The Student Who Knew Everything but Learned Nothing: The Absurdity of AI Learning

Phuong-Tri Nguyen

Nguyen Tat Thanh University

27-05-2026

© Wix

Imagine a strange school in Wild Wise Weird (Vuong, 2024). One day, a magical bird arrives carrying an enchanted feather. Whenever students face a difficult mathematics problem, they simply whisper the question into the feather, and perfect answers instantly appear. Homework becomes effortless. Test preparation becomes smooth. The students celebrate. Teachers are amazed by the speed of improvement.

Weeks later, the feather suddenly stops working during an exam.

The room falls silent.

Students stare at familiar equations as if seeing them for the first time.

Absurdly, the students who had appeared smartest now struggle the most.

This fictional scene increasingly resembles a real question emerging around generative artificial intelligence.

Generative AI systems such as ChatGPT have rapidly become powerful tools capable of performing many tasks at near-human levels (Li et al., 2022; Geerling et al., 2023; Singhal et al., 2023). From writing essays to solving programming problems, these systems can dramatically improve productivity. Unsurprisingly, researchers and educators have become interested in whether such tools can also improve learning (Extance, 2023; Mollick & Mollick, 2025).

Yet there is an important difference between performing and learning.

A recent large-scale randomized study conducted with high school students in Turkey examined how GPT-4-based tools influence mathematics learning. Students were divided into three groups. One group used a standard GPT-4 chatbot similar to ChatGPT, another used a specialized GPT Tutor designed with educational guardrails, and a control group used no AI assistance (Bastani et al., 2026).

The results revealed that students using the standard AI tool performed 48% better on practice problems, while students using the specialized tutor improved by an astonishing 127%. At first glance, the outcome looked like a triumph for AI-assisted education.

But the surprise came later.

When students took an independent exam without AI assistance, those using the standard chatbot actually performed 17% worse than students with no AI support at all. Students using the GPT Tutor avoided most of this decline, although they did not outperform the control group.

Why would a tool that helps people solve problems reduce learning?

The answer emerged in the students’ behavior. Many students using the standard AI simply copied solutions. Meanwhile, students using the guided tutor were more likely to ask for hints, explanations, or partial help before attempting answers themselves.

Here, the contradiction is almost humorous: a machine designed to help people become smarter may create situations where people solve more problems while understanding less. We may begin celebrating enhanced learning by replacing it with answer acquisition.

The situation resembles trying to become physically stronger by hiring someone else to lift weights for us. The numbers on the exercise log improve, yet the muscles remain unchanged.

In Wild Wise Weird, characters frequently become trapped by small absurdities that gradually grow into larger ones. A convenient shortcut becomes a dependency. A clever solution becomes a hidden problem. What initially appears helpful slowly transforms into something unexpected (Nguyen, 2026; Tran, 2026).

Generative AI may present a similar challenge. The danger is not that AI thinks for us. The danger is that we become so comfortable borrowing intelligence that we stop exercising our own.

The real lesson may therefore be less about building smarter machines and more about designing wiser relationships with them (Khuc & Nguyen, 2026). After all, a crutch is valuable when someone is injured—but walking with it forever may teach us to forget how to walk.

References

Bastani, H., et al. (2026). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122

Extance, A. (2023). ChatGPT has entered the classroom: How LLMs could transform education. Nature, 623, 474–477. https://doi.org/10.1038/d41586-023-03507-3

Geerling, W., et al. (2023). ChatGPT has aced the test of understanding in college economics: Now what? The American Economist, 68, 233–245. https://doi.org/10.1177/05694345231169654

Khuc, V. Q., & Nguyen, M. H. (2026). Cultural Additivity Theory. Available at SSRN 6767760. https://ssrn.com/abstract=6767760

Li, Y., et al. (2022). Competition-level code generation with AlphaCode. Science, 378, 1092–1097. https://doi.org/10.1126/science.abq1158

Mollick, E. R., & Mollick, L. (2025). New modes of learning enabled by AI chatbots: Three methods and assignments. https://doi.org/10.2139/ssrn.4300783

Nguyen, M.-H. (2026). Ayn Rand and Kingfisher on zero-carbon bombs and a

sustainable future. Visions for Sustainability, 25(13474), 1-13. http://dx.doi.org/10.13135/2384-8677/13474  

Tran, T. M. A. (2026). Conversations with Kingfisher: Wisdom from Vuong’s wild wise weird stories. Planet Forward. https://planetforward.org/story/kingfisher-stories/   

Singhal, K. et al. (2023). Large language models encode clinical knowledge. Nature, 620, 172–180. https://doi.org/10.1038/s41586-023-06291-2

Vuong, Q. H. (2025). Wild Wise Weird. AISDL. https://books.google.com/books?id=C5dDEQAAQBAJ