Productive Failure: How AI Reshapes Critical Thinking - and What Universities Must Do

It was a 40-minute technical interview, and the difference between the two candidates became visible at minute twelve. 

Both had similar backgrounds: AI master's degrees from top European universities, comparable publication records, active GitHub profiles, strong signals on every recruitment metric. We gave them the same task that day: a low-bit CUDA kernel with a subtle bug, to be debugged within 40 minutes. Any AI assistant was allowed. 

Candidate A opened Cursor, pasted the full code, and typed something close to: "Find the bug." The AI produced a confident-sounding explanation. He applied the suggested change, ran the test, failed. He asked the AI for the next hypothesis, applied it, failed again. When time was up, he told us: "The AI seemed very sure, but it was wrong. I'm not sure why." 

Candidate B did not touch AI for the first 12 minutes. He just read the code. Then he asked us a question: "Can I assume the dequantization path is correct?" I said yes. He spent several minutes drawing the data flow on paper. Only around minute twenty did he open an AI assistant. But his prompt was specific: "I suspect the sign-extension happens before the scale factor is applied. Can you walk me through this specific tensor shape to confirm?" This time, the AI was right. He used the remaining 18 minutes to write the fix and the test. 

B passed the exercise. But what stayed with me afterwards was not why B won. It was what A was missing. 

Why Higher Education Must Shift From Product-Based to Process-Based Assessment?

I have been teaching AI at a university for many years, and I have supervised several cohorts of master's and PhD students. Recently I have been following several discussions on how higher education should reform assessment in the age of generative AI. One argument in these discussions resonates strongly with me: in an age when 92% of UK undergraduates already use AI tools regularly (HEPI, 2025), traditional product-based assessment can no longer distinguish a student's own thinking from the contribution of an AI. Universities must shift from assessing the product to assessing the process. Process-folios, continuous assessment, AI transparency statements — the direction is correct. 

But that interview made me realize that something is missing from this conversation. What "process over product" protects is far more than academic integrity. It protects a specific neurological experience — an experience that labor markets are becoming increasingly willing to pay for. 

In other words, the stakes of assessment reform may be much higher than the educational community has yet acknowledged. 

What Is Productive Failure? The Learning Science Behind Struggle?

This neurological experience has a name in the learning sciences: productive failure. It was introduced and systematically studied by Manu Kapur at ETH Zurich, starting in 2008. Its design is counterintuitive: students are asked to attempt a problem that requires a concept they have not yet been taught. They will almost certainly fail. Only afterwards does the teacher step in with the formal explanation. 

The 2021 meta-analysis by Sinha and Kapur covers 166 comparisons drawn from 53 studies, with more than 12,000 participants. The conclusion is clear: students who experienced productive failure significantly outperformed those who received direct instruction first, both in conceptual understanding and in transfer to new problems (Hedge's g = 0.36; ranging from 0.37 to 0.58 in implementations with high design fidelity). In Kapur's own words, this effect size is roughly three times what a very good teacher can achieve in a year. His 2024 book, Productive Failure: Unlocking Deeper Learning Through the Science of Failing, presents the evidence and the design principles to a broader audience. 

The mechanism itself is not mysterious. When students struggle with a problem they cannot yet solve, they are building something in their minds: a conceptual scaffold. They discover the boundaries of their own knowledge, activate relevant prior understanding, and work out through repeated failure which analogies help and which do not. This scaffold is the learning. By the time the teacher finally provides the correct solution, the student is no longer receiving an unfamiliar piece of information — they are filling in a space that their cognitive map has already prepared. 

Robert Bjork calls this more general principle "desirable difficulties": the difficulties that make learning feel slower and more effortful in the moment are often precisely the conditions that make memory durable and knowledge transferable to new contexts. 

Put more directly: struggle is not the cost of learning. Struggle is the learning. 

The Interview Question That Reveals Cognitive Offloading 

About a year ago, we added a question to the interview process for every AI/ML role: 

"Tell me about a technical problem in the past two years where you were genuinely stuck for more than three days. What did it feel like? What did you try? When did you get unstuck?" 

Three years ago, this question produced detailed, concrete stories. In our recent interviews, around 50% of young candidates give a vague or evasive answer — not because they do not want to talk, but because the memory simply is not there. One candidate answered with complete honesty: "I don't really get stuck for that long anymore. If something is hard, I ask Claude or ChatGPT, and I usually get unstuck quickly." 

He did not realize this was a red flag. In his view, "not getting stuck anymore" was progress. From my perspective on the hiring side, it meant that his brain did not contain the neural pathway that forms when someone searches in the dark for three days and then sees a small point of light. And in the kind of work we do — low-bit quantization, MoE architectures, new optimization objectives — we encounter this kind of moment every week. 

When an optimization convergence problem holds you captive for two weeks, no AI tool can rescue you. No training data covers the specific edge case you are tuning. In those moments, an engineer who has never independently survived a productive struggle freezes. An engineer whose brain does contain that pathway knows how to keep moving forward, even when the next step is still in the dark. 

AI-as-Tool vs. AI-as-Shortcut: The Distinction Universities Must Teach 

The candidates most fluent in AI might be the least suited to technical work at an AI company. Back to university assessment, my understanding of "process over product" has changed. 

Process-folios protect more than academic integrity. They protect the narrow window in which a student actually experiences productive failure. The value of continuous assessment is not that it is harder to cheat — it is that it normalizes struggle and makes failure a low-cost event. An AI transparency statement is not a compliance instrument; its real function is to force the student to reflect on their own cognitive state before reaching for AI. 

What universities must do in the age of AI is not choose between "allow AI" and "forbid AI" — that binary decision is too cheap. The real work is to draw a precise distinction between AI-as-tool and AI-as-shortcut. The first lets a student, once the cognitive scaffold is already in place, amplify their output with AI. The second allows AI, before the scaffold has been built, to place an answer directly in front of the student — letting them skip the very struggle that was supposed to build the scaffold. 

The Research: How AI Use Correlates With Declining Critical Thinking 

Recent empirical research is confirming the importance of this distinction. Gerlich's 2025 study of 666 participants found a significant negative correlation between frequent use of AI tools and critical thinking ability, mediated by cognitive offloading. A 2025 experiment from the MIT Media Lab by Kosmyna and colleagues, conducted over four months, observed that participants in the LLM group consistently lagged behind the control group across neural, linguistic, and behavioral indicators, and reported the lowest sense of ownership over their own work. 

AI itself is not the problem. AI without the scaffolding phase is the problem. The real design task of assessment reform, then, may not be to "detect AI" — but to protect struggle. 

I am increasingly convinced that the real differentiation in higher education over the next five years will not come from "whoever embeds AI most deeply into the curriculum." That is a commodity; no institution can win on that ground. The real differentiation lies in a different dimension: which university can clearly tell its students when to use AI, and when to turn it off. 

Back to the scene at the beginning. Candidate B waited 12 minutes before opening Cursor. Those 12 minutes were not hesitation, nor unfamiliarity with the tool. They were an instinct — an instinct that had been trained into him, at some stage of his education, by someone, through some method. 

I would like to know where that stage, that person, and that method were. 

And I hope the next generation of Candidate Bs will be trained in a university that is genuinely willing to teach this. 

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. P. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press. 

Freeman, J. (2025). Student Generative AI Survey 2025 (HEPI Policy Note 61). Higher Education Policy Institute & Kortext. https://www.hepi.ac.uk/wp-content/uploads/2025/02/HEPI-Kortext-Student-Generative-AI-Survey-2025.pdf 

Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006 

Kapur, M. (2008). Productive failure. Cognition and Instruction, 26(3), 379–424. https://doi.org/10.1080/07370000802212669 

Kapur, M. (2024). Productive failure: Unlocking deeper learning through the science of failing. Wiley. 

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv preprint, arXiv:2506.08872. https://doi.org/10.48550/arXiv.2506.08872 

Sinha, T., & Kapur, M. (2021). When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5), 761–798. https://doi.org/10.3102/00346543211019105