AI is deleting the first rung, not the senior job

This morning we ran the Bank of Korea's headline count. The note underneath it is the story: the same industries shedding juniors are still adding people in their 50s.

A second print, not a first scare

On Tuesday the central bank's employment team published Issue Note 2026-19. The raw count is the one we already flagged in Korea lost 285,000 youth jobs in AI-exposed sectors: 285,000 fewer jobs for people aged 15 to 29 between June 2022 and June 2026, and 268,000 of those — 94 percent — sat in high-AI-exposure industries. Information services dropped 31.4 percent; publishing, which includes software and web design, dropped 27.4 percent; programming and systems integration fell 16.6 percent; professional services fell 11.6 percent.

This is not the first print. In late 2025, Jinsu Han and Samil Oh used the same National Pension Service file — about 16 million regular workers — and found 211,000 youth jobs gone over three years, 98.6 percent of them in high-exposure sectors. Twelve more months added roughly 74,000 losses. Workers in their 50s kept gaining: 209,000 then, 230,000 now, and 173,000 of the latest gain still sits inside those same industries. The bank's phrase, carried forward from 2025, is seniority-biased technical change. The update's job was to check whether the bias faded. It did not.

The authors set mid-2022 as the start date because that is when ChatGPT shipped. That is a research choice, not a proof of causation, and they say so.

What the machine actually replaces

Oh Sam-il, who heads the employment team, put the mechanism in one line: AI raises a young worker's productivity a great deal, which is another way of saying it can do the job. Junior work is codified — first drafts, tickets with a playbook, the assignment a textbook already described. Senior work is the organization's context. Models are trained on text. Text is what juniors produce.

The note's most useful cut is not age. It is how the firm uses the tool. Sectors that hand the task itself to the model — automation, in the bank's language — showed the steep youth drop. Sectors that keep the human on the job and use the model for drafts and checks — augmentation — did not. "We adopted AI" is a useless statistic. Automation versus augmentation is the one that predicts who still has a first assignment.

That split now has an independent twin. On August 12, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at the Stanford Digital Economy Lab revised their ADP payroll study, Canaries in the Coal Mine. They still find no economy-wide displacement. They do find that employment among 22-to-25-year-olds in the most AI-exposed occupations now sits about 19 percent below where it would be if those workers had kept pace with same-age peers in less-exposed jobs — up from 15 percent in their July 2025 vintage. In levels, the two most-exposed quintiles are down about 11 percent since November 2022; the less-exposed quintiles are up about 10 percent. The adjustment is mostly a hiring freeze, not a layoff wave, and it concentrates where AI use automates tasks rather than complements them.

Two datasets, two continents, one diagram: the model eats the first assignment. It leaves the person who already knows how the company actually works.

Korea's education split is the political version of that diagram. From January 2019 through October 2022, college-educated youth and everyone else were essentially tied on unemployment — 8.2 percent versus 8.0 percent. After November 2022 the college rate sat at 7.0 percent, 1.6 points above the 5.4 percent for people with an associate degree or less. The bank used schooling as a proxy because graduates cluster in AI-exposed white-collar work. A four-year degree used to be the on-ramp. In the sectors that degree was built for, it is now a marker of exposure.

The ladder is rusting from both ends. Comparing 2016–2019 with July 2022–June 2026, monthly youth outflows from high-exposure industries rose about 32 percent, from 3,700 to 4,900. Inflows fell about 11 percent, from 32,600 to 29,100. Fewer people get in. More of the ones who do get in leave. That is not a one-time hiring pause. It is a thinner apprenticeship.

Who keeps the job, and who never gets one

The winners, for now, are workers in their 50s who already hold tacit knowledge, and the firms that can swap a junior headcount for a subscription. We have already watched the US version of that substitution at the task level — One in five US workers now hands tasks to AI, not colleagues. Korea's note is what that looks like when a central bank can see the payroll file.

The losers are not "workers" as a class. They are 24-year-olds who did what the system asked — university, then a programming or professional-services first job — and found the first job is the one the model does cheapest. China's shock absorber has been the gig platforms: we covered that last week in China's jobs squeeze: 53M now deliver or drive. Korea does not have that cushion. It has a shrinking youth cohort and too many graduates chasing too few large firms.

The longer-horizon loser is the firm that thinks it can keep harvesting seniors without growing any. Entry-level work was never just output. It was how tacit knowledge got made. If the 2036 version of today's 50-year-old never spent two years on the tickets the model now drafts, the experience premium becomes a depleting stock. The authors say young people could still be the long-run winners if they adapt faster and if productivity creates new demand. That is a hope. The finding is that the current use pattern does not train anyone.

The case that this is not about the models

Take the skeptics seriously. South Korea's youth cohort is shrinking, so a falling youth population will show up as falling youth employment even if every firm is a saint. Post-pandemic hiring pullbacks, a shift toward experienced recruits, thinner internal training, and remote work all arrived in the same window, and remote-capable jobs overlap heavily with AI-exposed ones. The bank says this out loud: do not read the slump as a pure model effect. Read it as a tool speeding up a career-ladder weakening that was already underway. Degree inflation and a bottleneck at the large firms were the official diagnosis years before ChatGPT. If you already had a rusted ladder, a tool that does the first-year workbook will make the rust look like a robot.

That critique does not erase the sectoral concentration. A purely demographic story should scatter losses. These losses pile up in information services, publishing, programming, and professional services — the same industries where the Stanford payroll file shows the hiring freeze. Same technology, two uses, two youth-employment paths.

Policy, in the BOK's view, should stop trying to preserve last decade's internships. Build apprenticeships that mix real work, model use, and a mentor in the room. Subsidize the training and the senior's time, not the headcount the software already replaced. A tax on firms that cut jobs after adopting AI would be easier to pass, and would reward the company that never hired the junior in the first place. The thing that actually rebuilds tacit knowledge is a first assignment a human still has to finish — not a preserved title on an org chart.

What to watch: whether Korean firms that treat the model as a junior-replacement tool start posting different hiring numbers than firms that treat it as a draft assistant. The BOK split predicts they will. So does Stanford. If both keep being right, the missing first job is no longer a Korean curiosity. It is the default.

If the first job is how seniors get made, should firms that automate it have to fund the apprenticeship that replaces it? Tell us in the comments.

Sources: Bank of Korea — Issue Note 2026-19 · The Korea Times · The Herald Business · Stanford Digital Economy Lab · Bank of Korea — 2025 blog