An explanation can be perfect and still leave nothing behind. Most of us mistake the feeling of understanding for the fact of knowing. AI has made that mistake much easier to make.
Aarav is in Class 10 in Lucknow, and for three months he has had the best maths tutor he has ever had.
The tutor never gets tired. It never sighs when he asks the same question twice, never makes him feel slow in front of his friends, and it's there at 11:30 at night when the trigonometry homework finally makes him want to throw the book out of the window. He types a question, and a clear, patient, step-by-step explanation appears. If he doesn't follow a step, he types "explain like I'm 12" and it does. If he wants an example from cricket, he gets one from cricket.
His mother has noticed the change. The tantrums over homework have stopped. He finishes faster. When she asks how it's going, he says something she hasn't heard from him in years: "Maths is actually easy now. I understood everything."
Then the pre-board results come back. Trigonometry, the chapter he had "understood everything" in, is where he loses the most marks. He scores 47 out of 80. Not a disaster. Not the transformation either. Average, just as he was before the three months of brilliant explanations.
He isn't lying when he says he understood. He did understand, every night, sitting in front of the screen. What he couldn't do, alone in an exam hall with a question paper and no one to ask, was remember. He couldn't pull the method out of his own head, in the right order, without the explanation in front of him.
His mother asks the question every Indian parent eventually asks: "But you were studying every day. What happened?"
What happened is one of the most important and least discussed problems in education today. It isn't new. Teachers have seen it for a hundred years. But AI has quietly made it much worse, and much harder to see. Understanding and remembering are two different things. They feel the same from the inside, and our new tools are extraordinarily good at producing the first while leaving the second untouched.
(Aarav is a composite character. His evening, and his mother's question, will be familiar in a great many Indian homes.)
Let's be fair to the tools first, because they deserve it.
The explanations an AI model produces today are, very often, excellent. They're patient, personalised and available in many Indian languages at any hour. For a child in a small town without a good maths teacher, or whose parents can't afford tuition, that's not a small thing. In The Tuition Trap I wrote about the enormous sums Indian families spend on coaching, much of it driven by anxiety. A free, tireless explainer in every pocket could, in principle, take a great deal of that pressure off.
And Indian students have embraced it. At its AI for Learning Forum in New Delhi in January 2026, Google said India leads the world in daily use of its Gemini AI for learning. Indian users generated more than three million NotebookLM outputs (quizzes, flashcards, audio overviews) in a single month. Any parent with a teenager can see the rest for themselves: the chatbot tab open next to the homework, the photo of a textbook question uploaded for a step-by-step answer.
Now put one more number next to those. PARAKH Rashtriya Sarvekshan 2024, NCERT's national survey of more than 21 lakh students across 781 districts, found that the average Class 9 student scored 37% in mathematics. In Class 3 the figure was 60%. Learning falls away as children climb through school, and it falls away fastest in exactly the subjects where step-by-step methods have to be held in memory and used without help.
I want to be careful here. The PARAKH survey was conducted in December 2024 and it doesn't measure AI use. Nobody can draw a line from chatbots to that 37%, and I won't. The learning crisis in Indian schools is much older than ChatGPT, and its causes are many. I wrote about some of them in The Learning Poverty Number Nobody Talks About.
But the two numbers side by side raise a question that every parent, teacher and policymaker should be asking right now. We've given millions of Indian students the most fluent explainer in history. Will it close the gap between Class 3 and Class 9? Or will it give students like Aarav the feeling of closing it, while the gap stays where it is?
The research that has emerged over the past two years gives a surprisingly clear answer. It depends almost entirely on one thing: whether the AI does the remembering for the student, or makes the student do it.
In the autumn of 2023, a team of researchers from the University of Pennsylvania ran a field experiment at a large high school in Turkey. Nearly a thousand students in Classes 9, 10 and 11 took part in four 90-minute maths sessions, covering about 15% of their semester's syllabus. The results were published in PNAS in 2025, and they read like Aarav's story told in numbers.
Each session followed the same pattern. A teacher reviewed a topic and worked examples on the board. Then students practised problems, and here the classes were split three ways. One group had only their textbooks and notes. One had a standard ChatGPT-style interface running GPT-4, which the researchers called GPT Base. And one had a version the researchers had carefully prompted to behave like a good tutor, nudging and hinting instead of handing over answers. They called that one GPT Tutor.
During practice, both AI groups did far better than the textbook group. GPT Base students scored 48% higher on the practice problems. GPT Tutor students scored 127% higher.
Then the researchers took the AI away and gave everyone the same exam, with no help.
The GPT Tutor students performed about as well as the students who'd never had AI at all. The GPT Base students did worse, by roughly 17%, than students who had only ever used their textbooks.
Read that again, because it's the most important finding in this article. Students who had access to a powerful, unrestricted AI did better while they had it and worse once it was gone. When the researchers looked at the logs, they found a simple explanation. Many GPT Base students had used the AI as a crutch, asking for answers and copying them. The practice looked brilliant. The learning didn't happen.
The study has limits that its authors acknowledge. It was one school, one subject and four sessions. But it's a randomised experiment, and its message is hard to argue with: the same AI can help or harm learning, depending on whether it's designed to do the thinking for the student or with the student.
In June 2025, a team at the MIT Media Lab led by Nataliya Kosmyna published a preprint with a title that went viral: Your Brain on ChatGPT. They asked 54 university students in the Boston area to write short essays to SAT-style prompts over several months. One group used ChatGPT, one used a search engine, and one used nothing but their own brains. Electrodes recorded their brain activity while they wrote.
The brain-wave findings have been widely debated, and Kosmyna herself has warned against overinterpreting them. The finding I keep coming back to is much simpler, and it needs no electrodes.
After the first session, the researchers asked each participant to quote a sentence from the essay they had just written. In the brain-only group, 16 of 18 could. In the search-engine group, 15 of 18 could. In the ChatGPT group, not one of 18 could produce a correct quote from an essay they had submitted minutes earlier.
This was a small, early study of adults, not Indian schoolchildren, and it hadn't been peer-reviewed when it went viral. But the quote test captures something every teacher will recognise. If you didn't do the work of producing something, it was never really in your head to begin with. You can't remember what you never actually made.
If those two studies were the whole story, the lesson would be "keep AI away from students." They aren't.
In the autumn of 2023, Harvard physics lecturers Gregory Kestin and Kelly Miller ran a randomised trial with 194 students in Physical Sciences 2, the university's largest introductory physics course. For one topic, students attended an excellent active-learning class. For another, they worked through the material at home with an AI tutor the team had built themselves. The design followed the research on good teaching: it asked questions, gave one step at a time, made students attempt each problem and wouldn't simply hand over answers. Then the groups swapped.
Published in Scientific Reports in June 2025, the result was striking. Median learning gains with the AI tutor were more than double those in the classroom, and students took less time to get there.
Two caveats matter. These were high-achieving Harvard undergraduates, and the study measured learning right after the lesson, not weeks later. The authors themselves are careful to say an AI tutor shouldn't replace in-person teaching. But put the Turkey study next to the Harvard one and a clear pattern appears. When AI makes students work, they learn. When AI does the work for them, they feel like they've learned.
None of this is entirely about AI. It's about a quirk of the human mind that psychologists have studied for decades. AI just happens to exploit it at a scale no textbook or teacher ever could.
When information is easy to take in (clearly written, well explained, smoothly delivered), our brains read that ease as a sign that we know it. Psychologists call this processing fluency, and it's a powerful source of false confidence. Re-reading a highlighted chapter feels productive because the words are familiar. Familiarity feels like knowledge. It isn't.
The Harvard physics group documented this beautifully in an earlier study. In 2019, Louis Deslauriers, Kestin and colleagues published a randomised experiment in PNAS comparing a polished lecture by a highly rated instructor with an active class where students struggled with problems themselves. The active class learned more on the tests. But the students felt they had learned less, and they rated the polished lecture higher. The smoother the explanation, the more learning it seemed to produce, and the less it actually did.
Now imagine a lecturer who is infinitely patient, perfectly clear, and available at 11:30 every night. That's the AI explainer. It's the most fluent teacher ever built, and so it produces the most convincing illusion of learning ever built.
In 2002, Yale psychologists Leonid Rozenblit and Frank Keil asked people to rate how well they understood everyday things like a zip, a flush toilet or a bicycle. People rated themselves highly. Then the researchers asked them to write a step-by-step explanation of how each one worked. Confidence collapsed. People had mistaken recognising an explanation for being able to produce one.
That's exactly Aarav's situation. Every night, he recognised the explanation of trigonometric identities as it scrolled past. In the exam hall, he had to produce it. The two skills feel identical from the inside, and they're completely different.
The psychologist Robert Bjork has a phrase for the kind of struggle that builds memory: desirable difficulties. Trying to recall an answer before you look it up. Spacing out practice so you have partly forgotten. Mixing different kinds of problems so you have to work out which method applies. All of these feel harder and slower in the moment. All of them produce more durable learning.
In The Finite Feed I described the best-known example: a 2006 study by Henry Roediger and Jeffrey Karpicke in which students who practised recalling a passage remembered about 61% of it a week later, against about 40% for students who simply re-read it. The re-readers felt more confident. The recallers knew more.
This is the uncomfortable truth at the centre of this article. The effort we're so eager to remove from learning is the learning. An AI that removes the struggle from every step of homework isn't making learning easier. Very often it's making learning not happen, while producing a powerful feeling that it has.
Every country is wrestling with this. But several features of Indian schooling make the gap between understanding and remembering especially costly here.
Our exams are still closed-book, and the AI isn't allowed in. Board exams, entrance exams and most school tests ask a student to produce answers alone, from memory, under time pressure. Whatever else we think of that system (and I've been critical of it in pieces like I Am Because We Are), it's the system our children are judged by today. An AI that helps a child understand at home but leaves nothing in memory for the exam hall is, in the most practical sense, a trap.
Homework has become the easiest thing to outsource. Photo-solve apps were already common before generative AI. Now any question, in any subject, can be photographed and answered in seconds, with working shown. The homework gets done. The notebook looks perfect. The one thing homework exists to produce, practice at retrieving and applying a method alone, quietly disappears.
Teachers can't see the gap until it's too late. A teacher with 50 students in a class sees completed homework and confident faces. The first real signal that something is missing often arrives at the half-yearly or pre-board exam, months after the habits have formed. In Holistic Progress Cards I argued for richer, continuous evidence of what students actually know. The AI era makes that argument urgent.
Parents read the wrong signals. Faster homework, fewer tears and "I understood everything" all look like progress from the kitchen. They're exactly the signals the fluency illusion produces. Most parents have no simple way to tell the difference between a child who understood something and a child who can now do it alone.
The competency shift raises the stakes. NEP 2020 and the boards are moving, slowly but deliberately, toward questions that test application rather than recall of a textbook paragraph. That's the right direction, and I've argued for it in From Rote to Why. But applying a concept in an unfamiliar question requires the concept to be in your head, retrievable and flexible. You can't apply what you can't recall. Competency-based exams don't reduce the need for memory. They raise the bar for the kind of memory that counts.
I want to head off one misreading right away. Remembering isn't rote learning. Rote learning is reciting words you don't understand. What the research in this article describes is the opposite: understanding something and then practising pulling it out of your own head until you can use it without help. The child who can explain why the area of a circle is πr², in their own words, a week later, with the book closed, hasn't memorised a formula. They own an idea.
Regular readers will know my habit of naming ideas from our own languages: DARPAN, ANKUR, SETU, UTSAH and, most recently, VIRAAM. This one is SMARAN (स्मरण), which means remembrance or recollection. It's six habits that any student, parent, teacher or edtech product can use to make sure that understanding doesn't evaporate before the exam.
S — Struggle first. Before asking the AI (or the answer key, or the tuition teacher), spend five honest minutes trying the problem yourself. Write down what you know, where you're stuck and what you'd try next. Even a wrong attempt prepares the brain to absorb the explanation that follows. Researchers call this the pretesting effect: attempting something before you're taught it helps you learn it, even when the attempt fails.
M — Make it your own. After reading or hearing an explanation, close it and explain it back in your own words, on paper or out loud to a sibling or parent. If you can't, you haven't understood yet; you've only recognised. This is the antidote to the illusion of explanatory depth, and it takes two minutes.
A — Ask the AI to ask you. Change the direction of the conversation. Instead of "explain this to me," try "don't give me the answer, ask me questions until I get there," or "quiz me on this chapter, one question at a time, and don't tell me the answer until I've tried." This one sentence turns a GPT Base into something much closer to a GPT Tutor. Many tools now have a guided or study mode built in. Use it.
R — Return on a schedule. Come back to what you learned after one day, then after three, then after a week. Each return should be a recall attempt (questions, a blank page, a quick quiz), not a re-read. The forgetting curve is steepest in the first few days, so that's when a return does the most good.
A — Answer without the crutch. At least once a week, test yourself the way the exam will: no AI, no book, no notes, a timer running. This is the only honest measure of what you actually know. If the score is lower than you expected, that's not failure. It's the most useful information you'll get all week.
N — Notice the feeling. Learn to distrust the warm glow of "I get it." When an explanation feels easy, that's the moment to close it and check whether you can reproduce it. Feeling smart is information about how fluent the explanation was, not about what you'll remember on exam day.
None of these habits needs an expensive tool, a fast internet connection or a tuition fee. Most of them need a notebook, a pen and the discipline to close the screen for five minutes. That's the point. SMARAN isn't a product. It's a way of using every product, including the free ones, so that they leave something behind.
The tools aren't the villain in this story. The default way of using them is. Here's how the tools already in Indian students' hands can be turned from explainers into remembering machines.
Gemini and other chat assistants. Their default mode answers questions, which is the GPT Base pattern. But Google, like several other AI companies, now offers a guided learning mode that asks questions and builds understanding step by step. Students should know it exists, and parents should know to ask for it. Even without it, the single prompt "quiz me, don't tell me" changes everything.
NotebookLM. Because it works only from the sources you upload, it's well suited to exam preparation from your own textbook and notes. Its audio overviews are a lovely first pass, and they're still consumption. The flashcards and quizzes are where SMARAN happens. A good routine: listen once, then do the quiz the next day without listening again.
Google Classroom. With Gemini now built into Classroom at no extra cost for Workspace for Education schools, a teacher can generate a short quiz on last week's topic in minutes. That's the "return on a schedule" habit, delivered by the school instead of left to the student's willpower.
DIKSHA. India's national platform carries more than 3.6 lakh pieces of content and over 8,000 QR-linked textbooks, and much of it already includes practice questions and assessments, not just explanations. The opportunity, for any platform at that scale, is to put recall at the front of the journey: ask first, explain second.
Revision-first tools. A newer category of apps starts from memory rather than explanation. VidyaLoop, an AI study and revision app built for Indian school and college students, is one example: it brings back what a student has already studied in short recall loops, so the work of remembering happens before the exam rather than during it. Old-fashioned flashcards in a shoebox, sorted into "know it" and "not yet," use the same principle and cost nothing. What matters is the shape, not the brand: attempt first, recall on a schedule, and check yourself without help.
"Isn't this just an argument against AI?" No. The Harvard trial shows a well-designed AI tutor can more than double learning gains. The problem isn't AI in education. It's AI that removes the student's effort by default. My argument is for using AI the way the GPT Tutor group did, not for banning it.
"Students need understanding, not memorisation." Agreed, and nothing here says otherwise. But understanding that can't be retrieved when it's needed is fragile. The goal is understanding that lasts, and that requires practice at recall, which is the opposite of rote.
"Struggle is demotivating, especially for weak students." Struggle without support is demotivating. That's why the habit is five minutes of struggle followed by help, not an hour of frustration. The pretesting research suggests even a failed attempt makes the explanation that follows more effective. For a student who's already behind, that's good news: their attempts count, even the wrong ones.
"These studies are small and mostly foreign." True, and I've said so throughout. The Turkey study is one school; the MIT study is 54 adults; the Harvard study is elite undergraduates. India urgently needs its own research on how AI is changing learning in our classrooms, in our languages, under our exam system. Until it arrives, the direction of the existing evidence, backed by decades of memory research, is the best guide we have.
Aarav didn't give up his AI tutor. He changed how he used it.
His maths teacher had read about the Turkey study and told the class about it one morning: the students who'd scored brilliantly with the AI and worse without it. Something about the 17% stuck with him. That night, before asking for an explanation, he tried the problem himself for five minutes. He got stuck halfway. Then he typed something new: "Don't solve this. Ask me questions until I get there."
It was slower. It was irritating. On some nights it made him feel stupid in a way the old explanations never had. He started a small notebook of problems he'd got wrong and redid three of them, from scratch, every Sunday. Twice a week his mother would point at one and say, "Show me, without the phone."
He didn't feel like he was getting better. He felt like he was getting worse, because for the first time in months he could see what he didn't know.
In the board exam, trigonometry was the section he finished first.
We've built the most fluent teacher in history and handed it to a generation of Indian children. That's a genuine gift. But fluency was never the goal. The goal is a child who, alone in an exam hall, in a job interview or in a hard moment of their life, can reach into their own mind and find what they need.
Understanding opens the door. स्मरण is what keeps it open.
If this piece was useful, share it with one parent or teacher who's watching a child "understand everything" and score average, and subscribe to Chalk & Code for the next one in the series.
Understanding Is Not Remembering: Why AI Makes Students Feel Smart and Score Average