We Are the Last Generation to Have Learned Without AI. Here's What That Means.
Learning without AI built understanding through friction and failure — why the last generation to learn the hard way carries something now becoming scarce.
I learned to code in 2013 by being confused, searching for help, being confused again, and then eventually understanding something. The understanding was slow. It was uncomfortable. It was also, I now think, irreplaceable.
My nephew is twelve. He has never written code without an AI suggesting the next line. He has never had to hold a problem in his head for hours, turning it over, waiting for the angle that makes it solvable. He does not know what that feels like. He will never know what that feels like.
I do not think this makes him worse. I do not think it makes him better. I think it makes him different in a way we do not yet have the language to describe.
What Does Learning Without AI Build That AI-Assisted Learning Doesn’t?
It builds structural understanding — the ability to recognize not just the answer, but why the wrong approaches were wrong. That kind of knowledge is not in any documentation. It is built through the experience of a problem actively resisting you.
When I was learning, the friction was the entire education.
I would get stuck on a problem for an afternoon. I would try six wrong approaches before finding one that worked. I would understand, eventually, not just the answer but the shape of the problem — why the wrong approaches were wrong, what constraints made the correct approach correct. This understanding was not in any documentation. It was built through the experience of the thing resisting me.
Cognitive science calls this desirable difficulty — the finding that learning is more durable when it requires effort. UCLA psychologist Robert Bjork, who coined the term in the 1990s, demonstrated that introducing manageable obstacles during practice produces better long-term retention than frictionless instruction. Problems that are easy to solve are easy to forget. Problems that cost you something tend to stay.
I am not saying the cost was worth the suffering. I would have happily had an AI in 2013. But I am saying that the cost produced something: a kind of structural understanding that is different from knowing the answer. The difference between having found your way through a city on foot and having been driven through it. You know the city differently. You cannot always explain how, but you do.
What Do AI-Native Learners Gain — and What Do They Miss?
They gain speed, fluency with AI systems, and the ability to direct tools toward goals. What they may miss is the transferable depth that comes from having solved problems in genuinely unassisted environments — where no prior pattern applies and no tool can substitute for reasoning from scratch.
This is the question nobody is asking clearly: if the friction built something, and we remove the friction, what replaces what it built?
The honest answer is: we do not know yet.
What the next generation will have is something different. They will interact with AI systems from the beginning of their cognitive development. They will learn to frame problems, to evaluate outputs, to direct systems toward goals — skills that my generation is learning in our thirties and forties after learning the previous way first. They will likely be better at certain things than we are. They will likely be worse at others.
A 2024 study from Wharton’s Operations, Information and Decisions department tracked exactly this split. Researchers gave nearly 1,000 high school students access to an AI tutor for math practice. Students who used it performed substantially better during AI-assisted sessions — but scored significantly worse on subsequent assessments taken without AI assistance. The paper’s title is direct: Generative AI Can Harm Learning. The finding was not that AI made students worse overall; it was that AI dependency formed quickly enough to undermine independent performance on tasks the AI had not seen. Put plainly: the tool that raised their scores in the room lowered them the moment it left.
The question is what we are optimising for. If we optimise for task completion, AI-assisted learning is superior. If we optimise for the ability to operate in genuinely novel situations where no prior pattern applies, the picture is less clear.
Is AI Now Being Trained on Thinking That Was Already Shaped by AI?
Yes — and nobody knows exactly what that produces at scale. The training data being generated today reflects a cognitive texture already shaped by AI assistance, which means future models will increasingly learn from human thought that has been pre-filtered through AI tools.
Here is something I think about regularly.
The AI systems being built today were trained on content produced by people who learned without AI. The code on GitHub, the explanations on Stack Overflow, the documentation written by people who understood systems through years of working with them — this is what current AI knows.
That data captured something: the thinking of humans who built their understanding through friction.
The data being generated now is different. Developers who use AI agents think differently. They reach for a prompt before they reach for a solution. They work in handoffs and reviews rather than line-by-line construction. A 2023 study published in the Communications of the ACM found developers using GitHub Copilot completed tasks 55% faster than those working unassisted — a speed gain that inevitably reshapes how problems are approached before they are even fully understood. The texture of that work is changing too: GitHub reported in 2023 that Copilot was already writing around 46% of code in files where it was enabled. This is not worse. But it is a different cognitive texture, and that texture is becoming the training data for the next generation of models.
We are not, in any simple sense, training AI on human thought anymore. We are training AI on human thought that has already been shaped by AI. The feedback loop is closed. Nobody fully knows what it produces.
Should We Deliberately Keep AI Out of Early Learning?
Probably yes, in specific contexts — not as punishment, but as deliberate friction. The same logic that makes athletes train without competition-day equipment applies here: the body, and the mind, need to know what they are capable of without the tool before they can use the tool well.
I have a position on this, though I hold it carefully because I could be wrong.
The thing worth preserving is not syntax knowledge. Let syntax go — I am happy to never memorise another API. The thing worth preserving is the experience of genuine difficulty. Of holding a hard problem until it yields. Of building understanding through resistance rather than through assistance.
This probably means designing learning environments where AI is deliberately not available — not as punishment, but as deliberate friction. The same way athletes train without equipment that they use in competition. Not because the equipment is bad but because the body needs to know what it is like without it.
It probably also means being honest with young people about what AI can and cannot do — specifically, that it can produce answers without building the understanding that makes answers useful in new contexts.
I do not know how to preserve this at scale. I am not sure it is possible at scale. But I think the attempt matters.
What Do People Who Learned Before AI Have That Others Won’t?
The advantage is not knowledge — knowledge has always been abundant. It is judgment: the instinct that something is wrong before you can articulate why, earned through years when slowness and friction were the only methods available.
My generation — the last to have learned without AI — carries something that will become scarce.
Not knowledge. Knowledge is abundant now. The ability to search has always made knowledge abundant. What we carry is something harder to name: the feel of a system, the instinct that something is wrong before you can say why, the judgment built in years when there was no alternative to building it slowly.
This will matter most in the situations AI cannot handle — the genuinely novel problems, the edge cases outside any training distribution, the moments when what is needed is not a pattern from prior experience but the ability to reason from first principles in a context that has no precedent.
Those moments will still come. They always do. They are, arguably, the only moments that really matter.