There is a specific kind of discomfort that arrives when someone you respect says something true that isn’t quite true enough.
Dan Meyer’s recent interview in Education Week is careful, honest, and worth reading. A former math teacher, a curriculum developer, someone who has spent years at the intersection of technology and learning — he says what many educators feel but struggle to articulate: AI has not transformed math instruction, and the evidence that it will is far weaker than the enthusiasm suggests. Only 3% of teachers report using it significantly. The studies claiming big gains dissolve under scrutiny. The relationships students actually need — high-trust, reciprocal, demanding, and affirming at once — are not relationships AI can form.
He is right about all of it.
And something in that rightness stopped me. Because the question underneath his answer is one I don’t think the interview ever asks.
What if the instruction we’re trying to improve was already broken?
Not broken in the way people usually mean it. Not underfunded, or undertested, or underscaled. Broken at the level of theory. Broken in how it understands what mathematics actually is.
For most of its institutional history, mathematics education has operated on a single premise: mathematics is a body of content to be delivered and a set of procedures to be practiced. The teacher carries the knowledge. The student demonstrates receipt of it. Correct answers confirm transmission. Tests measure coverage.
That premise is not a failure of effort. It is a failure of imagination — a category error so old and so widespread that it became invisible. We confused the instrument for the music.
Mathematics, at its core, is not computation. It is not even problem-solving, though problems are part of its grammar. Mathematics is a discipline of attention. It is the practice of turning raw perception into shareable judgment — of noticing what has structure, naming what deserves focus, holding competing possibilities long enough to choose a wise next move. It sharpens the senses. It gives magnitude to feeling. It makes the invisible navigable.
When you understand it this way, something shifts.
Meyer describes, with precision, the classroom moment AI cannot replicate: a teacher nods at a tentative idea a student has offered. Other students build on it. Something becomes possible in that room that wasn’t possible before. He calls it “socialized” learning, in contrast to the personalized isolation of one child and one chatbot.
He’s right that this moment matters. But I want to stay with why it matters, because I think the reason goes deeper than relationship, important as relationship is.
That moment — the nod, the building, the room holding a thought together — is the emergent sensemaking of intellectual behavior made visible. Sense. Model. Move. Update. The student who offered the tentative idea sensed something (intuition/perception). The teacher and the localized learning community helped them model it, and then the community's emergent possibilities were updated.
This is what mathematics education is supposed to train. Not the emergent system’s outputs — not the correct answers — but the sensemaking orientation itself. The capacity to notice, to represent, to act, to revise. The willingness to hold uncertainty long enough for it to become information. AI cannot train that emergent process. Not because it lacks warmth, though it does. Because it operates only in the middle — one part of the community — and even there, it needs a well-formed question to work from. It needs someone who already knows how to sense. It needs an expert.
This is what Meyer observes when he notes that AI benefits experts far more than novices. I think he’s right, and I think the reason is structural, not temporary. To use AI well in any domain, you have to already know how to pay attention in that domain. You have to be able to ask the right question, evaluate the response, detect the moment the model has failed. In mathematics, that means you have to have already developed the very capacity that mathematics is supposed to build.
The tools have been wrong because the theory was wrong first. We built information-distribution machines to serve an information-distribution theory of learning. And when those machines failed — when the 95% of students who needed something other than transmission didn’t get it — we built better machines. Faster. More adaptive. More personalized.
And we kept not asking the question: personalized toward what, exactly?
If the answer is correct answers, we have a technology for that. It is called a calculator. We have had it for fifty years, and it has not taught a single child to think mathematically. If the answer is sensemaking — attention, perception, the felt sense of when a representation fits the moment and when it doesn’t — then we don’t need better machines. We need a different theory of what we’re doing. One that treats mathematical concepts not as content to be covered but as perceptual territory to be inhabited. One that sees the classroom not as a delivery system but as the epistemological structure within which attention is actually trained locally — through friction, through other minds, through the slow, irreplaceable practice of noticing together.
Meyer ends his interview with a line I keep returning to.
“The work is still the work.”
He means it as sobriety. Don’t let the noise of AI distract you from the real challenges: keeping students in school, making them feel known, and helping teachers find meaning in a difficult profession.
I hear all of that. And I want to hear something else in it, too.
The work is still the work. Meaning: we have not done it yet. Not because the tools have been wrong, though many have. Because the theory has been wrong. We have been teaching procedures and calling it mathematics. We have been measuring outputs and calling it learning. We have been building instruments and forgetting the music for so long that most people — students, teachers, administrators, policymakers — have never heard what it actually sounds like.
AI will not transform math instruction. Not because it lacks the power to distribute content more efficiently — it is extraordinary at that. But because math instruction, at its best, is not content distribution. It is the slow cultivation of human attention. And that is not a problem any tool can solve. It is a problem only a different way of seeing can solve.
That’s the work. It has always been the work. We are only now, perhaps, running out of excuses not to do it.