Fudan University uses 'stumping AI' as final exam, students design questions to challenge models

Fudan University's 'Data Mining' course recently held a unique final exam. Instead of sitting in a test room, students became question creators, designing 10 computational problems to challenge three AI models at different levels. The more the AI got wrong, the higher the student's score, testing their deep understanding of the subject.

The exam used three models: DeepSeek V4-Flash, MiniMax M2.7, and Claude Sonnet 4.6, each with a different difficulty tier. Each wrong answer by an AI earned students between 1.5 and 3 bonus points. Out of 51 exam papers, 50 students managed to stump at least one AI, but only 4 got a full zero on any model. The strongest model, Claude, was never fully beaten.

Professor Xiao Yanghua said traditional testing is outdated in the AI era, as AI outperforms humans on standard algorithm problems. The new format aims to show students that truly understanding knowledge helps find AI weaknesses—it's a skill, not luck. The class average was 85.7 points, with a median of 88.