Learning in the AI Era Isn't "Teach Then Practice" — It's "Struggle First, Then Ask for Help"
Professor Zhao Bin of the School of Life Sciences at Fudan University has proposed a systematic approach to education reform in the AI era. His core insight: the traditional "teach first, practice later" model has been pushed to a more dangerous extreme — when answers are instantly available, "efficiency" becomes a breeding ground for *ineffective success*.
Background
Zhao Bin, an ecology researcher and longtime educator, published a long-form essay in April 2026 on his Sciencenet blog: "In an Era When AI Answers in Seconds, Failing First Is the Best Way to Learn." The post has been widely reposted, including by Tencent News, and has sparked broad discussion in education circles.
The Core Problem: AI Amplifies "Ineffective Success"
Zhao argues that the decades-old "teacher-centered, textbook-centered, classroom-centered" model — explain, demonstrate, practice, correct, consolidate — trains algorithmic thinking: solving problems as fixed, callable procedures. It works for routine exams but fails in unfamiliar or creative situations.
AI magnifies this into cognitive offloading. When AI delivers complete answers in seconds, the brain switches to "power-saving mode," skipping the reflective effort needed to grasp deep structure. The result: you learn fast and forget fast. The knowledge is AI's, not yours. AI isn't the problem — it's a magnifying glass exposing flawed teaching logic.
Theoretical Foundation: Manu Kapur's "Productive Failure"
Zhao cites research by Manu Kapur (head of the Learning Sciences Lab, NUS), who classifies learning outcomes into four types:
| Type | Short-term | Long-term | Description | |------|-----------|-----------|-------------| | Effective success | Good | Good | Solved correctly after one's own exploration | | Ineffective success | Good | Poor | Correct by copying or peeking at answers | | Effective failure | Poor | Good | Got it wrong, but thought deeply | | Ineffective failure | Poor | Poor | Wrong and unreflective |
Traditional teaching, Zhao argues, produces precisely *ineffective success*. In Kapur's 2014 randomized controlled experiment (*Cognitive Science*), students who attempted problems before instruction scored similarly on procedural fluency but far exceeded the teach-first group in conceptual understanding and transfer. Even students who merely *observed* peers' failed attempts learned more than those who listened to lectures directly.
The mechanism: struggling with an unsolvable problem creates cognitive conflict — the confusion of "why doesn't this path work?" is the cognitive fuel for later understanding why the correct path is correct. This step cannot be skipped; direct answers from teachers or AI bypass the brain's biological opportunity to rewire itself.
Four Prescriptions for AI-Era Learners
1. Struggle first, then ask AI. Close all AI tools and references; spend 15–30 minutes working with paper and pencil. The goal is not to get it right, but to walk through dead ends (aligned with the *generation effect*). 2. Multi-round deep follow-up questions. Ask how formulas are derived, what happens if conditions change, when a method fails, whether simpler solutions exist. 3. Redo without AI. After AI helps solve a problem, close it and reproduce the solution from scratch. "Understanding" is not "learning" until you can reproduce it independently. 4. Correct AI's mistakes. Deliberately look for flaws in AI answers. If you can find AI's errors, you truly understand the problem.
Classroom Revolution: Writing a Book in One Semester
Zhao's practice goes beyond "AI for quiz generation." He positions AI as the fifth participant in the classroom (after four human roles), and leads students to write their own introductory ecology textbook over one semester.
A "content endorsement" system requires students to review, verify, and rewrite AI-generated content; published work carries students' names, not AI's. Endorsement means responsibility — if AI content you endorsed is wrong, you answer for it. AI is a tool, but accountability is human.
This also makes personalized teaching practical: each student chooses chapters, explains concepts their own way, and pursues their own ecological questions. The professor shifts from knowledge transmitter to learning designer and cognitive coach.
Feynman's Lens
In Zhao's spirit: "The brain doesn't thank you for skipping detours." The efficiency paradox of education — the detours we bypass are the most valuable cognitive processes. And per Feynman's famous test: if you can't explain it clearly, you don't understand it. Being able to *write it out*, and withstand challenges from classmates and AI, is deep understanding.
Supporting Cognitive Science Chain
| Research | Finding | Year | |----------|---------|------| | Schmidt & Bjork | "Desirable difficulties" benefit long-term retention and transfer | 1992 | | Miwa / Nohda | US students' "clumsy exploration" built stronger conceptual intuition than Japan's structured problem-solving | 1987 | | Kapur | Productive Failure: practice-first teaching up to ~3x more effective | 2008–2014 | | Generation effect | Self-generated information is remembered far better | classic | | Chen Liuqing et al. | Observing a robot peer's failures enhances classroom learning | 2025 (*Science Robotics*) | | Kapur | *Productive Failure: Unlocking Deeper Learning Through the Science of Failing*, Harvard Education Press | 2025 |
The common thread: learning is not information transfer, but cognitive construction. Information can be injected from outside; cognitive structures only grow from within.
Limitations and Open Questions
1. Boundary conditions: Productive Failure is mainly validated in math and science; applicability to humanities, arts, and language learning needs more evidence. 2. Emotional resilience: Not every student can withstand repeated failure; designing failure levels and emotional support is a practical challenge. 3. Scalability: "A book per semester" works in small classes; large lectures and unequal access to AI tools pose hurdles. 4. Endorsement details: Copyright, academic integrity, and responsibility procedures for AI-generated content remain underspecified.
References
- Zhao Bin. In an Era When AI 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. *Psychological Science* (1992).
- Chen Liuqing et al. Observing a Robot Peer's Failures Facilitates Students' Classroom Learning. *Science Robotics* (2025).