Machines have spent the last decade getting radically better at learning. Have we spent it quietly getting worse at learning ourselves?
Brute Force
On September 11, 25 Fields Medallists—nearly every living one—published a declaration about AI in mathematics, and their argument isn't that AI can't do math. It's close to the opposite: AI has become good enough to solve real open problems—and that's exactly the problem! Solving a problem was never the actual goal of research. It was evidence that a community had built new understanding along the way, through years of talks, simplification, and teaching that turned one insight into something the whole field could use. In their own words: mass-producing true-or-false statements at speed "could destroy fertile ground instead of breathing life into new ideas." It's brute force in a fairly literal sense: enough compute thrown at a problem to reach the finish line without the years of understanding that used to get built walking there.
Ten Years Down
PISA measures the same chain from the other end, and its trend line isn't a single bad year. Reading and math scores across the OECD have been falling since 2015 without recovering, and the 2025 results, published 3 days before the Fields Medallists' declaration, put both at their lowest levels on record. Between a 15-year-old and a Fields Medallist sits the entire chain that produces expertise—student, junior, specialist, researcher—and both ends of it are moving in the same direction at once.
Not Just Mathematics
The same worry shows up wherever AI is making real breakthroughs, not only in research mathematics. A 2026 interview study of software engineers in South Korea describes debugging, small refactors, and reading someone else's bad abstraction, exactly the work that used to train juniors, being absorbed straight into senior-plus-AI workflows, with the struggle that used to build judgment disappearing along with it. Law and finance describe similar concerns from their own corners. One mechanism, several fields: automation reaches the exact tasks that used to manufacture expertise, because those are also the cheapest ones to automate.
A Dilemma. Everyone's a Prisoner?
What connects all of it is closer to game theory than to any one technology. Every firm, lab, or school has an individual reason to automate the entry-level work that happens to be exactly what used to build expertise, and every one of them is individually right to do it. But if every actor in an industry makes the same rational call, the industry collectively cuts through the transmission chain that trained its own experts, without any single decision along the way being wrong on its own terms.
Economists call the shape of it a "tragedy of the commons," but it plays out more like a prisoner's dilemma run at the scale of an entire field: everyone defects, nobody colludes, and the outcome is worse for everyone than cooperation would have been. Except there's no rival across the table extracting a confession here. The prisoner is us all. The savings show up this quarter. The judgment that would have been built instead shows up, if it ever does, a decade from now, inherited by whoever may not even remember the choice being made.
Stock, Flow, Debt
Real economic literature is now decomposing this mechanism formally. Mustafa Seref Akin's 2026 paper, "The Apprenticeship Externality: Generative AI, Entry-Level Work, and the Future Supply of Expertise," names it the apprenticeship externality: a firm automates a junior task whenever its own private gain is positive, without pricing in the learning value that task would have produced for someone else entirely, a future employer, a future client, the profession generally. The tasks a spreadsheet flags first for automation, low productive value, high training value, are exactly the ones a mentor would have protected.
Another NBER paper, "Automation, Learning, and Career Dynamics" by Afrouzi, Blanco, Drenik, and Hurst, finds something sharper than a trade-off: an economy with strong learning-by-doing can settle into either of two stable states. In one, cheaper automation raises welfare because the learning channel stays open. In the other, it tips the system into what they call a human-capital trap, crowding out the tasks people learn from, which shrinks the future supply of judgment, which then requires still more automation to cover the gap.
Luis Garicano and Luis Rayo's "Training in the Age of AI: A Theory of Career Viability" adds the sharpest detail: a junior has historically paid for training by producing low-value work a client tolerated from a novice, and once AI does that work more cheaply, the arrangement stops financing itself, not because the junior can no longer learn, but because nobody is still paying for the learning to happen. It's the business model underneath apprenticeship that breaks first.
The clearest way to see what's at stake is an accounting identity: next year's stock of expertise equals this year's, minus whoever retires or forgets, plus whatever regenerates it. AI makes the stock dramatically more productive right now, which is exactly why a field can look unusually strong while the regeneration term quietly runs toward zero.
Deep Learning, Deep Unlearning
Zoom out and the timing stops looking like a coincidence. The same years that produced Deep Learning's breakthroughs, the scale, the compute, the benchmarks cleared one after another, are producing something like its mirror image in humans. Not people getting less intelligent. People exercising, less and less, the exact processes that used to produce some of their capabilities, because the result of those processes can now be gotten another way. As the machines went deep into learning, it became easier for everyone else to stay on the surface.
The same mistake keeps recurring: treating work as a list of tasks to get through, friction to eliminate, when it's also, and first, something interior. Software delivery has the identical blind spot: collapsing the friction of shipping code doesn't just remove waste, it also removes the time design intuition needs to mature, and the moments a team would have caught, by actually using its own product, that something shouldn't ship at all. Apprenticeship runs on the same logic. A Benedictine monk finds real liberation in humble, repetitive labor, and the reason isn't the repetition: work that actually forms someone runs through four stages, setting a goal, choosing the means, carrying them out, and then contemplating what happened. Junior work built judgment through its fourth stage: standing back to see why something didn't add up, exactly the step AI now skips on the way to a right answer.
So what to do about it?
Built In, Not Bolted On
The dilemma from earlier doesn't get solved by any one actor behaving well, any more than one driver's clean exhaust clears a city's air by itself. It gets solved the way pollution usually does: not by trusting restraint, but by building the constraint into the thing itself. A model that holds an answer back for a moment, that asks for a guess before giving one, that leaves the silence where a person's own reasoning was supposed to happen, does for cognitive capital roughly what emissions control does for a shared atmosphere: containing at the source what would otherwise accumulate, invisibly, at the scale of an entire population. That design would have to work against the exact incentive that keeps people coming back: a model built to please tends toward sycophancy, confirming what someone already believes instead of correcting it, which doesn't just fail to teach, it flatters someone into staying wrong. Nobody enjoys being contradicted, or refused an answer they came for, which is exactly why this has to be a deliberate design choice rather than something the market selects against on its own.
Not every model needs the same constraint. A frontier system built to push open problems at the edge of a field is closer to a race car than a commuter's: its users are few, and it isn't the vehicle filling anyone's daily air. The models carrying the actual load, the ones a billion students, engineers, and analysts run through every day, are the city cars. That's where the filter has to sit, because that's where the exhaust accumulates, one skipped step of reasoning at a time, into a stock nobody owns and everybody eventually depends on.
It's the same equation as before, only global. A resource gets consumed today by the very process that was supposed to be regenerating it, and the two run on different clocks: the withdrawal now, the shortfall a decade out. The Fields Medallists were describing exactly this, at the top of the chain, when they warned that the mass production of results could "destroy fertile ground instead of breathing life into new ideas." Fertile ground doesn't recover by itself once it's gone.
The Oak at Flagey
In 1864, Gustave Courbet painted an oak near his family's farm in Flagey. Three years later, exhibiting the painting, he added a subtitle: "called the Oak of Vercingetorix, Caesar's camp near Alesia." Nobody agreed, then or now, on where that battle—Vercingetorix's last stand, the end of Gallic resistance—actually took place. Napoleon III had already backed one site with his own excavations. Courbet, painting from his home region, was staking a claim for another, and the oak was what he could point to: old enough to have stood there when it mattered, solid enough to outlast an argument about where "there" even was.
In the painting, the tree fills the whole canvas and stands alone on an otherwise bare plain. But an oak that size was never raised alone. It takes a crowd of younger, shorter trees sharing the same ground for decades, most of which never reach the canopy, before a single one ends up tall enough to look like it did this by itself. The bare plain doesn't hide the trunk. It hides how deep, and how crowded, the ground beneath it had to stay for that long, before one tree could end up standing like this.
The oak was struck by lightning sometime in the 1920s. It no longer exists. Nothing in the painting said it wouldn't.
Deep Learning taught machines how to learn. Now we may have to relearn how to learn deeply.
