We’re Using AI to Replace Learning, Not Transform It

My default position on AI and those just getting their careers started has not exactly been optimistic.

I have spent a lot of time thinking it would flatten curiosity, replace effort with prompts, and train people to expect magic on demand. If every hard problem gets met with an instant answer, what happens to patience? What happens to struggle? What happens to the part of learning where you have to sit in confusion long enough for something to click?

That version of the future still worries me.

It is not hard to imagine a generation that becomes less capable because they have been taught to demand a response to every blank page, every equation, every essay, and every question, without ever understanding what it takes to get there. If the machine always rescues you, eventually you stop building the muscles required to rescue yourself.

That is the cynical view, and for a while I thought it was the only honest one. Lately I have been thinking there is another version of this story, and it is a lot more interesting. The best case for AI in education is not that it makes learning easier. It is that it makes learning fit.

The boring science nobody argues with

None of this is a new argument. It has just never been practical at scale.

In 1984, Benjamin Bloom published what became known as the “2-sigma problem.” Students who got one-on-one tutoring with mastery checks along the way dramatically outperformed students in a standard classroom. That is not a small lift. Roughly the gap between an average student and one in the top 2%.

The follow-up research never fully reproduced that leap, but the direction held. VanLehn’s 2011 review found that even intelligent tutoring software moved students from average toward roughly the top quarter of the class. Human tutors landed in the same neighborhood. Dunlosky’s 2013 review catalogued what actually moves the needle and kept landing on the same boring winners. Retrieval practice. Spaced repetition. Interleaving. Worked examples. Not lectures. Not worksheets. Not personality quizzes.

The bottleneck was never knowing what works. The bottleneck has always been attention per student. One teacher cannot simultaneously tell thirty different stories about the same concept, pace thirty different timelines, and diagnose thirty specific misunderstandings in real time.

AI might be the first tool with a realistic shot at that kind of scale.

Nobody is a “visual learner”

One thing out of the way first, before this starts sounding like a BuzzFeed quiz.

The idea that some people are “visual learners” while others are “auditory” or “kinesthetic” is not supported by the evidence. The Association for Psychological Science put it plainly. There is no credible case that matching teaching style to a student’s preference improves learning. A 2008 paper walked through the research and found the effect was close to zero. A 2017 follow-up put it even more bluntly. The title was “Stop Propagating the Learning Styles Myth.”

Around 90 percent of teachers still believe in it anyway.

So when I say “fit,” I do not mean matching some fixed sensory preference in a student’s head. I mean something more grounded. Meeting the student at their actual prior knowledge. Pacing to their actual speed. Swapping the representation of an idea when the first version fails to land. Content determines the best representation, not the learner’s self-reported type. Geometry needs diagrams. Poetry needs sound. History needs narrative. None of that is controversial, and none of it requires the learning-styles myth.

What “fit” actually looks like

For as long as school has existed in its modern form, students have mostly been asked to adapt themselves to the delivery mechanism. Same lecture. Same worksheet. Same pacing. Same explanation. Same tests. Some students happen to match the teacher and thrive. Some students survive. Some students quietly decide they are bad at a subject when really they were just handed it in the wrong language.

That part matters more than people admit.

A good AI-driven learning system does not just re-explain the material. It checks understanding constantly. It adapts difficulty in real time. When the first example does not land, it pulls up another one from a different angle. It drops the level when the student is lost and lifts it when they are bored. It catches the specific error, like “you forgot the negative sign in step two,” instead of just marking the whole answer wrong.

That is the part that keeps sticking with me.

So much of education gets blamed on the student when the real failure is the delivery system. We treat mismatch like inability. We confuse boredom with incompetence. We let one bad fit with a teacher shape an entire relationship to a subject.

I’ve generally done well in school and beyond. Geometry was the one class that never really clicked. It wasn’t a lack of effort, and it wasn’t ability. The pace and the delivery just never lined up. Looking back, there was no mechanism for translating the material into a form that actually worked for me. No way to try the concept from a second angle. No way to check whether the first angle had even landed. No way to notice that I had been quietly lost since the first two weeks of the semester.

That is not a small thing.

There are probably millions of people walking around with fake conclusions about themselves because school delivered a subject badly and called the result merit.

The uncomfortable part

The current evidence on AI in education is not a clean story. It is also not all on my side.

A 2025 MIT Media Lab study put students through essay-writing tasks while measuring their brain activity with EEG. The group using ChatGPT showed the weakest neural connectivity in the study. Their sense of ownership over their own work dropped. By the final session, many were mostly copy-pasting. The paper describes the LLM users as “consistently underperforming at neural, linguistic, and behavioral levels.”

Wharton researchers, led by Bastani in 2024, tested GPT-4 as a math tutor with Turkish high-school students. In-session performance jumped by roughly 48 percent with the AI available. Then they pulled the AI away and sat the students for a closed-book exam. Performance dropped about 17 percent compared to students who had never used it. Classic pattern. Scaffolding becomes a crutch.

In a 2025 survey of 319 knowledge workers, Microsoft Research and Carnegie Mellon found that the more confident people were in the AI, the less critical-thinking effort they applied, and the more homogenized their output became.

And Khan Academy’s own 2024 efficacy report on Khanmigo, plus follow-up analyses like this one in EdWeek, show some engagement gains and mixed learning outcomes. Not the revolution the initial pitch suggested.

Read together, those studies do not kill the thesis. They sharpen it. The same thing that makes LLMs feel magical is the thing that undermines learning. Fluent, confident prose on demand. Instant answers. No friction. That is the opposite of what the research says actually works. If the AI hands the answer over too quickly, the student never builds the thing in their own head.

So the real question is not “does AI help learning.” It is whether we can build products that enforce retrieval, productive struggle, and mastery instead of shortcutting them. Those are very different systems. Most of what is shipping today is the wrong kind.

The pattern is older than AI

Audrey Watters has been writing about this for years in Hack Education and Teaching Machines. Every wave of educational technology since B.F. Skinner’s teaching machines in the 1950s has promised some version of “personalized learning at scale” and has mostly delivered worksheets in a new package. The warning in her work is not that the promise is fake. It is that the hype keeps running about a decade ahead of the reality, and the casualties in between are real kids.

Dan Meyer, who has spent a long time thinking about how people actually learn math, has made a related point about current AI tutors. They default to hint-giving and answer-shaping that short-circuits the productive struggle that is where math learning actually happens. A fluent, confident chatbot that helps a student past every moment of discomfort is not a tutor. It is an escape hatch.

None of that means the idea is fake. It means the first generation of products is mostly missing the point.

The garbage around AI is the bigger threat

Honestly, the bigger danger to younger generations is not AI in classrooms. It is the garbage surrounding AI everywhere else.

The real poison is clickbait. It is bullshit videos, fake gurus, thinly disguised vaporware, and an endless stream of people promising impossible outcomes because hype pays better than honesty. That ecosystem trains kids to chase shortcuts, confuse marketing with substance, and expect results without contact with reality. It is not education. It is monetized delusion.

And the effect of that stuff is not abstract. It changes plasticity. It changes what people believe work even is. If all you see is “build an app in one prompt,” “make passive income with no skill,” or “launch a company this weekend with AI,” eventually your brain starts calibrating itself to fantasy instead of effort.

I see it up close too. Someone has an app idea. Might even scrape together a little bit of funding. Goes hard on mockups and designs. On paper, that all sounds great. In reality, it is window dressing. No grasp of what a real payment system involves. Security is an afterthought. No plan for maintaining the thing after launch. Just enough surface area to look real to someone who does not know what they are looking at. And if one of these does make it to production, the risk lands on actual customers. Some of these ideas can absolutely come to life. I just see far more “this is going to be a disaster” than “this is a real thing.”

And the depressing part is that plenty of companies will still reward that kind of thing because they are also being trained by the same hype loop. The issue is not ambition. The issue is a culture that keeps teaching people presentation is the product.

That is where I think a lot of the real danger lives. Not in AI as a learning tool, but in a broader social willingness to lie, posture, and prey on the uninformed because there is always money on the other side of the click.

Where this leaves me

My hot take is not that AI will save education. It is that it might finally expose how bad we have always been at personalizing learning, and how much wasted human potential has been written off as laziness, disinterest, or low ability when the real issue was poor translation.

The risk is still real. The magic-thinking problem is still real. The temptation to use AI as a substitute for effort is still real. The current crop of products mostly makes those risks worse, not better. If the design goal stays “engagement and fluent answers,” we will end up with a generation that is more entertained, less capable, and more convinced they understand things they do not.

But if we are serious and honest about this, there is another possibility on the table. A system that enforces retrieval, insists on productive struggle, gates progression on actual mastery, and adapts the representation of a concept until it lands. Not a chatbot that does homework for you. A translation layer between the curriculum and the student.

Maybe AI does not make people dumber than a brick.

Maybe, if we build it with any respect for the actual science of how people learn, it makes it a lot harder for broken systems to keep convincing smart people they are.