The last generation of experts

Everything valuable I know how to do, I learned by doing things that had, taken in isolation, almost no value. My first years as a mobile developer were filled with glue code, display bugs, crashes to reproduce patiently — the routine work a coding agent now dispatches in minutes. Yet it was in that thankless material that my architect's intuition took shape, along with my nose for regressions and my ability to sense that an elegant solution is hiding a problem. Every expert I know tells the same story. And it is exactly that story that is becoming impossible to live.

The bridge is giving way

The 2026 numbers no longer describe a hypothesis; they describe a fork in the road already being taken. Erik Brynjolfsson's team at Stanford, which tracks payroll records for 4.6 million American workers, finds that employment of 22-to-25-year-olds in occupations highly exposed to AI now sits 19% below its expected trajectory — a gap that stood at 15% a year earlier and is widening by half a point per month, while their experienced elders show no comparable decline^[1]. In tech, the venture fund SignalFire estimates entry-level hiring down about 65% at the sector's twelve largest employers compared with 2019, and down 76% at startups; a graduate of the top computer-science programs is 45% less likely to land a job there than in 2022^[2]. The World Economic Forum, which surveyed 9,400 entry-level workers across 48 countries, puts it soberly: entry-level roles “have long been the bridge between education and employment,” and that bridge is shifting^[3]. Shifting is an understatement: it is giving way.

The cause is no mystery. More than a quarter of Google's new code was AI-generated by late 2024, more than 30% six months later^[4]; Microsoft reports 20 to 30%^[5]; at Anthropic, the chief financial officer cites more than 90%^[6]. Dario Amodei was warning as early as May 2025 that AI could eliminate half of all entry-level white-collar jobs within a few years^[7]. One can quibble with the numbers, all self-reported and sometimes self-serving. One can no longer quibble with the direction.

AI absorbs the formative work first

“AI absorbs tasks before it absorbs jobs,” as Brynjolfsson sums it up^[1]. But look at which tasks. Andrej Karpathy, reflecting on his own sessions with coding agents, observes that they excel at anything resembling the millions of examples seen on the Internet — boilerplate, familiar patterns — and go off the rails on novel, intellectually dense code^[8]. In other words: the machine succeeds at precisely the work on which we used to grow our juniors, and stumbles on the work that demands the expertise those juniors will no longer build.

This trap was not born with language models. As early as 1983, in a paper that became the founding text of forty years of human-factors research, Lisanne Bainbridge described the “ironies of automation”: by relieving the operator of routine tasks, you deprive him of the very practice that maintains the skills he will need on the day the automation fails^[9]. Forty years later, the sociologist Matt Beane documented the same mechanism in operating rooms: with robotic surgery, the resident — once hands-deep in the patient alongside the senior surgeon — becomes “completely optional,” a spectator to an expert more productive than ever. Only one learner in eight manages to train anyway, through workarounds Beane calls “shadow learning”^[10]. What agentic AI is doing to software engineering, robotics had already done to surgical apprenticeship: it optimizes today's production while silently divesting from tomorrow's competence.

The illusion of competence

The objection writes itself: young graduates will learn differently, with and through AI. Perhaps — but the first rigorous measurements counsel caution, because they reveal something more troubling than slowness: illusion. In a randomized trial run by METR with seasoned open-source developers, participants equipped with AI tools turned out to be 19% slower on their real tasks — while remaining convinced, afterward, that they had saved 20% of their time^[11]. Among novices, an award-winning study from the ICER conference shows that struggling students who code with generative AI finish with “an illusion of competence,” persuaded they did better than they actually did^[12]. And a survey of knowledge workers by Microsoft Research and Carnegie Mellon establishes that the greater the confidence in AI, the less critical thinking gets exercised^[13].

In all fairness: Anthropic's usage data notes that heavy delegators report learning as much as everyone else^[14]. But that is self-reporting — precisely the measure METR and ICER have just shown to be the least reliable there is. When you feel faster while being slower, and more competent while being less so, the thermometer of felt experience is broken. That is in fact the heart of the problem: the erosion of expertise is painless for the person undergoing it.

Who will sign?

Yet our entire edifice of trust rests on expert verification. The quarter of Google's code that AI writes is, Sundar Pichai specifies in the same sentence, “then reviewed and accepted by engineers”^[4] — human review remains the mandatory checkpoint. Simon Willison, one of the most closely followed voices on AI-assisted programming, says it plainly: AI tools “amplify existing expertise,” and engineers must remain “proudly accountable for the software they produce”^[15]. But amplifying zero yields zero. Ethan Mollick, a professor at Wharton, asked the awkward question in a line I never tire of quoting: “every time we hand work to a wizard, we lose a chance to develop our own expertise — to build the very judgment we need to evaluate the wizard's work”^[16].

The psychology of expertise has told us, ever since Ericsson, that expert judgment is born of years of deliberate practice — thousands of hours of confrontation with reality under immediate feedback^[17]. You do not acquire it in a classroom, or by watching, or by approving pull requests you could not have written yourself. A graduate propelled into “agent manager” without years of practice is not an accelerated expert: he is an approver. And in the industries my clients inhabit — energy, transport, healthcare, defense — the question is not rhetorical. Embedded software gets certified; a structural calculation carries a commitment; someone, at the end, signs. The signature of an approver who cannot retrace the reasoning protects no one: it merely shifts the risk to the first incident.

What if the models become perfect?

Let me raise the objection I raise against myself: models are improving fast, and human verification of code may well become as anecdotal as the manual verification of compilers. Granted. Two things will nonetheless survive that generous hypothesis. First, accountability, which is a social and legal construct, not a technical performance: our clients, our regulators and our courts will long demand that a natural or legal person answer for what is delivered, and answering for what you do not understand is not a tenable position. Second, arbitration: even served by perfect agents, someone will have to decide what to build, to what standard, with which trade-offs — a judgment that is itself formed by practice. Verification will change altitude; it will not disappear. It will climb from code to architecture, from architecture to intent. But those upper floors still need people capable of occupying them — and there is no elevator.

Organizing the succession rather than enduring it

I do not believe in turning back: depriving teams of agents to preserve pedagogy would be as futile as training accountants on the abacus. But I do believe we must now treat the formation of future experts as infrastructure, with a budget and requirements, because the market will no longer produce it for free as a by-product of work. Concretely: hire juniors anyway, and accept that their immediate profitability is no longer the point; deliberately set aside a share of the formative work — a production incident analyzed by hand, a module written without assistance, a paired review with a senior — the way airlines reserve simulator time for pilots who nonetheless fly on autopilot; make verification a discipline that is taught, tooled and assessed, not a vague “human in the loop”; and finally, reinvent apprenticeship around the agent, senior and junior working side by side on what the machine produces, as Beane observes among that one-in-eight of learners who always find a way.

The title of this paper is an alternative, not a prophecy. My generation may be the last to have been formed by the old world — by thankless code, demanding peers and late-night production releases. If we do nothing, it will be the last generation of experts, full stop, and around 2035 we will discover what an industry is worth when no one left can tell whether the work is any good. If we organize the handover, it will merely be the last generation to have been formed this way — and the first to have formed the next one differently, deliberately. The window for choosing is open. It is closing at half a point per month.