Beyond bans and permissions: Designing AI that teaches students to think

Not long ago, watching a group of high school students work through a thorny calculus problem, I noticed something that has stayed with me. When the problem got hard enough, no one reached for a textbook or sketched a graph. The instinct was a reflexive reach for the phone. Within seconds, the answer appeared. The tension in the room dissolved, but so did the learning. AI as a cognitive calculator has made many human efforts more productive, but one of those efforts demands the very friction that AI ends up removing – learning something new.

We are living through the most consequential pedagogical experiment in human history. As someone who has spent decades building technology at scale and shaping digital platforms used by millions, I am instinctively an optimist about what technology can do for access and equity. But what I am watching in classrooms troubles me. We are building a generation that knows how to find answers but has quietly forgotten how to think.

The scale of AI adoption in education is extraordinary. Google’s 2025 Our Life with AI survey across 22 countries found that 85% of students and 81% of teachers are now regular AI users, well ahead of the general public. Learners and educators, the report concluded, are AI’s new super users.

Cognitive debt

But look closer and a different story emerges. AI usage among students drops sharply during summer vacations; the one period when homework dries up. Students are not using AI out of curiosity. They are using it to complete tasks. The moment the task disappears, so does the AI. That is not a learning revolution. That is a more efficient homework machine.

Neuroscience is beginning to confirm what teachers have suspected. A 2025 preprint study from MIT Media Lab, “Your Brain on ChatGPT”, tracked participants over four months using EEG-based brain activity monitoring. Students who used AI assistance showed measurably lower neural connectivity than those who worked without it, particularly in regions tied to memory and reasoning. When asked to recall sentences they had written minutes earlier, 83% of AI users could not, because, in any meaningful sense, they hadn’t written them at all.

The researchers coined the term “cognitive debt”, the gradual erosion of the mental effort we are willing to make when AI is reliably present to spare us from it. Like financial debt, it accumulates quietly. For a student preparing for the JEE or NEET, cognitive debt is not abstract. It is the gap between what the algorithm has done on their behalf and what their brain can independently produce when the examination booklet opens and the screen goes dark. They may have mastered the prompt. They have not mastered the concept.

For India, this matters more than most places. The country produces the world’s largest cohort of STEM graduates, and its students sit some of the most demanding examinations on earth. We have also spent the last decade — through the NEP 2020 and broader reform — trying to move education away from rote memorisation toward conceptual, application-based thinking. A no-rules AI environment risks replacing one form of rote with another: not repetition of textbook lines, but prompt-and-paste. High-tech, and just as hollow.

The demographic dividend we speak of so often is not guaranteed. It is conditional on whether the next generation can reason, build, and solve; not merely retrieve.

Bypassing productive struggle

In every other industry, we optimise to eliminate friction. In education, friction is the point. There is a concept in pedagogy called productive struggle — that zone where a student is challenged enough to stay engaged, but not so overwhelmed that they quit. It is there that new neural pathways are built. When AI delivers the answer in three seconds, it bypasses the zone entirely.

The right design principle for AI in education is not “give the correct answer faster.” It is “help the student arrive at the correct answer themselves.” These are architecturally different products. One needs to be smart. The other needs to be pedagogically wise knowing when to prompt, when to withhold, when to ask a counter-question, and when to let the student sit in difficulty long enough for something to click.

Early evidence suggests this works. In pilots across student cohorts in India and the United States, students who engaged in guided, stepwise practice, without being handed solutions showed average mastery gains of nearly 50% from their first attempts to their last, with reliance on hints falling steadily over time. They were learning to think without the crutch.

Beyond the binary of ban vs allow

The conversation about AI in education cannot be a binary of ban versus allow. The only productive question is: what kind of AI, and to what end?

AI tools in schools should be audited for their pedagogical logic – does the tool show the answer, or does it require the student to show their work? AI should operate on a coaching-first mandate: surfacing misconceptions, not just errors. And it should serve the teacher, giving them an X-ray vision into where student reasoning is breaking down, so human judgment can step in where it is most needed.

We can use AI to automate the student, or use it to augment the student’s intellect. The choice, once stated plainly, is not a hard one. But it requires intention, and it requires rules.

The struggle is the lesson. We cannot afford to engineer it away.

(Peeyush Ranjan is the CEO of Fermi.ai, an AI-first learning platform for high-school STEM education. He has previously served as GM & VP at Google, VP of Engineering at Airbnb, and CTO of Flipkart.)

(Sign up for THEdge, The Hindu’s weekly education newsletter.)

Published – February 11, 2026 02:38 pm IST