Fudan University's "Data Mining Technology" course recently held an unconventional final exam. Unlike traditional exams where students sit in a classroom answering questions, this final exam had the entire class take on the role of "question setters."
According to the assessment rules, each student needed to independently design 10 data mining calculation problems with unique correct answers and complete derivation processes, and use these questions to test three AI models at different ability levels. The scoring logic was highly innovative: the more AI models were stumped, and the harder the questions that AI got wrong, the higher the score the question setters received.
Traditional Exams Are Ineffective in the AI Era, Forcing Changes in Teaching Methods
The course leader, Professor Xiao Yanghua from Fudan University's School of Computing and Intelligence Innovation, pointed out that traditional question-setting methods have lost their original technical testing significance in today's AI era. If teachers still set standard algorithm questions, AI's computing speed and accuracy can easily surpass any human student.
Professor Xiao believes it is meaningless to directly compete with AI in areas where AI excels, so he decided to completely reverse the assessment direction. The core of this "human vs. AI" design lies in guiding students to firmly believe that as long as they truly deeply and thoroughly understand professional knowledge, they will be able to keenly identify the logical blind spots of AI.
The Strongest Model Wasn't Cracked, Only 4 Students Made AI Score Zero
In the 51 final answer sheets submitted, the test results showed extremely dramatic data comparisons. Out of the 50 students in the class, 50 managed to make at least one AI model fail in answering questions, while only one student failed to stump any model. The class overall average score was 85.7 points.
Notably, among the three models facing the challenge—DeepSeek V4-Flash, MiniMax M2.7, and Claude Sonnet 4.6—the top-tier Claude model demonstrated strong resilience and was not completely stumped by any student. Ultimately, only 4 students created high-level original questions that successfully made the weaker AI model score zero on the entire test paper, securing full marks for the final exam.
