Every profession has its bad-first-draft ritual. The memo that comes back bleeding comments. The research summary quietly rewritten at midnight by the person who commissioned it. The pull request returned with forty notes, one of which is just “no.” For as long as offices have existed, the arrangement has been the same: the senior does the interesting part, hands the tedious residue to a junior, the junior does it badly, the senior corrects it, and after a few thousand repetitions the junior stops doing it badly. That loop is not a side effect of professional life. It is the mechanism — the only known way to turn a bright graduate into someone whose judgment you would trust with your name at the bottom of the page.
Which makes it awkward that the tedious residue is precisely what generative AI does best. First-draft research, routine memos, basic analysis, boilerplate code: the tasks most exposed to automation are almost exactly the tasks beginners were given. The first draft was never really the product. The junior was.
No boardroom voted to abolish apprenticeship, and none had to. Each firm faces the same small arithmetic: a model produces an acceptable first draft in seconds for the price of an API call, while a graduate produces a worse one in days, for a salary, and then consumes a senior’s expensive hour learning why it was worse. Automating the rung is individually rational to the point of being compulsory. The interesting question is what all that rationality adds up to.
The numbers are real; the blame is contested
The honest answer is that the evidence is early, contested, and narrower than the headlines suggest. The study that set the tone is a Stanford Digital Economy Lab working paper from November 2025, not yet peer-reviewed. It found that workers aged 22 to 25 in the occupations most exposed to AI experienced a 16 percent relative decline in employment after generative AI spread, controlling for other factors. More experienced workers in the same occupations saw no such decline, and neither did young people in entry-level jobs with low AI exposure.
Official statistics sketch the softer graduate market without explaining it. In the fourth quarter of 2025, the Federal Reserve Bank of New York put unemployment among recent college graduates at 5.6 percent, and underemployment — graduates working jobs that don’t typically require a degree — at 42.5 percent, the worst since the pandemic. On its own that proves little: hiring has cooled broadly, and no single data series can separate AI from an ordinary weak market. Aggregate employment in developed economies remains broadly stable, and assessments to date, including MIT Technology Review’s in May 2026, find limited evidence that AI has moved the headline numbers at all.
The cleanest evidence is occupational. An IZA discussion paper from June 2026, applying event-study and difference-in-differences methods to near-universe U.S. vacancy data, estimates that after ChatGPT’s November 2022 release, junior software-developer postings fell 14 to 15 percent relative to senior ones — a drop larger than in comparable technical fields and absent in mechanical engineering. Two details matter more than the headline. Employers did not rename the jobs; they raised the experience demanded inside the same titles. And the junior postings that survived shifted toward problem-solving, communication and attention to detail — not AI skills.
Then there is the strongest dose of doubt. In a 2026 policy briefing from CAGE, a research centre at the University of Warwick, the economists Lambert and Schindler draw on 243 million employer–employee matches and 407 million job postings across the U.S., U.K., Canada and Australia between 2017 and 2025. They confirm the pattern — the junior share of new hires fell 8 to 11 percentage points below 2019 levels by 2025 — but complicate the culprit. AI exposure and remote-work exposure are nearly the same variable in their data, with a rank correlation of 0.77 across 683 occupations. Tested separately, a two-standard-deviation increase in either predicts a four-to-five-point fall in junior hiring; tested together, the AI effect shrinks to statistical insignificance while remote work remains significant. It may be the chatbot. It may be the kitchen table. Right now the data cannot cleanly tell them apart.
The work was never the point
Oddly, both suspects break the same loop. Whether the junior was displaced by a model or stranded alone at home, the missing ingredient is identical: nobody close enough to say the draft is bad, and here’s why. Feedback is the curriculum, and it is expensive in a particular way — it costs the senior an hour tonight and pays out over years, mostly to other firms.
Economists have argued for decades that firms underinvest in training, because trained employees walk out the door carrying the investment with them. That is theory, not a measurement from any study cited here — but AI has sharpened it. Junior work was always a bundle: useful output plus slow training, sold as a single transaction, the output quietly subsidizing the training. AI unbundles it. The output now costs almost nothing, which means the training must be bought on its own, as a pure cost with an uncertain and portable return. No budget line exists for that. A recent paper in the journal AI and Ethics describes the result as a “selective erosion of junior roles” across finance, law, marketing, journalism, software and customer service, and argues that the career ladder is best understood as training infrastructure — infrastructure being the thing everyone uses and nobody wants to pay for.
Read that list of surviving skills again: problem-solving, communication, attention to detail. Those are precisely the qualities the deleted tasks existed to produce. Employers are advertising for the output of a training process they have stopped funding. Writing in Quartz, Harvard’s Amy Edmondson and Tomas Chamorro-Premuzic pose the pipeline question bluntly: “if we don’t have people coming in at the entry level, how are we going to develop the pipeline to become managers?” That is expert opinion rather than data, but the arithmetic beneath it is hard to argue with.
None of these studies projects more than a few years out, so the fifteen-year fuse is my arithmetic, not theirs. But demography is patient. The seniors doing today’s reviews were built by someone’s red pen around 2009. The people who should be replacing them in 2040 are, at this moment, not being corrected by anyone. A profession can run for a surprisingly long time on its stock of trained judgment before anyone notices that the factory closed.
Somewhere tonight, a senior will stay late reviewing a first draft. It will be a competent draft, produced in seconds; she will fix the two things that matter and go home. No one will read her corrections. No one will feel the small, useful sting of them, or remember them, or be promoted fifteen years from now on the strength of everything they taught. The work is getting done. The people are not getting made.