At Fudan, the new exam: students grade the AI, not the other way around

At Shanghai's Fudan University, the harder Claude and DeepSeek struggle, the higher the student scores. In a June 2025 data-mining course taught by Prof. Xiao Yanghua, 51 undergraduates each designed ten questions intended to stump three frontier AI models, and grading rewarded the question, not the answer. The stronger a question broke a model, the more points the student earned. It is the first published scoring rubric of what Beijing calls "AI Plus Education," the national push to weave AI into every layer of schooling by 2030 (Fudan University).

"If we ask students to solve a standard algorithm, AI will always be faster," Xiao said. "We must move toward training them to judge, direct, and identify the structural flaws of these models." His read on the labor market follows the same line. "The most important competitive advantage in the AI era is the ability to use AI and judge AI" (TMTPOST).

Each student submitted ten questions. Grading started at 60 points for a compliant set, with bonus weight when a target model got the question wrong. DeepSeek V4-Flash failures were worth 1.5 points each, MiniMax M2.7 failures 2 points, and Claude Sonnet 4.6 failures 3 points, with a 100-point cap. Out of 51 students, 50 stumped at least one model on at least one question. Only four forced any single model to score zero across their full paper. Claude, the strongest of the three, was not fully stumped by anyone. Class average landed at 85.7, median 88, a result that grades the cohort's adversarial reach rather than their ability to recall a textbook (Fudan University).

The shift from prompt-and-use to adversarial audit is the operating theory behind China's "AI Plus Education" roadmap, which targets a comprehensive AI-driven education system across all learning levels by 2030. Xiao's exam is not a national rollout, but it is the kind of course design the roadmap points toward. A newscentralasia policy layer puts the stakes in plain English: a generation schooled to challenge the model, not defer to it (newscentralasia.net).

Fendou Primary School in Beijing uses AI-generated interactive 3D models to teach geometry, putting a manipulable shape on a concept younger students struggle to hold in the head. Beijing No. 13 High School has deployed AI agents that track each student's digital learning footprint and produce individualized "learning health reports." Beijing No. 35 Middle School runs an AI code reviewer that diagnoses student programming errors in under two minutes, faster than most teachers can scan the same line. At Shandong College of Tourism and Hospitality, a "Baize" AI digital assistant simulates hotel check-ins and guest services so vocational students can rehearse service scripts in a sandbox (newscentralasia.net).

"Traditional testing methods are facing new challenges in the age of AI," he said, casting the inverted exam as a response to a labor market that no longer rewards recall (TMTPOST Chinese). The Fudan scoring tiers encode that response. A student who can find a hole in Claude picks up bonus weight that a correct textbook answer never does.

The Fudan mechanism is one course with one cohort, and the 2030 roadmap is policy direction, not audited rollout. The 50-of-51 result and the 85.7 average describe a class of students who were already strong enough to engage the models at all. The next test is whether a national curriculum borrows the Fudan scoring key, or whether the inverted exam stays a single-course experiment.