The Broken Ladder: AI Is Doing the Apprentices' Work — So Who Is Building the Masters of 2040?

THE THESIS. AI is not destroying expert work: it is destroying the path toward it. Today's systems are good at precisely the tasks professions used to train their novices on — reading normal X-rays, reviewing standard contracts, fixing trivial bugs. That rung is disappearing, and the data is starting to confirm it. Our thesis is twofold. First: the expertise of 2040 is not doomed, because routine work was never the magic ingredient of learning; it was merely the cheapest way to get it. Second: nobody will get that expertise for free anymore. The market, on its own, will not pay to train juniors. Training will have to be designed — with simulators, explicit mentorship, and new apprenticeship contracts — just as aviation redesigned its own after discovering that the autopilot was atrophying its pilots.

THE NUMBER THAT MATTERS. Two recent studies measure the phenomenon with real data, not surveys. The Stanford study (Brynjolfsson, Chandar and Chen), built on payroll records from ADP — America's largest payroll processor, covering some 25 million employees — finds that employment for workers aged 22 to 25 in the occupations most exposed to AI has fallen 13% since late 2022. In entry-level software and customer service roles, the drop approaches 20% through mid-2025. Meanwhile, workers over 30 in those same fields grew between 6% and 12%. The Harvard study (Hosseini and Lichtinger), using résumés and job postings from 285,000 firms, supplies the mechanism: at companies that adopted generative AI, junior employment fell roughly 9% within six quarters relative to non-adopters — driven by slower hiring, not layoffs — while senior employment barely moved. They call it "seniority-biased technological change": the technology doesn't replace a profession, it replaces a rung. One necessary caveat: a May 2026 London School of Economics study argues that remote work predicts the entry-level decline better than AI does, and the unemployment rate for recent US graduates (5.6%, per figures reported by NPR) has several causes at once. But the pattern repeats across every source: the base of the ladder is narrowing while the top holds.

WAS ROUTINE WORK ACTUALLY ESSENTIAL? Here it pays to return to the science of expertise. Anders Ericsson, the psychologist whose work inspired the famous 10,000-hour rule, demonstrated something subtler than the rule itself: mastery doesn't come from repetition, but from "deliberate practice" — exercises at the edge of your ability, with immediate feedback and rising difficulty. Mechanical repetition, he insisted, teaches almost nothing. Seen that way, the junior's routine task was never the essential ingredient: it was the vehicle. It was the economic excuse that let a company pay a salary while the learning came along for free. But beware the easy conclusion that nothing is lost. That boring volume hid two valuable things: practice with real consequences, and massive exposure to cases. The resident who reads thousands of normal X-rays learns what normal looks like — a precondition for recognizing the abnormal. The programmer fixing trivial bugs absorbs how a real system is built. If AI removes the rung, both things must be replaced, not just one.

WHEN THE LADDER WAS REBUILT — AND WHEN IT WASN'T. History offers both endings. Aviation lived through this first: the autopilot flew so well that pilots stopped practicing. The 2009 crash of Air France Flight 447 — 228 dead after a perfectly recoverable aircraft that the crew could not recover by hand — turned "skill fade through disuse" into official doctrine. The answer was not to remove the automation but to redesign the training: regulators mandated manual flying hours and simulator-based stall-recovery training. The ladder was rebuilt because someone paid to rebuild it. Calculators tell the gentle version: manual arithmetic turned out to be scaffolding, not expertise, and engineers did not get worse when they let it go. Robotic surgery tells the dark one. Researcher Matt Beane (UC Santa Barbara) spent years observing operating rooms and documented that with the surgical robot, the attending surgeon operates alone while the resident watches: the apprentice's role evaporated. Only rule-breakers learned — what Beane calls "shadow learning": specializing too early, rehearsing covertly on simulators and videos, practicing without supervision. The result: a generation of surgeons with fewer real chances to learn. The difference between these cases is not the technology. It is whether an institution absorbed the cost of redesigning the learning — or nobody did.

OUR READ. The market will not fix this on its own, and it's worth saying plainly. Training a junior is an investment whose return is usually captured by someone else: the trained junior leaves; the AI doesn't. For each individual firm, replacing apprentices with models is rational; for the economy as a whole, it is sawing off the branch. It is the classic commons problem, like basic research. And its danger is that it produces no visible collapse — only an invisible deficit: today's seniors are still at their desks and everything seems fine; the bill arrives in 2035-2040, when they need replacing. That said, we reject fatalism — not on faith, but on logic: the same AI that breaks the ladder is the best material ever invented for rebuilding it. An infinitely patient tutor, available around the clock, generating tailored hard cases and giving instant feedback is, almost point for point, what Ericsson described as the ideal conditions for deliberate practice. The thing that was always scarce and expensive — one-on-one expert attention — is getting cheaper by orders of magnitude. The bottleneck is no longer pedagogical. It is one of incentives.

WHAT COMES NEXT. The first experiments point in the right direction. Bank of America is keeping close to 4,000 interns in 2026 while building AI simulations designed to compress years of on-the-job learning into months, according to industry reporting. Other firms are inverting the workflow: the junior solves the problem by hand first, then compares their version against the AI's, forging judgment before dependence; or the AI produces and the junior interrogates and verifies the output with a senior acting as coach. There is also talk of "paid residencies" on the medical model, with structured training replacing incidental learning. These are pilots, not proven solutions, but they share the right design: making explicit and deliberate what used to be implicit and free. Implications. For those starting out: the entry ticket is no longer knowing how to do the task, but knowing how to audit the machine that does it; seek employers who invest in training you — they are the ones thinking about 2035. For companies: protecting the talent pipeline is strategy, not charity; whoever trains juniors with AI will have the seniors their competitors cannot hire. For regulators: aviation proved that mandating training works without slowing the technology. Long term, our usual optimism, duly qualified: the ladder is not disappearing, it is being redesigned — and in the abundance horizon we defend, mastery will increasingly be pursued out of passion, not just payroll. The question for 2040 is not whether there will be masters. It is whether we start building them now, on purpose, or wait until their absence shows. By then it will be too late.