The Algorithmic Crutch: How AI Tutoring Is Quietly Dismantling Mathematical Understanding

Albus Duc Hoang, VCTLI

There is a particular kind of fraud being committed in mathematics classrooms right now, and it is being committed with the best of intentions. Across schools at every level, AI-powered tutoring platforms are being deployed on the premise that personalized, immediate, friction-free guidance will finally democratize access to mathematical learning. Students who would once have sat with a blank worksheet now have a tireless digital companion stepping in the moment confusion appears. Frustration, that old enemy of the math classroom, is to be engineered away. Investors are delighted. Administrators are reassured. And mathematics education, I am afraid, is being quietly gutted.

Let me be precise about what is happening. These systems are not simply automating the drill. The more sophisticated among them watch every step a student takes, model their understanding in real time, and intervene with hints, prompts, or scaffolded sub-problems at the first sign of unproductive hesitation. The design philosophy is coherent and, on its own terms, humane: reduce cognitive overload, keep students in their zone of proximal development, and maintain engagement. What could go wrong?

Quite a lot, as it turns out. And the evidence has been accumulating for years, largely unread by the people who build and buy these tools.

The foundational problem is that the kind of cognition these systems are optimized to eliminate is precisely the kind that builds durable mathematical understanding. In 2021, Tanmay Sinha and Manu Kapur published a meta-analysis in the Review of Educational Research drawing on 53 studies across mathematics, physics, chemistry, biology, medicine, psychology, and environmental science. The core question was deceptively simple: Should students engage in problem-solving before receiving instruction or receive instruction first? The results were unambiguous. Students who struggled with problems before being taught significantly outperformed those who were taught first, demonstrating a meaningful effect on conceptual understanding and transfer without any loss of procedural fluency. When the instructional design adhered closely to the principles Sinha and Kapur had termed “productive failure,” the effect size tripled, reaching a margin roughly three times that of the annual gains attributable to an outstanding teacher. The research is not whispering. It is shouting, but the AI tutoring industry is not listening.

What the meta-analysis confirms runs counter to every instinct of a comfort-seeking education market: students learn more deeply when they are first allowed to fail. The failure is not a regrettable side effect of poor instruction. It is a feature of learning. Encountering a problem without a scaffold, sitting with the gap between what you know and what the task demands, trying approaches that collapse, and only then receiving instruction on the concept, activates cognitive processes that instruction alone cannot trigger. Students who struggle first develop richer knowledge structures, greater awareness of their own gaps, and a far stronger readiness to absorb and retain formal explanations. This is exactly the mechanism of how we learn. And the AI tutor, optimized to minimize exactly this discomfort, is dismantling it step by step.

Hiroko Warshauer arrived at a related conclusion from a different direction. In an observational study published in the Journal of Mathematics Teacher Education in 2015, she analyzed 186 episodes of struggle in middle school mathematics classrooms and documented what happened to 6th and 7th-grade students when teachers responded in different ways. The episodes in which students eventually achieved genuine understanding were not those in which teachers quickly reduced the difficulty of the task or provided direct guidance; they were those in which teachers pressed students to reason, built on their existing thinking, and gave them time. Warshauer named this the support of productive struggle, and her analysis showed how easily it is undermined by the instinct, however well-meaning, to rescue a student from difficulty.

Now scale that instinct. Build it into a system that responds in milliseconds and never tires. Deploy it across thousands of classrooms simultaneously. This is what we do when we learn.

This is precisely the terrain that the neuroscience of learning has been mapping with increasing urgency. Barbara Oakley, together with Michael Johnston, Ken-Zen Chen, Eulho Jung, and Terrence Sejnowski, argued in their 2025 chapter “The Memory Paradox: Why Our Brains Need Knowledge in an Age of AI” that constantly offloading cognitive work to external tools, rather than internalizing knowledge, results in shallow neural schemas that limit deep understanding and cross-domain thinking. The brain’s declarative and procedural memory systems are not passive storage devices; they are active structures that grow through use and atrophy through neglect. An AI tutor that perpetually intervenes before a student has time to engage those systems is not supporting learning. It is slowly dismantling the cognitive architecture that makes learning possible in the first place.

The companies selling these systems are not ignorant of the research. They have read it and found a way to accommodate it. Their response is essentially this: our system does not give away answers; it scaffolds students toward understanding with carefully calibrated hints. This is a more defensible position than simple answer-provision, but it misses the point. Research on productive failure does not merely argue against providing answers. It argues that the struggle itself, including the experience of being stuck, of trying an approach that fails, of feeling the gap between what you know and what the problem demands, is constitutive of learning. A scaffolded hint, which does not give the final answer but smaller bites of the final answer, prevents a student from sitting with that gap, which, in cognitive terms, is not much different from a provided answer. It short-circuits the very process it claims to support.

The stakes here are not abstract. Students who pass through mathematics courses guided at every step by AI systems that maintain their confidence and minimize their frustration may well achieve the platform’s performance metrics. However, what they may not achieve is the capacity to transfer their knowledge, to tolerate the ambiguity of an unfamiliar problem, or to persist when no hint is forthcoming. These are precisely the capacities that distinguish students who can use mathematics from students who can only perform it under supervision. And the irony is sharp: a technology marketed as preparation for a complex, uncertain, AI-saturated future may be producing learners who are spectacularly ill-equipped to navigate it.

There is a question that the companies building these tools, the administrators procuring them, and the governments funding the procurement do not want to ask: who benefits from a system that produces legible, measurable, short-term gains in student performance at the cost of the less visible, harder-to-measure, longer-term development of genuine mathematical thinking? The answer, I suspect, is not the students.

Before the next wave of AI tutoring adoption is locked into school contracts, it is worth demanding something the research already knows how to study: not whether students perform better on in-system assessments, but whether they understand mathematics better six months later, in a different context, without the algorithm’s assistance. Until that evidence exists, we are not deploying a pedagogical innovation. We are running an uncontrolled experiment, at scale, on other people’s children.

References

Oakley, B., Johnston, M., Chen, K.-Z., Jung, E., & Sejnowski, T. (2025). The Memory Paradox: Why Our Brains Need Knowledge in an Age of AI. In M. Rangeley & N. Fairfax (Eds.), The artificial intelligence revolution: Challenges and opportunities. Springer Nature.

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

Warshauer, H. K. (2015). Productive Struggle in Middle School Mathematics Classrooms. Journal of Mathematics Teacher Education, 18(4), 375–400. https://doi.org/10.1007/s10857-014-9286-3