Productive Failure in the Age of AI: Fudan Professor Zhao Bin's Learning Reform
One-Sentence Conclusion
Fudan University School of Life Sciences professor Zhao Bin published a systematic AI-era education reform proposal in spring 2026. His core insight: the traditional "teach-then-practice" model has been pushed to a more dangerous extreme in the AI era — when answers are instantly available, "efficiency" becomes the largest incubator of "ineffective success." Drawing on Singapore-based scholar Manu Kapur's Productive Failure theory, Zhao notes that "practice-then-teach" substantially outperforms "teach-then-practice" on conceptual depth and transfer (effects can reach three times the conventional baseline). He prescribes four remedies for AI-era learners: struggle first before asking for help, multi-round deep questioning, switch off AI to self-review, and correct AI errors. He additionally proposes treating AI as the classroom's "fifth participant," establishing a "content endorsement" system, and guiding students to write a personal ecology textbook over a semester.
Background: Who Is Zhao Bin?
Zhao Bin is a professor in Fudan's School of Life Sciences, an ecology researcher, and a long-standing classroom teacher. His ScienceNet blog records sustained thinking that progresses from PBL (Problem/Project Based Learning) to AI-era education reform. In April 2026, he published the long-form post "In an Age When AI Spits Out Answers in Seconds, Failing First Is the Best Way to Learn," systematically articulating his AI education philosophy grounded in Productive Failure theory. The post has been widely circulated and discussed in Chinese education circles.
The Core Contradiction: AI Takes "Ineffective Success" to an Unprecedented Scale
Problems with Traditional Teaching
Zhao sharply observes that the "three centers, five links" teaching model (teacher-centered, textbook-centered, classroom-centered) used for 76 years has been pushed to a more dangerous extreme in the AI era.
The core logic: explain → demonstrate → practice → correct → consolidate. The goal is clear — minimize mistakes and weld "standard actions" for correct answers into memory.
In cognitive science this is called Algorithmic Thinking — codifying problem-solving into callable, fixed procedures. It handles routine exam questions but breaks down in unfamiliar or innovative scenarios.
AI's Amplifying Effect
Zhao's observation: AI escalates the problem from "low teaching efficiency" to "cognitive offloading." When AI delivers complete answers in seconds, the brain unconsciously enters a "battery-saver mode" —
It no longer actively engages reflective thinking to grasp the underlying structure of a problem; it merely calls up an existing procedure from memory.
This is not AI's fault. AI is a magnifier — it stretches the pre-existing flaws of our teaching logic to the extreme.
The "Learn Fast, Forget Fast" Trap
Zhao borrows a concept from cognitive science: the efficient answers AI provides are fundamentally a form of Cognitive Offloading. You outsource memory and reasoning to the machine; the brain does not undergo the necessary cognitive conflict, and the knowledge never enters long-term memory.
The result: learning feels fast, and after the exam you forget everything. This is not "having learned" — this is "having borrowed." The knowledge belongs to the AI, not to you.
Theoretical Foundation: Manu Kapur's Productive Failure
Four Learning Outcomes
Zhao systematically cites research from Manu Kapur, who directs the Learning Sciences Lab at the National University of Singapore. Kapur divides learning outcomes into four categories:
| Type | Short-term Performance | Long-term Effect | Description | |------|------------------------|------------------|-------------| | Productive Success | Good | Good | Solved correctly after self-exploration | | Ineffective Success | Good | Poor | Completed correctly by copying or peeking at answers | | Productive Failure | Poor | Good | Did not get it right, but engaged in deep thinking | | Ineffective Failure | Poor | Poor | Neither correct nor thoughtfully engaged |
Traditional teaching cultivates precisely "ineffective success" — high scores from fluency with standard procedures, but only surface-level conceptual understanding. Error-avoidance training fundamentally obstructs students from entering the "productive failure" state.
Kapur's Randomized Controlled Experiment
In a 2014 *Cognitive Science* paper, Kapur randomly split students into two groups:
- Teach-then-practice group: instructors explain the concept first, then students practice.
- Practice-then-teach group: students attempt the problem first (likely failing), then receive the explanation.
- The two groups scored similarly on procedural knowledge (fluency with solution steps).
- But the practice-then-teach group significantly outperformed the teach-then-practice group on conceptual depth and transfer.
- "How is this formula derived?"
- "What happens to the result if the conditions change?"
- "Under what circumstances does this method fail?"
- "Is there a more elegant solution?"
- AI-generated content must be reviewed, verified, and rewritten by students.
- The final published work is signed by the student, not the AI.
- Endorsement means accountability: if endorsed AI content contains an error, you are responsible.
- Zhao Bin. "In an Age When AI Spits Out Answers in Seconds, Failing First Is the Best Way to Learn." *ScienceNet 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. "New Conceptualizations of Practice: Common Principles in Three Paradigms Suggest New Concepts for Training." *Psychological Science* (1992).
- Miwa, T. (1987); Nohda, N. (1987). Comparative study of Japanese "structured problem-solving" vs American classrooms.
- Chen Liuqing et al. "Observing a Robot Peer's Failures Facilitates Students' Classroom Learning." *Science Robotics* (2025).
- Fudan University Education Reform 3.0 reporting (2025).
Results were striking:
Zhao highlights an even more counter-intuitive finding: even when students merely observe other students' failed attempts, their learning still outperforms listening to a direct lecture. Failed attempts contain richer learning material than correct answers.
The Mechanism Behind "Fail Miserably, Learn More"
Why does failing first work better? Zhao explains:
When we struggle with an intractable problem, the brain undergoes intense cognitive conflict. That confusion — "why does this path not work?" — is precisely the cognitive fuel required for subsequent understanding of why the correct path is correct.
Cognitive conflict triggered by failed attempts → activates prior knowledge → distinguishes known from unknown → primes students for subsequent precise instruction → deep understanding
This process cannot be skipped. Traditional teaching and AI's direct-answer shortcut bypass not "the long way around" but the biological window in which brain cells rebuild themselves.
Zhao Bin's Four Remedies: What Should AI-Era Learners Do?
Remedy 1: Struggle First, Then Ask AI for Help
This aligns tightly with the Generation Effect: information you actively produce is encoded far more deeply than information you passively receive.
Operating procedure: 1. Block the answer: close all AI tools and reference materials. 2. Set a struggle window: 15-30 minutes. 3. Use pen and paper: sketch diagrams, list hypotheses, reason backward, ask and answer yourself. 4. The goal is not getting it right, but traversing dead ends: at minimum, identify clearly why a path is closed.
The "struggle" itself already activates the brain and primes it for subsequent learning.
Remedy 2: Multi-Round Deep Questioning
Do not ask AI "what is the answer?" Ask instead:
Each round of questioning deepens a cognitive conflict.
Remedy 3: Switch Off AI and Self-Review
This is the most critical step — after AI helps you solve a problem, turn it off and redo the problem from scratch.
If you cannot do it, that means "understanding" is not yet "learning," and you need to return to the previous step and keep asking. Continue until you can independently reproduce the solution.
Remedy 4: Correct AI Errors
Zhao reminds us: AI is not omniscient. When it answers, deliberately look for faults — "Is there a problem here?" "Is this assumption reasonable?"
Being able to spot AI's errors proves you truly understand the problem. That is more valuable than getting a question right.
A Teaching-Practice Revolution: Write a Book in One Semester
Beyond Surface-Level "AI-Assisted Instruction"
Zhao's classroom practice is not simply "use AI to generate test items" or "let AI grade assignments." It restructures the fundamental architecture of teaching.
He guides students through a semester to write a personalized ecology textbook. In this process, AI becomes the "fifth participant" — students and instructor are the first four, and AI joins as an equal fifth creator.
The "Content Endorsement" System
In an age when AI can generate any content, "who is responsible for this content" becomes a critical question. Zhao has built a content endorsement system:
This system resolves a core AI-era teaching problem: AI is a tool, but responsibility lies with people.
Personalized Teaching Becomes Real
Past "personalized teaching" was empty talk — how could one professor personalize instruction for dozens or hundreds of students? AI now lets each student pick their own chapter, articulate concepts in their own way, and explore ecology questions that interest them. The professor's role shifts from "knowledge transmitter" to "learning designer" and "cognitive coach."
A Feynman's-Eye View: "Your Brain Will Not Thank You for Skipping the Long Way"
Feynman would say: "You think learning is the straight-line distance from A to B, but the brain actually grows while it is lost."
Zhao's thesis describes an efficiency paradox in education.
We pursue "fewer detours," "rapid mastery," "efficient learning," but cognitive science tells us: the detours we skip are precisely the most valuable cognitive process.
When AI levels those detours, learning becomes apparently "efficient" yet genuinely "ineffective." Students reach B from A faster, but they never build the intuition that links A to B — the inner conviction that "this path is correct because…"
Feynman would also say: "If you cannot explain it to a smart high-school student, you have not really understood it."
Zhao's "write a book in a semester" practice essentially requires students to explain ecological concepts in their own language, their own logic, their own examples to an imagined reader. Being able to write it out is what real understanding looks like. Being able to write it out and survive peer and AI scrutiny is what deep understanding looks like.
The Cognitive Science Support Chain
Zhao's reform rests on a solid chain of cognitive science research:
| Research | Finding | Year | |----------|---------|------| | Schmidt & Bjork | "Desirable Difficulty": introducing difficulty benefits long-term retention and transfer | 1992 | | Miwa & Nohda | Japanese "structured problem-solving" vs American "clumsy摸索" — American students show far stronger conceptual intuition | 1987 | | Kapur | Productive Failure: practice-then-teach is 3x teach-then-practice | 2008-2014 | | Generation Effect | Actively produced information is encoded far more deeply than passively received | cognitive science classic | | Zhejiang University's Chen Liuqing | "Robot-peer productive failure pedagogy" — observing a robot fail beats direct lecture or the student's own failure | 2025 (*Science Robotics*) | | Kapur's New Book | *Productive Failure: Unlocking Deeper Learning Through the Science of Failing* | 2025 |
The shared direction: learning is not information transfer but cognitive construction. Information can be injected from outside, but cognitive structures can only grow from within.
Limits and Open Questions
1. Boundary Conditions of Productive Failure
Kapur's experiments have been validated mainly in math and science. Do humanities, the arts, and language learning follow the same "practice-then-teach" pattern? Zhao's ecology classroom provides natural-science evidence, but broader validation is needed.
2. Students' Emotional Resilience
"Practice-then-teach" requires students to fail first. Not every student has the psychological stamina to endure repeated frustration. Kapur reports the method works for all students, including low performers — but designing the right degree of failure and providing affective support are central implementation challenges.
3. Scalability of "Write a Book in a Semester"
This model works in Zhao's small ecology classes, but how does it scale to large lectures of hundreds of students? "AI as the fifth participant" requires every student to be fluent with AI tools, raising a fresh digital-divide concern.
4. Operational Details of Content Endorsement
Copyright of AI-generated content, the boundaries of academic integrity, and concrete workflows for responsibility attribution are not detailed in Zhao's post. In practice, these procedural details may be harder to land than the principle itself.