Education in the age of AI

I want to start with a confession that most people who grew up in the Indian education system will recognize: I spent the most formative years of my intellectual life optimizing for the wrong thing. Not learning. Recall. The two get conflated so thoroughly in our schools that by the time you realize they aren't the same, you've already spent a decade training the wrong muscle. You can recite the laws of thermodynamics and have no intuition for why a hot cup of coffee cools. You can score a 95 in chemistry and be unable to explain, in your own words, what an orbital actually is. The system rewards the performance of knowing, not knowing.

I think about this a lot now, partly because I work on AI systems for a living, and partly because I have a two-year-old. By the time she's old enough to sit for an exam, the world she sits in will be unrecognizable. And I don't mean that in the breathless, conference-keynote way. I mean it in the specific, mechanical way that anyone watching the frontier closely sees — the way that a sufficiently capable AI compresses decades of specialist labor into months, and turns the question "what should a human spend their finite years learning?" into a question we genuinely have to re-answer from scratch.

So let me try.

What education is actually for

There's a tidy answer and a real answer.

The tidy answer is that education exists to transmit knowledge across generations, prepare people for economic participation, and cultivate citizenship. This is the answer policymakers give. It is not wrong, exactly, but it's the kind of true thing that obscures more than it reveals.

The real answer, I think, is that education is supposed to do four distinct things, and we've been pretending they're one thing for so long that we've forgotten they can come apart:

Transfer information — facts, formulas, dates, vocabulary.

Build skill — the ability to actually do something with that information: write a proof, fix an engine, diagnose a patient, debug a system.

Shape judgment — taste, ethics, the ability to tell a good argument from a flashy one, the ability to know what's worth working on.

Develop the person — curiosity, resilience, the capacity for sustained attention, the ability to disagree without disintegrating.

The first one — information transfer — is the part the Indian system, and many systems like it, did with brutal efficiency and almost no joy. Mug it up. Vomit it out on the answer sheet. Move on. We did this not because anyone thought it was the best way to learn, but because at scale, with one teacher per sixty students and a single high-stakes exam at the end, it was the only mechanism that could be enforced. Rote learning is a coordination technology, not a pedagogical one. It exists because evaluating actual understanding at scale was, until very recently, impossible.

That constraint has now broken.

The constraint that just broke

Here is the thing that I don't think most people in education have fully internalized yet: the reason your child has been graded on what they remember rather than what they understand is that we did not have the technology to grade understanding cheaply. A good Socratic dialogue, the kind where a tutor probes, asks "but why?", reframes the question, watches the student stumble and recover — that is the gold standard of pedagogy and has been since Athens. We have known this for two and a half thousand years. We did not adopt it because it requires roughly one trained adult per student, and humanity could not afford that.

We can afford it now.

A model that can hold a coherent, patient, infinitely-tolerant conversation with a child about why the coffee cools — that can ask the right next question, notice when the child is parroting versus reasoning, and adjust — exists today. Not in some speculative future. Today. The version five years from now will be substantially better than any tutor most of us have ever had access to, and it will run on a phone.

Bloom's famous "two sigma problem" — the finding that one-on-one tutoring produces results two standard deviations above classroom instruction — was, for forty years, an interesting curiosity that policy could do nothing about. It is now a deployment problem. And deployment problems get solved.

What this changes, concretely

Let me try to be specific, because the abstract version of "AI will transform education" has been said so many times it's lost meaning.

The mug-up era ends, but not the way people think. The reflex among smart people is to say, "Good, no more rote learning, kids can finally focus on creativity and critical thinking." This is half right and half a fantasy. You cannot think critically about a domain you don't know anything about. A child who has never memorized the multiplication table will not have number sense; they will have a calculator dependency. The end of mug-up doesn't mean the end of memorization. It means memorization gets put in its proper place — as scaffolding for understanding, not as a substitute for it. The AI tutor's job is to make sure the kid actually internalizes the times table and knows why 7×8 is what it is, instead of letting them skate by on either one.

Subjects stop being the unit of learning. The fact that we organize twelve years of a child's life around "physics," "chemistry," "biology," "history," "geography" is a nineteenth-century artifact of how textbooks were written and how teachers were trained. Real curiosity doesn't respect these boundaries. A kid who wants to know why the sky is blue is asking a question that touches optics, atmospheric chemistry, evolutionary biology (why our eyes see those wavelengths), and the history of how anyone figured it out. An AI that can follow that thread wherever it goes, and that can pull the child back when they're going in circles, doesn't need to pretend the world is divided into forty-five-minute periods.

Assessment becomes continuous and invisible. The exam, as an institution, exists because we needed a discrete moment to measure something that was otherwise unmeasurable. If a system is in dialogue with a learner every day, it knows what they understand. The terror of the final exam — the thing that defined my own adolescence and the adolescence of every Indian kid I grew up with — becomes obsolete in roughly the same way that the village storyteller became obsolete when the printing press arrived. There will still be moments of evaluation, but they'll be more like a doctor's checkup than a public execution.

The teacher's job changes, but doesn't disappear. I want to be careful here, because I don't believe the breathless version of this where every teacher gets replaced by a chatbot. The teacher's job, in a world with AI tutors, becomes something closer to what the best teachers were always trying to do anyway — mentor, role model, designer of experiences, person who notices when a kid is going through something at home. The information-transfer part of the job, which was always the most soul-crushing part, gets handed off. What's left is the human part. This is, I think, a better job. It is also a job that we will need fewer people for in a narrowly transactional sense, and more people for in any sense that takes children seriously as humans. How we resolve that tension is a political question, not a technical one.

What's worth learning

Here's the harder question, the one I keep circling. If a model can do, in thirty seconds, the kind of analysis that took me four years of engineering school, what exactly am I supposed to teach my daughter?

I don't fully know. But I have some guesses, and I'll put them down honestly.

Taste. This is the unfashionable answer but I think it's the most important one. The ability to look at two pieces of work and know which one is better. The ability to look at a problem and know which version of it is worth solving. AI is going to flood the world with competent output. The scarce thing — the thing that's always been scarce, but is about to become the only scarce thing — is judgment about what's worth making. Taste isn't taught directly. It's absorbed, by exposure to good things and by being around people who can articulate why those things are good. We should be far more deliberate about this than we are.

The ability to ask. Most of school, in India and elsewhere, trains kids to answer. AI inverts the value: anyone can get an answer now, but very few people can pose a question that's worth asking. Question-asking is a skill. It can be taught. It mostly isn't.

Real making. Not "projects" in the school sense, where the rubric is so well-defined that the project is just an exam in disguise. Actual making — building a thing that has to work, where the world tells you whether you got it right, not the teacher. This was always the most educational thing a child could do. It is now also the thing that most clearly marks the boundary between a person and the AI they're using, because making something is fundamentally about caring whether it's good, and caring isn't yet something we know how to outsource.

Friendship with hard things. I mean this almost literally. The thing school did to me, and to most of my friends, is teach us that hard things were obstacles between us and a grade. Not a single teacher, in twelve years, communicated to me that struggling with a problem for a long time was the point, not a failure mode of an otherwise efficient process. Mathematics was something to get through. So was poetry, and so was Sanskrit. With AI as a patient companion that doesn't get bored or frustrated, you can finally let a child sit with a hard thing for as long as it takes, and you can teach them to like it.

Bodies and other people. I notice that almost everything I've written so far is about thinking. The longer I work in this field, the more convinced I become that the parts of being a person that aren't about thinking are the parts that matter most, and the parts that AI is least going to help with. Sport, music, cooking, dancing, looking after another human being, being looked after — these are not "soft skills" to be relegated after the real curriculum is done. They might be the curriculum. The real one.

What I worry about

I don't want to write the version of this essay where everything turns out fine if we just adopt the technology thoughtfully. The honest version doesn't write that essay, and I don't want to either.

The thing I worry about is not that AI tutors will be bad. The early ones already aren't, and the trajectory is steep. The thing I worry about is that we'll get the deployment wrong in ways that look exactly like the deployment of every other powerful tool we've built. The kids who already have engaged parents and good schools will get personal tutors that turbocharge them. The kids who already have nothing will get a lower-cost version of the worst parts of the existing system, automated. The gap between these two outcomes is enormous, and bridging it is not a technical problem. It's the same political problem we have always had, dressed in new clothes.

“Caring isn't yet something we know how to outsource.”

I also worry — and I say this as someone who builds these systems — about a generation of children who learn to outsource the parts of thinking that build a self. There is something that happens to a person who has to genuinely struggle with a hard idea, alone, until it gives. Something gets formed. I don't yet know what it does to a child to never have to do this, because the patient AI is always there to break the problem into smaller pieces. It might be fine. It might not. We are running this experiment on real children right now, and we don't get to do a second take.

The version I want for my daughter

I'll close with this, because abstractions are easy and parenting is not.

I want my daughter to grow up in a world where her curiosity is taken seriously. Where the question she asks at dinner — about why the moon follows the car, about whether ants get bored, about why her amma's accent is different from her appa's — is answered by something patient enough to follow it wherever it goes. I want her to have access, by default, to the kind of one-on-one attention that, in the world I grew up in, was reserved for the children of the rich.

I also want her to know how to be bored. I want her to have a relationship with a hard thing — a musical instrument, a sport, a craft, a body of literature — that has nothing to do with being optimal at it. I want her to have friends she made by being physically near them for long enough that something happened. I want her to have the experience of caring for another person who needs her, before any AI tells her caring is a thing she can do.

I think the version of education that's coming can give us both. The patient infinite tutor, and the room to be a child. The end of the cruelty of mug-up, and the preservation of the dignity of struggle. It won't happen by itself. It's the kind of thing that happens because we decide, as parents and teachers and societies, what we actually value, and then we build for it.

The technology, finally, isn't the constraint anymore.

We are.

— Aravind