Productive Failure in the AI Era: Fudan Professor Zhao Bin's Guide to Learning Before Asking AI

Core Thesis

Professor Zhao Bin of Fudan University's School of Life Sciences argues that the traditional "teach first, practice later" teaching model is pushed to a dangerous extreme in the AI era: when answers are instantly available, "efficiency" becomes the breeding ground for what cognitive scientist Manu Kapur calls ineffective success — getting correct results without genuine understanding.

Citing Kapur's Productive Failure theory, Zhao contends that "practice first, teach later" substantially outperforms direct instruction (reportedly up to three times more effective) in conceptual understanding and knowledge transfer.

Why AI Amplifies the Problem

• The decades-old "three centers, five stages" model (teacher-, textbook-, classroom-centered) trains algorithmic thinking: fixed procedures that handle routine exams but fail in novel situations.
• When AI answers in seconds, the brain shifts to "power-saving mode" — cognitive offloading. Memory and reasoning are outsourced; knowledge never enters long-term memory.
• "You learned fast and forgot fast. You didn't learn it — you borrowed it."

Kapur's Four Quadrants of Learning Outcomes

| Type | Short-term | Long-term | Description | |------|-----------|-----------|-------------| | Productive success | Good | Good | Solved correctly after own exploration | | Unproductive success | Good | Poor | Copied or peeked at answers | | Productive failure | Poor | Good | Got it wrong but thought deeply | | Unproductive failure | Poor | Poor | Neither correct nor thoughtful |

Traditional teaching, Zhao argues, optimizes for unproductive success. In Kapur's 2014 *Cognitive Science* randomized experiment, both instruction orders scored similarly on procedural skill, but the practice-first group far exceeded the lecture-first group in conceptual depth and transfer. Even students who merely *observed* peers' failed attempts learned more than those who heard a lecture directly.

The mechanism: failed attempts create cognitive conflict → activate prior knowledge → distinguish known from unknown → prepare the mind for deep understanding. This step cannot be skipped.

Four Prescriptions for AI-Era Learners

1. Struggle first, then ask AI. Block all answers, spend 15–30 minutes with pen and paper. The goal is not to be correct but to understand why dead ends are dead ends (aligning with the generation effect). 2. Ask multi-round, deep questions — "How was this derived?" "When does this method fail?" "Is there a simpler solution?" 3. Turn off AI and redo the solution yourself. If you cannot reproduce it, "understanding" was illusion. Repeat until you can. 4. Find errors in AI's answers. Spotting AI mistakes signals genuine understanding — more valuable than a correct answer.

Classroom Revolution: Writing a Book in One Semester

Zhao's teaching practice goes beyond "AI for quizzes": students spend a semester writing their own ecology textbook, with AI as a "fifth participant" (students and teacher being the first four). A "content endorsement" system holds students responsible: AI-generated content must be reviewed, verified, and rewritten, and students — not AI — sign their names. The professor's role shifts from knowledge transmitter to learning designer and cognitive coach.

Supporting Cognitive Science

• Schmidt & Bjork (1992): desirable difficulties improve retention and transfer
• Miwa & Nohda (1987): exploratory American classrooms produced stronger conceptual intuition than structured Japanese instruction
• Kapur (2008–2014): Productive Failure research; new book *Productive Failure* (Harvard Education Press, 2025)
• Generation effect: self-produced information is remembered far better
• **Chen Liuqing et al. (2025, *Science Robotics*): observing a robot peer's failures outperforms direct instruction

Limitations and Open Questions

1. Boundary conditions: Productive Failure is mostly validated in math and science; humanities applications need more evidence. 2. Emotional resilience: repeated failure requires psychological support and careful design. 3. Scalability: "write a book in a semester" works in small classes but is hard in large lectures; AI access raises digital-divide concerns. 4. Endorsement details:** copyright, academic integrity, and accountability procedures remain underspecified.

Sources

• Zhao Bin, "AI秒出答案的时代，先失败才是最好的学习方法," Science Net blog (2026-04-26). https://blog.sciencenet.cn/blog-502444-1532159.html
• Kapur, M. Productive Failure in Learning Math. *Cognitive Science* (2014).
• Kapur, M. *Productive Failure: Unlocking Deeper Learning Through the Science of Failing*. Harvard Education Press (2025).
• Schmidt, R.A. & Bjork, R.A. *Psychological Science* (1992).
• Chen Liuqing et al. *Science Robotics* (2025).