From an AI Answer to True Understanding

During our school’s annual School Without Walls trip to Tibet, we visited the Sera Monastery in Lhasa with a group of students. At the monastery, we witnessed the debates that take place in the courtyard. Standing monks challenged seated ones. A standing monk asked a question and clapped to hand it over; the seated monk answered, and others watched and weighed in. The debates draw on five great treatises: valid cognition (logic and how we know what we know), the perfection of wisdom, the middle way, phenomenology (the analysis of mind and matter), and monastic discipline. 

The seated monks had to defend and support their answers. They could do that because of years of memorizing and synthesizing texts under a teacher (Dreyfus, 2003). Given my work in STEM subjects and my focus on the future of education and cognition, what caught my attention was this long-held practice of challenging and questioning opinions. This article is not about a religious perspective, but about how learning is enhanced and why these cognitive practices may still be relevant in the artificial intelligence (AI) era. 

Students can now get an answer from AI in seconds. That does not automatically give them understanding. The Tibetan courtyard is not a model for a STEM classroom, but it elicits two questions I want to explore for schools: How do students move from having an answer to developing deep understanding? And how can students learn to guard their own thinking when AI can answer for them?

More Than Completing a Process

Let’s deal with the first question: How do students move from having an answer to developing deep understanding? Take a perfectly completed project; it may hide shallow student understanding. My concern is that, for some students, completing a process in aesthetically pleasing form becomes proof of learning. Similarly, teachers asking students to “evaluate” does not teach evaluation. Calling a solution innovative does not establish evidence of improvement. Understanding shows when students can explain and defend their choices under questioning, as the seated monks had to.

The AI era makes this gap harder to see. For example, in a randomized study at one Turkish school, students using a basic GPT-4 interface performed better during mathematics practice but worse on unassisted assessments than students without AI. An AI tutor that gave hints instead of answers mostly prevented this drop, but it did not raise test scores either (Bastani et al., 2025). The test was taken the same day, so we don't know if the effect lasts. Still, doing well with AI's help is not the same as learning.

How Understanding Develops

So, what should we teachers emphasize? Teach the knowledge students need, then ask how its parts relate to each other. In mathematics, discuss why a formula works, not just how to use it.

Cognitive science research supports spaced practice and retrieval practice as ways to help students remember what they learn. Spaced practice means returning to a topic across separate sessions. Retrieval practice means answering from memory before consulting notes or rereading an answer (Carpenter et al., 2022). Some examples show what these approaches can achieve. Students learning vocabulary in a foreign language remembered about 80% of the words a week later when they kept being quizzed on words they had already gotten right. When they only kept rereading those words with the answer shown, they remembered about 35% (Karpicke & Roediger, 2008). In seventh-grade science and high school history, rudimentary quizzes with corrective feedback improved performance on later assessments. A comparison with retrieval practice also showed that the benefit was not simply from seeing the information again (McDermott et al., 2014). In mathematics, spreading practice across sessions improves learning compared with doing all the practice at once (Murray et al., 2025).

However, remembering is not the whole of understanding. With appropriate guidance, explanations, and visual representations, students can connect new information with prior knowledge (Fiorella, 2023). Simply asking students to explain does not ensure that they can.

For teachers, combining these practices purposefully with instruction and feedback can improve learning. Work through an example, ask students to justify selected steps, then have them attempt a related problem. Use their responses to identify what needs further teaching, and allow them to correct their reasoning. Return to the concept over several lessons, asking students to recall what they know and explain how it applies. The aim is not simply to include a learning strategy, but to examine whether a strategy is helping students learn. Check what students can remember, explain, and use later, not just whether they completed the work.

Guardians of Their Own Thinking

Now to the second question: How can students learn to guard their own thinking when AI can answer for them? Whenever AI is permitted, if at all, in the classroom, students need to learn when assistance supports their understanding and when it replaces reasoning they still need to practice. In the classroom, ask students to attempt an explanation after instruction, use assistance for a specific difficulty, then check the result independently. Model that check through a calculation, a credible source, or a counterexample. Later, ask students to explain again without AI supplying the reasoning. Retain appropriate language and access supports. 

Have students compare what they expected to explain with what they actually could, then identify what needs further revision. Check confidence against evidence of understanding; students are often poor judges of what they will remember (Karpicke & Roediger, 2008). Being guardians of their own thinking also implies noticing the limits of their knowledge. Students need to ask not only whether they understand an answer, but whether they know enough about the subject to judge it. Knowledge does not transfer automatically. Critical thinking depends on subject-related knowledge and practice (Willingham, 2007). A student who can find the derivative of a function may need instruction to interpret velocity in physics: what is changing, with respect to what, and what do the sign and units mean?

When transfer is an objective, teach the connections explicitly. Ask what principle applies and what additional knowledge is needed. An interdisciplinary project does not, by itself, demonstrate transfer.

The enduring lesson of Sera Monastery is that, despite technological advancement, the fundamental needs of the human brain remain the same.To develop deep understanding, as in Tibet’s ancient debate tradition, the brain still needs spaced practice and retrieval to retain information, and it needs human relationships: a teacher to guide and peers to challenge. Being guardians of our thinking means knowing what we can explain and justify, where our knowledge is incomplete, when we need help from a guide, and when we need to challenge our perspective in light of others’ points of view. Teaching and learning are a social, interactive endeavor, and human interaction matters more now than ever in the era of AI.

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122

Carpenter, S. K., Pan, S. C., & Butler, A. C. (2022). The science of effective learning with spacing and retrieval practice. Nature Reviews Psychology, 1, 496–511. https://doi.org/10.1038/s44159-022-00089-1

Dreyfus, G. B. J. (2003). The sound of two hands clapping: The education of a Tibetan Buddhist monk. University of California Press.

Fiorella, L. (2023). Making sense of generative learning. Educational Psychology Review, 35(2), Article 50. https://doi.org/10.1007/s10648-023-09769-7

Karpicke, J. D., & Roediger, H. L., III. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968. https://doi.org/10.1126/science.1152408

McDermott, K. B., Agarwal, P. K., D’Antonio, L., Roediger, H. L., III, & McDaniel, M. A. (2014). Both multiple-choice and short-answer quizzes enhance later exam performance in middle and high school classes. Journal of Experimental Psychology: Applied, 20(1), 3–21. https://doi.org/10.1037/xap0000004

Murray, E., Horner, A. J., & Göbel, S. M. (2025). A meta-analytic review of the effectiveness of spacing and retrieval practice for mathematics learning. Educational Psychology Review, 37, Article 75. https://doi.org/10.1007/s10648-025-10035-1

Willingham, D. T. (2007). Critical thinking: Why is it so hard to teach? American Educator, 31(2), 8–19.