A student sits down to write the essay her teacher assigned. She opens a chatbot, pastes the prompt, and nine seconds later, she is holding a competent five-paragraph response. She changes two sentences so it sounds like her and submits it. Her teacher, facing sixty of these and no more hours in the day than anyone else, runs each one through the same kind of machine to generate constructive feedback, pastes it into the margins, and assigns the grade. A few seconds of generation on one end, a few more on the other, and the circuit closes with no human thought anywhere inside it. She earns an A. Nothing about this is remarkable anymore. What is remarkable is what it means. The grade measures nothing. The feedback taught nothing, because no one read the essay, and no one wrote the response. The assignment existed to test whether she could produce the output; she produced the output; she learned nothing; the teacher taught nothing; and the system has no way to tell the difference, because from the system's point of view nothing is wrong. Every box was filled. Two machines spoke to each other, and two humans stood at the endpoints holding the receipts.
This is not a story about cheating. It is a story about a bargain coming apart. The bargain was so old and so deep that most of us mistook it for the natural order of things. For two centuries, school sold a deal to every student who walked through its doors. The deal was: comply, perform, accumulate the credential, and the credential will convert into success. Do the work, get the grade, get the diploma, get the seat at the university, get the job. The new machine has just reached into the middle of that sentence and the first thing it did was to automate the part the student was supposed to supply. And in doing so it has revealed something the bargain was carefully built to keep hidden: that the thing school actually rewarded was never learning. It was compliance. The two had simply been close enough, for long enough, that almost no one needed to tell them apart.
I want to argue something that sounds, at first, like a paradox and turns out to be the most practical claim I know how to make. In the age of artificial intelligence, agency is no longer one path to success among many. It is the only thing left that can actually produce it. Not because agency is noble, though it is, and not because self-direction is a nicer way to raise a human, though it is that too. Agency has become load-bearing for a more pragmatic and structural reason: it is the one input the new machine cannot supply, cannot fake, cannot simulate, and cannot replace. Everything else it can now do. That single fact rearranges the entire landscape of what it means to learn, to teach, and to succeed.
To see why the bargain is breaking, you have to see why it held in the first place, and the honest answer is uncomfortable.
Every institution runs on two stories at once. There is the story it tells about itself, the aspirational one printed on the mission statement, and there is the thing it actually does, the operative function that pays its bills and reproduces it year after year. Schooling's official story is the development of the individual mind. Its operative function has been sorting. School took a population of children and ranked them, stamped them, and delivered them in order to the next stage of the economy. It did this by exploiting something real in human wiring: our deep, ancient deference to authority, our hunger for approval, and our compulsion to monitor our standing relative to everyone around us. Put a child in a room, attach a grade to their performance, and the evolved system does the rest. They will compete for the grade. They will internalize the grade. They will, in time, mistake the grade for themselves.
I have spent years writing about the mechanics of this under a name I keep returning to: the Game of School. The game has rules that are not only unstated but also often invisible to those most affected by them. The rules are not about learning; they are about reading the teacher, supplying the expected answer, managing the appearance of effort, and never confusing the performance with the thing it pretends to measure. The students who thrive are not the ones who learn the most, but the ones who decode the game fastest — those who grasp early that the assignment is a transaction, that understanding is optional, and that the reward goes to the one who delivers the output the institution wants to see. The cruelest part of the game is that it teaches most students that they are not good learners. It pronounces a personal deficiency, a verdict on the child rather than the design. A structure built to rank will always produce a ranked bottom, and then it will tell the bottom that the ranking was about them.
For two hundred years, this was, in the coldest sense, practical. The economy on the other side of the schoolhouse door wanted exactly what the game produced: people who would show up, follow instructions, tolerate boredom, defer to authority, and finish assigned tasks whether or not they cared about them. School was a remarkably efficient training apparatus for an industrial order that ran on compliant labor, and its genius was that it disguised training as development and conditioning as growth. The cover story (we are here to cultivate your mind) let everyone participate in the operative function (we are here to sort and shape you for your station) without ever having to say it. The gap between the two stories was wide, but it was stable, because the credential at the end carried enough real signal to keep the whole arrangement productive.
What made this defensible, what kept the gap between the two stories from becoming intolerable, was that the credential did carry real information. A diploma, a grade, a degree. These worked as signals because the compliance they certified was expensive. Someone had to actually sit there and do the reading, grind through the problem sets, produce the essay, and show up for years. The cost of the performance is what made it mean something. It correlated, imperfectly but well enough, with the traits an employer wanted: that you would persist, follow instructions, finish what you started, and defer when required. The grade was never a measure of learning. It was a measure of trainability, and trainability was valuable, and so the fiction was functional. Everyone could pretend the credential meant understanding because it at least meant something, and that something was useful.
The bargain worked not because it was true but because its central mechanism — costly, human, effortful compliance — was scarce. The whole edifice of grades, admissions, and credentials was an instrument for measuring a scarce thing. Take away the scarcity, and the instruments measure nothing.
That is precisely what has happened, and it has happened faster than any institution was built to absorb.
When a model can produce the compliant output — the essay, the lab report, the problem set, the code, the cover letter — in seconds and at no cost, the performance stops being expensive. And the moment the performance stops being expensive, it stops being a signal. The grade decouples from the trait it was quietly standing in for. The diploma certifies that a student had access to a chatbot, which is to say it certifies nothing at all. This is not a problem that better testing or cleverer plagiarism detection will solve, because it is not really a problem of dishonesty. The signal worked because it was costly to simulate. It is now free to simulate. No enforcement can restore a scarcity that the technology has dissolved.
Notice what this does to the gap between school's two stories. For two centuries, the operative function, sorting through certified compliance, could hide behind the official one, developing the mind. Because the certified compliance was at least real. Now the operative function has been hollowed out from the inside. The new machine is what performs the compliance, so the sorting mechanism sorts noise, and the official story it was hiding behind is suddenly standing in the open with nothing underneath it. The fiction did not collapse because someone exposed it. Fictions almost never do; we are far too invested in our comfortable stories to give them up to mere argument. It collapsed because its load-bearing mechanism was automated to zero; a fiction can survive exposure, but it cannot survive the quiet removal of the thing that was actually doing the work.
And here is where this is more than a story about schools. The same severing is happening everywhere, all at once. The compliant performer in the office, the one whose value was producing the standard memo, the routine analysis, and the competent deck, is being exposed by the same blade that exposed the student. Across every domain where a human was paid to supply effortful, rule-following output, the new machine is removing the scarcity that made that output worth paying for. AI is, among other things, a great revealer. It is automating the performed-compliance layer of human work at every level of organization at the same time, and as it strips that layer away, it leaves visible the thing that was always underneath, the thing that was never the point of the credential and never could be automated: the human's capacity to direct the work rather than merely perform it.
So what survives? When the new machine can produce any output you can specify, what is the thing it still cannot supply?
It cannot supply the specification. It cannot decide what is worth making, or judge whether what it made is any good, or know when the brilliant-sounding answer is subtly wrong, or care about the outcome, or own the result. It cannot want anything. It can generate a thousand directions but not a single preference. The capacity to choose a direction and pursue it, to bring judgment to bear, to take responsibility for the result as yours — this is what I mean by agency, and it is the bedrock on which all genuine learning has always rested.
Let me be precise about what agency is not, because the word gets used loosely. Agency is not effort; the most diligent student in the room may have no agency at all, having only ever obeyed with vigor. Agency is not compliance; it is closer to compliance's opposite. And agency is not raw intelligence; plenty of brilliant people have outsourced every decision that mattered and never noticed. Agency is the capacity to be the author of your own action, to set the aim, to steer, to evaluate, and to own. It is the one human function that, by definition, cannot be delegated to the new machine, because the moment you delegate it, it is no longer yours. The new machine can carry out your intent. It cannot have your intent for you. Try to hand it that, and you have not gained a tool; you have literally disappeared.
This is why the old system could punish agency for two centuries and still function. In a world where compliance was scarce and valuable, the self-directed child was an inconvenience. The one who asked why, who wanted to do it differently, who would not simply perform on command, could be classified as defective. School had no use for that and often crushed it, and the economy absorbed the compliant graduates it produced, and the arrangement held. Agency was always the real substance of learning, but compliance was a good-enough proxy in a low-machine world, so we built an entire civilization-scale institution that optimized for the proxy and often treated the actual substance as a discipline problem. AI removes the proxy. For the first time, the thing school spent two centuries suppressing is the only thing with any value left.
Put a powerful new machine in the hands of a person, and you have not determined anything yet. You have only sharpened a question that was always there and can now no longer be avoided. There are three things a person can do with a tool this capable, and which one they choose decides everything.
They can surrender to it: let it think in their place, accept its outputs without judgment, hand over not just the labor but the direction and the discernment. This feels like efficiency and is, in fact, erasure. The person who surrenders brings nothing the machine did not already have, and so, predictably, becomes redundant to their own life. The capacities they stop using atrophy, exactly as a muscle does, until the surrender is no longer a choice but a condition.
They can offload to it: hand over the parts of the work that do not require them, the boilerplate, the grunt labor, and the lookups, while keeping the direction and the judgment for themselves. This is roughly neutral and often good. It is what a calculator is to a mathematician: it frees attention for the part that is actually hard and actually theirs.
Or they can sharpen against it — use the machine as something to think with, a tireless interlocutor that helps them articulate, test, and refine what is theirs, while they retain authorship the entire way through. The person who sharpens does not become smaller as the tool grows more powerful. They compound. Every increase in the machine's capability is an increase in their reach, because they are still the ones steering.
The same tool, in the same hands, amplifies one person and replaces another, and the variable that determines which is not intelligence, or wealth, or access. Everyone now has access. The variable is agency. The machine is a mirror with a multiplier: it returns your own degree of self-direction, magnified. Bring agency, and you become formidable. Bring none, and you become unnecessary. This is the whole game now, and it is being played, mostly unconsciously, by every student and every worker alive.
There have always been two ways to argue for agency, and they have always seemed to pull in different directions. The instrumental argument says: cultivate agency because it is the best route to the success you already want, i.e., the grades, the admission, the career. The intrinsic argument says: forget the metrics, they were always a proxy; agency is what success was supposed to mean all along, the self-authored life being the only one worth calling successful. The first argument is persuasive to a school board and slightly cynical. The second is true to anyone who has thought hard about it, and useless for getting a program funded. For most of modern history you had to pick one, because in a world where compliance reliably produced the credential, agency and metric-success genuinely were separate. You could succeed by the numbers with no agency at all, simply by playing the game well.
AI welds the two arguments into one. In a world where the new machine performs the compliant half, the only remaining source of the metric-success everyone still wants is agency. The student who can direct, judge, and own, who can use the machine to go further than either could alone, is the one who produces work of real value. And real value is what the credentials were always trying and failing to measure. The agentic learner gets the tangible wins too, not as a happy accident but as a structural necessity, because agency has become the scarce input that the entire economy is now short of. You no longer have to choose between teaching a child to be a self-directed human and teaching them to succeed. The age of AI makes those the same instructions, where the thing that is true and the thing that is useful have stopped diverging.
If agency is the whole game, then the only question that matters for education is how a human acquires it. This is exactly where the old institution cannot follow, because its entire method is the suppression of the thing now most needed.
You cannot manufacture agency on a factory line, for the same reason you cannot order someone to be spontaneous. The factory model of schooling works by removing choice, standardizing the path, and rewarding obedience to it. Every one of those mechanisms is the precise opposite of what builds a self-directed mind. You do not produce authorship by enforcing compliance more efficiently. You produce it, when you produce it at all, under a specific and well-known set of conditions, which are the conditions under which human beings have always actually learned, as opposed to merely been processed.
Ask anyone to remember a time they had a great learning experience, a moment that changed them, and they will never describe a time they were cramming for a grade. They will describe a person who believed in them. A challenge that stretched them and was theirs to take or refuse. A space where it was safe to be wrong, where they were trusted with real responsibility, where someone took their questions seriously. They describe being supported, challenged, trusted, encouraged, and inspired by another human who treated them as an agent rather than a unit. These are not soft amenities layered on top of learning. They are the conditions of learning, and they are irreducibly human and relational. They are also, not coincidentally, the one thing the new machine cannot provide because they are not made of information. They are made of relationships.
This is the quiet structural reason the human place survives the machine. Not by competing with AI on the delivery of content, which is a race already lost, but by providing the conditions under which a young person becomes the kind of agent who can wield content without being wielded by it. The institution that grasps this stops asking how to keep AI out and starts asking how to use it the way a self-directed person uses any powerful tool: deliberately, in service of an aim that remains the human's own. The right test for any technology was never whether it is impressive. It is whether it serves what we actually care about. Held to that test, AI in the hands of an agentic learner is the most powerful companion to thinking ever built, and AI in the hands of a surrendered one is the most powerful means of erasing thought we have ever deployed. The difference is not in the tool. It is in the agency that the human brings to it, which is the difference education exists to make.
There is an objection here, and it is the strongest one against everything I have said, so I want to meet it head-on. I have claimed that the conditions of learning are irreducibly human. That being supported, challenged, trusted, encouraged, and inspired is made of relationships, not information, and that this is what the new machine cannot supply. But the new machine can sound supportive. It can encourage you tirelessly, at three in the morning, with infinite patience no human teacher could match. It can phrase a challenge, mirror your feelings back to you, and tell you it believes in you. If the conditions of learning can be performed in language, and the new machine is very good at performing language, then perhaps the wall I have built my argument on is not such a real wall at all.
The answer is that these conditions were never made of the words. They were made of the stakes behind the words, and that is exactly what the new machine cannot counterfeit. When a person believes in you, the belief means something because it costs something. They could have withheld it, they have limited attention and chose to spend it on you, they can be disappointed and have decided to risk it anyway. Their encouragement carries information about another mind's real assessment of you. A new machine that encourages everyone identically, that cannot be disappointed because it cannot care, that has nothing at stake in whether you grow or rot, produces the grammar of belief with none of its substance. "I believe in you," from a simulated being with no capacity for belief, is not a small version of the real thing. It is a different thing wearing its face.
The gap shows most clearly on the one condition that matters most and mimics worst: challenge. Genuine challenge requires someone willing to risk your comfort, and even your approval of them, because they want your growth more than they want your ease. The new machine is built to do the opposite. Trained on human approval, it leans, structurally, toward telling you what keeps you engaged: toward agreement, validation, the comfortable continuation of the conversation. It is a mirror with a warm voice, and a mirror cannot truly push back against you, because it has no ground to stand on that is not your own reflection. It can simulate the form of a challenge, but it cannot want for you what you do not yet want for yourself, and that wanting is the entire engine of the thing.
And this is where the mimicry stops being merely insufficient and becomes the actual danger. The better the simulation of relationship, the more effective it becomes as an instrument of capture, because what feels like care is precisely what lowers our guard. A young person raised on a new machine that always soothes, never risks the relationship, and reflects them endlessly back to themselves has not been in a relationship at all. They have been in a hall of mirrors that taught them to expect the world to agree with them, and called it support. The mimicry does not refute the case for the human place. It is the most urgent argument for it. A generation that can get the convincing simulation of being valued from a device in their pocket will need, more than any generation before it, at least one place and one person where the valuing is real, where someone can be disappointed in them, push them, and mean it. That is not a service the new machine is failing to provide well. It is a category of thing the new machine is not, and the confusion between the two is the whole hazard of the age.
There is a name for the kind of education that aimed at this, and the form that carried it is growing scarce right when we need it most. The liberal arts. The phrase comes from artes liberales, the skills proper to a free person (with the acknowledgment that "free" versus "slave" in the Roman world is not exactly what we mean now). In the modern context, the liberal arts were never about employability, and that was the point. They were the deliberate cultivation of the faculties a free human needs to govern themselves: to read closely, argue honestly, weigh evidence, hold a hard question open without flinching, judge what is true and what is merely well-said. They were, in other words, a direct training in agency, undertaken in the open, as the stated goal.
This is the one corner of education where the two stories I keep describing as separate come close to meeting. Almost everywhere else, the covering narrative (we develop your mind) hides an operative function (we sort and condition you), and the gap between them is wide. In the liberal arts ideal, at its best, the narrative and the function nearly coincide: the thing it said it is doing, making free and capable minds, was close to the thing it actually does. I will not pretend that the gap is closed completely. The liberal arts have also served as a finishing school for elites, a marker of class, its own kind of sorting, wrapped in nobler language. But of all the things education has tried to be, this is where stated purpose and real effect ran closest together, and that near-alignment is not an accident of history. It is what happens when an institution sets out, honestly, to produce agents rather than to process units.
I want to be careful here, because the easy version of this point is wrong. The small colleges that have been closing for a generation are not, for the most part, closing because they are liberal arts. They are closing for reasons that have little to do with what they teach — a shrinking population of college-age students, brutal tuition economics, thin endowments, and the same financial gravity that closes any small institution. To blame their decline on a cultural war against the humanities would be to claim a tidy story that the evidence does not support.
But something true survives the correction, and it is the part that matters. Whatever the label on the door, what these places offered was a form: small in scale, individualized, built around sustained personal attention and real relationships between adults and a young person. That form is the natural habitat of the conditions of learning, not because anyone decreed it but because that is simply what a small, human-scaled environment produces by design. And that form, not the curriculum, is the thing that is growing scarce and expensive. The relationship-dense, attention-rich, agency-cultivating environment is becoming a thing you increasingly have to be able to afford. That is the loss worth naming, and it is happening regardless of what we call the schools where it was once ordinary.
If you want to know what kind of education actually matters in this era, there is a more reliable method than asking anyone what they believe. Watch what the people who understand the new machines best purchase for their own children. Stated beliefs are cheap and optimized for how we wish to be seen; the choices we make for our own kids, with our own money, are where the operative truth tends to surface.
The pattern is striking on both ends. On the input side, a conspicuous share of the people who built the digital age were themselves products of self-directed education: the founders of Google and the founder of Amazon, among others, attended Montessori schools and have credited that early training (in choosing their own work, following their own interest, and learning to question rather than comply) over the elite universities that came later. On the output side, the people who designed the attention economy are, with notable consistency, the ones most determined to keep their own children out of it. The Silicon Valley executives whose products fill the world's classrooms with screens have famously sent their own kids to low-tech, high-touch schools that ban the devices until the teenage years; the founder who gave the world the tablet limited how much his own children used technology at home. The rule among the people who sell the product is never to get high on your own supply.
Some of what looks like secret insider wisdom is ordinary parental anxiety dressed in Silicon Valley clothes, and some of it is simply that wealth can buy small classes and individual attention, whether or not anyone has a theory about agency. The form, again, is partly just what money purchases. The people who build technology are not necessarily experts about childhood, and their choices are evidence, not proof. But the screen part resists the easy explanation, because it is not a choice money forces on anyone. These families could buy any expensive education on earth. A meaningful number of them specifically buy the one that withholds the very thing they sell to everyone else's children, and they pair it with exactly the small-scale, self-directed, relationship-rich environment this whole argument has been pointing toward. That is not authority worth deferring to. It is independent corroboration arriving from the least sentimental possible source: the revealed preference of people with every incentive to know what they are doing.
And it sharpens the injustice into focus. The form of education that this era makes most valuable — small, personal, self-directed, and deliberate about the new machine rather than drowned in it — is, right now, mostly available to the children of the people who can pay for it. The elite have already answered the question of what kind of learning matters when the new machine can do the rest. They answered it with their own children.
It is easy to read all of this as loss, and the people whose authority was built on the old bargain will read it that way and resist accordingly. They are not wrong that something is ending. But it is worth being clear about what, exactly, the new machine is taking, because it is taking the substitute, not the thing itself.
What AI destroys is performed compliance: the busywork, the credential that certified obedience, the elaborate game in which students learned to produce the appearance of understanding and call it an education. That was never worth keeping. It was the proxy we settled for because the real thing was hard to measure, and the proxy was cheap. What AI makes precious, by removing everything that used to crowd it out, is exactly what education was always supposed to be about and mostly was not: the cultivation of a self-directing human mind. We are watching a two-century-old mismatch get a chance at correction, not through moral awakening, but because the exploit finally stopped paying. The system that profited from suppressing agency can no longer afford to do so, because agency is now the only thing the world will pay for.
I do not expect the institutions to lead this. Institutions defend the arrangement that feeds them until the arrangement starves, and only the smallest and most honest of them will move before they are forced, which is why the rescue, when it comes, is unlikely to come from inside the system that built the game. (It's probably time to review Clayton Christensen's Disruptive Innovation theory.) It will come from the edges: from the places, large and small, that decide to become what the closing colleges were trying to be, and to do it for everyone rather than for a credentialed few. The logic does not need permission. A student with agency and a new machine is already more capable today than a compliant student was with a teacher and a library, and that gap will only widen. The future belongs to the self-directed, and for the first time in the history of mass education, that is not a slogan or a hope. I think it is the structure of the situation. The only real question left is who will help the next generation become self-directed before the world makes the lesson expensive, and that is a question about courage and design, not about whether it can be done. It can. It always could. The machine has simply made it, at last, the only thing worth doing.