Two sentences sit on our homepage, three-quarters of a century apart. The first is Turing’s, from 1950: instead of trying to program the adult mind, build the child’s, and subject it to “an appropriate course of education.” The second is ours, and it adds only one word of interpretation — for AI, that course of education is data.
It is easy to say a sentence like that. This essay is an attempt to mean it. Because if data really is an education, and not merely something a model consumes, then most of what our era says about data is wrong — and wrong in a way that decides what these machines become.
The miner and the schoolmaster
Our era’s favorite metaphor calls data the new oil. The metaphor confesses more than it intends. Oil is found, not made. It is drilled, refined, measured in barrels, priced by volume; it is indifferent to the engine that burns it, and the engine is indifferent in return. To call data oil is to ask of it exactly one question: how much?
An education answers to different questions entirely. Nobody praises a school for the tonnage of its syllabus. We ask what is taught, in what order, and why this and not that. A resource is valued by quantity. A curriculum is valued by judgment.
The first decade of large models was the era of found data. We let the child read whatever was lying around the house — and the house, being the internet, turned out to contain libraries. What happened next astonished everyone, including the people it happened to. But an era built on finding ends the way all extraction ends. The child has read the house. What is scarce now is not text but lessons, and a lesson, unlike oil, does not occur in nature. Someone has to decide what the student should meet next. That deciding has a name, and the name is teaching.
So the question of the coming decade is not a miner’s question but a schoolmaster’s: what should a machine read?
We spend our days composing courses of education for machines. The machinery changes weekly; the convictions underneath it have not changed at all. There are four.
A lesson is a question that can be failed
Information tells; a lesson tests. The difference between a paragraph and a lesson is that a lesson carries, inside itself, a way of being wrong.
So we hold to an old schoolroom rule: nothing reaches the student without an answer key. Not as bureaucracy — as definition. A question that cannot be failed is not a question; it is wallpaper. And the distinction matters more for machines than it ever did for us, because a model raised on unverifiable text learns to sound right, while a model raised on checkable consequences learns to be right. Nearly everything people mean by the word trust lives in the gap between those two sentences.
Turing chose his word carefully. He did not say the child machine should be exposed to things. Exposure washes over a mind. Education examines it.
Good questions are born from argument
A question composed by a single mind inherits that mind’s blind spots — and worse, it flatters its author. Whatever one intelligence finds hard to ask, it also finds hard to imagine asking.
The remedy is older than our field, older than any field: dialectic. One voice proposes; another resists, doubts, demands more. Questions that survive an argument come out harder and more honest than anything either voice would have written alone. And the same tradition hands down a second rule, cheap to state and expensive to honor: the mind that sets the exam must never be the mind that grades it.
A curriculum nobody argued over converges on the comfortable. And comfort teaches nothing.
Teach at the edge of failure
A student learns the least from what it already does well. The lessons that matter live in a narrow band just past the edge of competence — where the student fails, but fails intelligibly.
A curriculum, then, cannot be written once and administered forever. It has to watch. Where exactly does the student stumble? And what kind of stumble is it — a missing fact, a broken habit, two ideas wearing each other’s clothes? Every failure is a fragment of a map: the chart of everything education has not yet reached. The next lesson should be written on precisely that ground.
Education, in other words, is a loop, not a broadcast. The student’s failures revise the syllabus; the revised syllabus buys new and better failures. A teacher who never updates the lesson plan is not teaching. He is reading aloud.
A curriculum is defined by what it refuses
Of all that could be taught, nearly everything should not be. Collecting candidates is the easy part; the discipline lives at the gate. Most lessons that aspire to reach the student deserve to die on the way — too easy, too ambiguous, simply wrong, or merely a repetition of what is already known.
A library grows by acquisition. A curriculum grows by refusal. Each refusal is a small judgment about the kind of mind one is trying to build, which is why curation, done honestly, is a form of authorship. A dataset is not a warehouse of the world. It is a portrait of an intended mind, drawn in examples.
The weight of one word
Read Turing’s sentence once more and notice where the weight falls. Not on machine — he was confident of the machine. It falls on appropriate: a word he left sitting in the open, undefined, like a question on a desk waiting for the class to arrive.
The class has arrived. We have built the child beyond anything he imagined, and the appropriate course of education is still the open problem — no longer a footnote to the engineering, but the work itself. It is also the part that lasts. Models are trained, surpassed, retired; weights graduate like the classes of a school, year after year. What accumulates is the curriculum: the slow compounding of judgment about which questions are real, which difficulties are fertile, which truths can be checked. Students leave. The course of education stays, and every new student inherits a better one.
To choose a machine’s data is to choose what that machine becomes. A mind is made of what it has been given to read — for us a metaphor, for a machine the entire mechanism. So the question what should a machine read? turns out to be a larger question wearing work clothes: what do we want these minds to be?
No one will answer that by drilling deeper. It will be answered the way teachers have always answered it — one judgment, one lesson, one refusal at a time.