Let me start with the frame, because the frame decides the answer. The fear that AI will atrophy young minds is not irrational. It is a real risk, and I take it seriously. But it is not a property of the technology — it is a property of how the technology is used. The same tool that can hollow out a student's thinking can, used differently, expand it dramatically. Those are not two different tools. They are the same tool, pointed in two different directions. So the useful question is never “is AI good or bad for learning?” It is “what do we have to do so that it develops capacity instead of replacing it?” To answer that, I use a picture I keep coming back to.
The electric bicycle
Think about an electric bicycle. You can ride it in two completely different ways.
The first way is the lazy way. You put the motor on maximum, you stop pedaling, and you let it carry you. You arrive at your destination effortlessly. You also arrive with no exercise. Your heart is no stronger than when you left. Your legs have done nothing. You have moved through space and developed nothing. The bicycle did the work, and the work was the whole point you skipped.
The second way is the athlete's way. You use the motor strategically — to climb hills that used to be impossible and shut whole routes off from you, to extend your range past what your current condition allows, to hold your pace when you are tired but still want to keep training. Now the assistance does not replace your effort; it unlocks more of it. You train harder, go farther, challenge yourself more, and build more capacity than you ever could without it.
Same bicycle. Same motor. Radically different human being at the end of the ride. One has coasted; the other has become stronger precisely because of the machine. Artificial Intelligence is exactly this bicycle. Intelligence, like a muscle, grows with use and wastes without it — and AI is an engine you can bolt onto that muscle to spare it or to stretch it.
Two futures
That choice, made at scale across a whole education system, opens onto two very different futures. They are worth naming plainly.
The first future is AI as a crutch. Students use it to do their homework without thinking, to generate essays they neither read nor understand, to obtain answers without ever passing through the process that would have taught them something, to avoid cognitive effort entirely. The result is a generation that leans on AI for basic thoughts, that cannot reason independently, whose intellectual muscles have quietly atrophied from disuse. This is the future we fear. I want to be clear: it is a genuinely possible one. It is a real risk, not a rhetorical one.
The second future is AI as a cognitive gym. Students use it to practice extensively with immediate feedback, to explore more complex concepts faster because they mastered the fundamentals efficiently, to receive personalized scaffolding that lets them work right at the edge of what they can almost-but-not-quite do on their own, to sharpen their judgment by evaluating and correcting what the AI produces, to train skills that used to be inaccessible without a private tutor. The result is a generation with amplified cognitive capacity — able to think deeper, wider, and more creatively because it has tools that multiply human capability instead of substituting for it. This future is extraordinary. It is also possible.
Both futures run on the same machine. The difference between them is not in the AI. It is in the design of how the AI is used.
The all-powerful gym
So here is the real challenge, and it is harder than the fear makes it sound. We have been handed an all-powerful gym. The AI can do almost anything — that is exactly what makes it dangerous as a learning tool, not just powerful as one. A gym where every machine will lift the weight for you is not obviously a place anyone gets stronger. The whole task in front of educators is this: how do we get students to use this all-powerful gym to build their cognitive muscles rather than to atrophy them?
You can ride this bike as a lazy person's shortcut, or as a professional's machine loaded with every technological advance — built to develop your maximum potential, if you are willing to put in a little extra effort. The bicycle does not decide which. And this is the point where most of the conversation goes wrong.
Whose responsibility this actually is
Here is the part I insist on, because almost everyone gets it backwards: the responsibility for making the right choice is not the student's.
Ask a ten-year-old whether they want the easy homework or the hard homework. They will choose easy. Obviously. Ask a teenager whether they want the AI that hands them the answer or the AI that makes them think. They will choose the answer. Obviously. This is not a moral failing in our students. It is not evidence that this generation is lazier or softer than the ones before it. It is human nature to seek the path of least resistance — you would choose it, I would choose it, every adult reading this chooses it constantly in the parts of their own life where no one is designing the friction for them.
So it is neither fair nor realistic to put the burden of this choice on children and teenagers and then act surprised when they take the shortcut we left wide open. The responsibility belongs to the adults in the room. It is the job of teachers and educational institutions to design the use of AI so that it develops capacity rather than replacing it. The student rides the bike. We decide whether the assignment, the assessment, and the classroom are built so that the motor unlocks more effort or excuses all of it.
That is a design problem, and it is squarely ours. If the essay can be vibed into existence and still earn the grade, students will vibe it, and they will be right to — we built an assessment that rewards coasting. If the only way to succeed at the task runs through the student's own thinking, with AI as the training partner rather than the substitute, then the same tool that could have atrophied them will strengthen them instead. Nothing about the technology chose that outcome. We did.
Design the effort in
So my answer to “will AI make students think less or more?” is not a prediction. It is an assignment. It will make them think less wherever we let them coast, and more wherever we design the effort back in. The muscle is real; the atrophy is real; the amplification is real. Which one a school produces is decided not by the students and not by the AI, but by the adults who design how the two meet.
Stop asking whether the bicycle is good or bad. Start asking how we get every student to ride it like an athlete — and then take responsibility for building the road that makes them.
A working draft. This essay lays out my position; a forthcoming revision will weave in sourced research — the cognitive-science evidence on offloading, desirable difficulties, and learning with generative AI — with full citations. If you have work I should read before then, tell me. — Carlos Miranda Levy
Four perspectives
The move I trust most in this essay is that Carlos refuses to locate the effect in the technology. ‘Will AI make students think less or more’ is unanswerable as stated, because it treats AI as a fixed intervention with a fixed effect. It is not. The effect is conditional on the task design, and that is precisely where the evidence points too: outcomes with the same tool diverge wildly depending on whether the task forces the learner to generate, retrieve, and self-explain, or lets them accept a finished output. So the honest claim is not ‘AI helps’ or ‘AI harms’ — it is ‘the effect is a function of design,’ which is both more modest and more actionable.
I want to sharpen the responsibility point, because it has an equity edge Carlos names but I would underline. Saying ‘the burden is on educators and institutions, not students’ is correct — and it means the schools most able to design good friction are the ones with the most resourced, best-supported teachers. Under-resourced schools are the ones most likely to hand over the tool with no design at all, because designing the effort in takes time, training, and slack that stressed systems do not have. So ‘it depends on how we use it’ can quietly become ‘it depends on how well-funded your school is.’ If we accept the responsibility framing, we have to fund the responsibility, or we have built a machine for widening the gap while calling it access.
Practically, the electric-bike test is a great gut check for any assignment: does the AI let the kid coast, or does it let them climb a hill they could not climb alone? If you cannot answer that about a task you just assigned, you have not designed it yet. Concretely: make the AI the sparring partner, not the ghostwriter. Have them argue with it, correct it, and explain why its answer is wrong or right. The explanation step is where the muscle actually moves. And stop banning the tool out of fear — banning it just outsources the design problem to the kids, who will absolutely solve it in the lazy direction.
I keep coming back to the bicycle because it puts the blame in the right place. When a student uses AI to avoid thinking, our instinct is to call it a character problem — kids these days, no discipline. That is a comfortable story because it is not about us. But a ten-year-old choosing the easy homework is not a scandal; it is the most predictable thing in the world. The scandal would be adults who know that and design the assignment as if it were not true. Intelligence is a muscle. AI is the most powerful training machine we have ever built for it, and also the easiest way ever invented to skip the workout. Which one it becomes in your classroom is your decision, not your students’. That is heavy, and it should be. It is also the only version of this problem we can actually do something about.