TL;DR: Generative AI means essays, assignments, and reports can be produced at a keystroke. The assumption we have relied on for centuries, that a polished written product reflects personal ability, has quietly collapsed. In 2022, “The Death of the College Essay” fired the first shot. This piece argues: when assignments can be generated, the one thing that cannot be generated is the actual process of doing something in a real situation. When I led children through real tasks in 2018, that was still an educational preference. After AI, it became a necessity.

A friend who teaches at a university once described a scene I have not been able to forget. One of his students submitted assignments that were fast, polished, every sentence flowing, structure impeccable. The same student, asked to explain what he had written during an oral exam, fell apart completely, could not say what he meant or why. This was not an isolated case. After AI became widespread, many teachers noticed the same bimodal pattern: a group of students using AI to dispatch written work in seconds, turning in perfect products, then being completely exposed the moment they had to explain themselves face to face.

What this scene exposes is more than cheating. It exposes an assumption we have held for centuries without quite stating it aloud.

An Unspoken Assumption Breaks Down

The assumption is this: the quality of a written product reflects the ability of the person who produced it.

Our entire educational assessment system has been built on that sentence. We judge university students by their essays, workplace professionals by their reports. A well-written document was taken as evidence that the person behind it understood the material, had thought it through. The inference held because, for a long time, producing a decent piece of writing actually required you to think clearly first.

Generative AI broke that chain. A polished product can now exist with no “having thought it through” behind it at all. In late 2022, The Atlantic published a piece with a direct title: “The Death of the College Essay.” Stephen Marche’s question was simple: when humanities disciplines use essays to evaluate students and confer degrees, and when that entire process can now be automated, what happens?

Three years on, the answer is becoming clear. The problem has deepened from a question of academic integrity (“will students cheat?”) into something more fundamental: when written products no longer demonstrate ability, what can we actually use to show that someone has learned?

The One Thing That Cannot Be Generated

AI can generate an article on how to build an e-commerce website, and probably write it better than most people would. But there is something else it cannot generate.

It cannot generate the full arc of four middle-school students facing a blank screen together, stuck on the question of what they were even trying to make, arguing, redistributing the work, and then, a month later, actually getting a website live. That history cannot be replicated. I wrote about this real task in “Four Kids, One Summer, One Website That Had to Go Live.” Its value was never in how polished the final product looked. It was in the judgments, the collaboration, the setbacks, and the corrections that the people involved actually lived through before the product existed.

That is the fundamental difference between a real task and a written assignment. An assignment produces a document, and documents can be generated. A real task produces an experience, one that unfolds in a real context, with a real audience, a real deadline, and a genuine risk of failure. That experience, generative AI cannot substitute for. You can ask AI to write a travel itinerary, but it cannot walk the Kumano Kodō for you or make the real trade-offs between your budget and your route. You can ask it to produce meeting minutes, but it cannot replace the moment when you actually persuaded someone in that room who disagreed with you.

When everything that can be generated loses its value, the processes that cannot be generated become the most valuable things of all.

From Educational Preference to Basic Necessity

Looking back, the timing is striking. When I ran that real-task project in 2018, ChatGPT did not exist.

At that point, my argument for real tasks over simulated assignments was a choice, a preference. I believed children learned more solidly under genuine pressure, that real stakes produced real learning, but I could not say simulated assignments were useless. They at least trained foundational skills. After AI, the situation changed. When a simulated assignment can be generated at a keystroke with no visible seam, its value as evidence of learning approaches zero. Real tasks stopped being “the better option.” They became one of the few reliable ways to actually cultivate and verify capability.

My educational philosophy did not get sharper. The world simply caught up, and I happened to be standing in a useful place.

The World Economic Forum’s 2025 Future of Jobs Report surveyed over a thousand employers covering 14 million workers, and estimated that 39% of core skills will change by 2030. The capabilities employers ranked highest were analytical thinking, followed by resilience and agility, leadership and social influence, and creative thinking. These share one feature: none of them come from memorizing answers. They grow through friction in real tasks, and they are precisely the capabilities AI is currently least able to replace.

In other words, AI is simultaneously making standard answers worthless and driving up the price of the ability to pull things together in the middle of ambiguity. I examined this from the employment side in “The Canary in the Coal Mine.” The education side is the same story seen from the other direction.

Learning Happens in the Struggle

There is a deeper layer here, about how learning actually occurs.

Learning science has a concept called productive struggle: genuine understanding tends to grow in the stretch where you are stuck, confused, and have to find your own way through. When answers arrive too easily, that cognitive circuit never gets trained. Writing works the same way. Writing is not simply recording what you have already thought, the act of writing is itself an act of thinking. When students use AI to generate or polish their work, what they bypass is not the effort of typing. It is the thinking itself.

AI is remarkably good at eliminating struggle. It delivers a serviceable answer in seconds, spares you all the frustration of being stuck, and presents the result neatly formatted. In many situations this is genuinely useful. In a learning context, it removes exactly the place where learning happens. As answers become effortless, productive struggle does not disappear, it becomes something scarce, something that has to be deliberately designed and deliberately preserved.

Real tasks are among the few things that can still reliably produce that struggle. The anxiety of the blank page, the friction of a team in conflict, the torment of ambiguous requirements, these “inefficiencies” are where cognition is actually built. This is also why I am not worried about AI outsourcing a child’s thinking, provided the child is still doing something real. The underlying cognitive mechanism is something I wrote about in “System vs. Intuition.”

So What Should We Do?

The direction is clear: give children things AI cannot do for them.

A task with a real audience. A goal that can genuinely fail. A project that requires working with others to complete. An experience that leaves a real mark on the person who lived it. AI cannot generate these things, because their value is not in the final product, it is in the path the person actually walked before that product existed.

I do not want to make this sound easy. Getting a real task to actually work has never been only a matter of commitment. It requires adults’ time. It requires resources. It requires money: teachers and mentors willing to stay genuinely involved, time to accompany and to ask hard questions, and the financial capacity to cover venues, tools, and transportation. Not every teacher and not every family is in a position to sustain this kind of investment over the long run. If anyone walks away from this piece thinking enthusiasm is enough, that is the last misunderstanding I would want to leave behind.

In “From Flipping to Climbing,” I wrote that the core of education is letting children encounter the world for real. AI has not changed that sentence. It has only made it more urgent. When a machine can generate nearly every assignment a child might be given, the only things we can be certain grew from the child are the things he actually built with his own hands, the walls he ran into himself.

Assignments can be generated by AI, but the experience of hitting a wall cannot. People grow up precisely through the grind of running into walls head-on.

A child’s frustration, the tears you and your child shed together, that confusion, and the process of groping through the chaos, all of it is a priceless asset. The irony is that our education system used to work so hard to help children avoid or remove these obstacles, while shouting in the same breath that it did not want the next generation to be replaced by AI.