Where Did the Entry-Level Job Go?

In the payroll records of millions of American workers, researchers at Stanford found a pattern hiding in plain sight. Since late 2022 — the month generative AI reached the public — employment for workers aged 22 to 25 in the occupations most exposed to AI, such as software development and customer service, has fallen about 13 percent relative to other groups. Their older colleagues, in the same occupations at the same firms, held steady or grew [1]. The researchers called these young workers "canaries in the coal mine." The study is a working paper, not yet peer-reviewed, and we will treat it that way. But it put numbers to something families had been sensing for two years.

The official statistics rhyme with it. The Federal Reserve Bank of New York's tracker put unemployment among recent college graduates at 5.6 percent in early 2026 — up from 3.6 percent in March 2019 — with 41.5 percent of recent graduates underemployed, working jobs that do not require their degree [2]. For most of the history of such data, a fresh degree meant doing better than the labor market average. That reliable head start has, for now, eroded.

The question is not whether machines will take all the jobs. That question has been asked for two hundred years, and the record of confident answers is brutal in both directions. The investigative question is narrower and stranger: what happens to the bottom rung of the career ladder — the jobs we learn on — when machines learn to do precisely those?

Start with the most famous machine-breakers in history. The Luddites of 1811 were not ignorant technophobes — they were skilled textile workers correctly predicting that mechanized looms would destroy their livelihoods. They were right about their own lives and wrong about the arc: employment grew enormously in the industrial economy that followed. That double truth — real, concentrated pain; broad, delayed gains — is the recurring signature of automation [3].

The twentieth century repeated it at scale. In 1900, about four in ten American workers farmed; by 2000, fewer than two in a hundred did — a near-total automation of the largest occupation that ever existed, absorbed without permanent mass unemployment [3]. The economist David Autor's summary of two hundred years of this history: automation removes tasks, but it also raises productivity and incomes, which creates demand for new work — and commentators reliably overstate the substitution while missing the complements [3]. The canonical modern case is almost funny: after ATMs spread, the number of human bank tellers rose for decades — cheaper branches meant more branches, and the teller's job shifted from counting cash to handling relationships [4].

So what? History's pattern is specific: the bottom rung of the ladder has repeatedly moved — from field to factory floor to office cubicle — but there was always somewhere for a beginner to stand, because machines took the muscle work and the rote work while humans climbed toward judgment. The unsettling novelty of the present moment is that "rote cognitive work performed by a beginner" is exactly what the new machines do best. The historical escape route ran through the very rung now in question.

The measured facts: the New York Fed's series on recent graduates is official, long-running data, and its recent deterioration is real [2]. The Stanford payroll study is the largest-scale early evidence — millions of ADP payroll records — and its core facts are carefully drawn: the declines concentrate among the youngest workers, in the most AI-exposed occupations, and specifically where AI automates work rather than augments it [1]. Emerging evidence It remains a working paper: unreplicated, one data source, and its authors have themselves published follow-up analyses testing whether interest rates or tech-sector timing could explain the pattern instead [1].

A memory from the last hype cycle belongs here. In 2013, a widely cited Oxford study estimated 47 percent of US employment was at risk of automation [7]; an OECD team re-examined the same question task-by-task and got about 9 percent [8]. Same data, same decade, a five-fold gap — because "a job contains automatable tasks" is not the same as "a job disappears." Every number in this investigation should be read with that gap in mind.

Competing hypothesis Four explanations currently compete for the graduate numbers, each with credible backing, and none yet eliminated. AI substitution: the Stanford pattern — youngest workers, most-exposed occupations, automation-flavored uses — is what AI displacement would look like in its first years [1]. Competing hypothesis Remote work: New York Fed economists estimate that distributed teams — where teaching a novice is hardest — may account for much of the rise in youth unemployment specifically [2][5]. A boring explanation with real evidence behind it. Competing hypothesis Interest rates and the tech correction: the sectors that hire young graduates binged on hiring through 2021 and retrenched when money got expensive; some of the "AI effect" may be an echo of that cycle — a possibility the Stanford authors engage directly rather than dismiss [1]. Competing hypothesis A mixed, cyclical story: labor economists at EPI caution that the graduate picture is "more mixed than the headlines suggest" — weaker, yes, but within the range of past soft markets for new entrants [6].

So what? The honest description of 2026 is not "AI is destroying entry-level work." It is: the youngest workers in the most exposed jobs are measurably losing ground, the pattern matches AI substitution better than it matches nothing, and the alternative explanations have not yet been ruled out. These explanations have very different futures attached. If it is rates or cycle, the first rung grows back with the next expansion. If it is remote work, it grows back where offices do. If it is AI substitution, it does not grow back — and the 2026 cohort is the leading edge of a structural break.

The entry-level job is not an isolated labor-market variable — it sits at the intersection of hiring economics, higher education, and the interest-rate cycle, which is exactly why four different explanations remain alive at once [1][2][5][6]. Corporate hiring budgets that expanded during a cheap-money era and contracted when rates rose share the same graduates who are also the first cohort to meet AI at work, which means the labor data cannot yet cleanly separate a monetary-policy effect from a technology effect. Universities built four-year degree programs on the assumption that a diploma reliably converted into a wage premium over non-graduates — a system-level bet now under pressure from both automation and underemployment [2]. And because entry-level roles are also where organizations have historically trained the next generation of institutional experts, this investigation's findings connect directly to a broader question about pipelines: any sector losing its senior specialists to retirement depends on the same entry-level hiring this piece documents as thinning.

The four explanations in the section above haven't been narrowed down yet — and each one, if it turns out to be the dominant force, leads the career ladder somewhere different. None of this is a forecast. It's where today's evidence actually points, depending on which explanation wins out.

The career ladder is civilization's least-examined piece of infrastructure: expertise comes from the first thousand hours of supervised, low-stakes, real work, and societies are only now starting to ask who pays for those hours when a machine will do them for free. The answer will be visible first in hiring data, not in headlines — these are the indicators to track.

• Labor statistics releases — The New York Fed's quarterly graduate series [2] and the Stanford team's public dashboard [1]: does the gap between young workers in AI-exposed and less-exposed occupations widen, stabilize, or close? Whether falling interest rates revive junior tech hiring, which would favor the cyclical explanation.
• Academic replication — Peer review and independent replication of the payroll findings — the single most important upgrade or downgrade to this investigation's evidence base [1].
• Spread and cohort tracking — Whether the pattern spreads beyond the most-exposed occupations into law, finance, and media entry roles — or stays contained. First longitudinal evidence on the 2023–2025 graduate cohorts: scarring, or catch-up?
• Hiring & postings data — Whether new entry-level job categories appear at scale in postings data — the Autor test: past transitions created work no one had named yet [3]. The wage premium for experience: a sustained rise would be the hollow-ladder scenario becoming visible.
• Long-run structural signals — Whether the college wage premium — the anchor assumption of modern education — holds.

Think about your own first real job — the one that actually taught you how to work. Does that job still exist today? And if it does: would a machine be doing it by the time your kids need it? Tell us — the answers to that question, across a few thousand readers, would be a dataset worth having.