Korea's AI-Exposed Youth Jobs Fell 13%, Starting Before ChatGPT

Between 2022 and 2025, the number of Koreans under 30 working in the jobs most exposed to generative AI fell from 1,197,000 to 1,037,000. That is a drop of 13.4% in three years. The same group had already lost 54,000 workers in the single year before ChatGPT launched.

That second number is the one to keep in mind. It comes from a set of five studies that the Korea Employment Information Service (KEIS), a research body under Korea's Ministry of Employment and Labor, published on August 31, 2026 in the summer issue of its quarterly Employment Issues (고용이슈). The ministry announced the issue with a press release whose headline calls AI "not a job destroyer but a partner in reshaping work." [Fact]

The headline is a verdict. The studies underneath it are more careful than that, and in one place they point the other way.

What was actually measured

The five papers use different Korean datasets, and that matters, because each answers a slightly different question. [Fact]

• Yoon Jung-hye matched a Korean-built occupational AI exposure index to employment insurance administrative records, which cover people entering and leaving insured jobs.
• Kim Su-hyun and Lee Jung-a took the "observed exposure" measure published in 2026 by Anthropic economists Massenkoff and McCrory, which scores US occupations by how much Claude is actually used for their tasks at work, and mapped it onto Korea's occupational classification. They then joined it to the regional employment survey run by Korea's national statistics office for the second half of each year from 2021 to 2025.
• Jung Soon-ki reported a survey of 2,297 establishments across 21 industries.
• Kim Dong-kyu surveyed 305 employees in IT development, visual and digital design, and general office work.
• Kong Jung-seung used the 2025 youth panel survey to compare young workers who use generative AI with those who do not.

A disclosure before going further. The exposure measure in the second paper comes from Anthropic, and this site's drafts are written with Anthropic's model. We have tried to report that paper's weak spots as plainly as its strengths.

The whole-economy picture: no collapse

For the workforce as a whole, the Kim and Lee paper finds nothing that looks like displacement. Workers in the top quarter of exposed occupations numbered 7,196,000 in the second half of 2022 and 7,377,000 in 2025, an index of 102.5 against 2022. Workers in occupations with zero observed exposure went from 8,996,000 to 9,042,000, an index of 100.5. [Fact] The high-exposure group grew faster, not slower.

Exposure itself is thin in Korea. In 2022 the average worker's occupation had an observed exposure of 8.0%, the median was 2.0%, and about 30% of workers were in occupations scored at zero. [Fact] Of 490 detailed Korean occupations, 147 scored zero, including cooks, construction trades, pilots and firefighters.

The top of the list will look familiar to anyone who follows US exposure research: product planners, travel product developers and advertising and PR specialists at 64.8%, application software developers at 57.4%, data analysts at 45.5%, information security specialists at 45.0%, customer service agents at 43.4%, and translators and interpreters at 43.0%. [Fact] If you work in one of those, our pages for software developers, customer service representatives, translators and interpreters, data scientists, information security analysts, advertising and promotions managers and public relations specialists carry the occupation-level detail.

Exposure also tracks pay. Wage earners in the high-exposure group averaged 3.658 million won a month in 2025, against 2.706 million won for the zero-exposure group, about 35% more. [Fact] Women were more exposed than men (30.4% versus 21.4% in high-exposure jobs), and so were university graduates (35.7%) compared with high school graduates (20.3%).

The youth numbers, read twice

Now the under-30 table, which is where the press release's reassurance gets thin.

High-exposure employment among under-30s: 1,251,000 in 2021, 1,197,000 in 2022, then 1,169,000, 1,093,000 and 1,037,000. Zero-exposure employment for the same age group: 909,000, 915,000, then 895,000, 882,000 and 846,000. [Fact] Indexed to 2022, that is 86.6 against 92.5, a gap of 5.9 points.

The authors' reading is that the high-exposure decline started before ChatGPT, so it cannot be pinned on generative AI, and that the decline actually eased after launch. The first half of that is plainly in the data. The second half depends on how you measure it.

We ran the rates from their table. In the one pre-ChatGPT year, high-exposure youth employment fell 4.3% while zero-exposure youth employment rose 0.7%, a gap of about 5 points. Over 2022 to 2025, the high-exposure group fell about 4.7% a year and the zero-exposure group about 2.6% a year, a gap of about 2.1 points a year. [Estimate]

So two statements are true at once. The high-exposure group's own decline did not slow, measured as an annual rate. What narrowed was its gap with the comparison group, because young workers in zero-exposure jobs started shrinking too. Whether you call that "easing" depends on which of those you think measures AI. The paper's own conclusion is more guarded than its press release: it says the youth pattern deserves attention and that a sustained slowdown in young people entering high-exposure jobs could damage their skill formation over time. [Claim]

The age split is sharper still. Taking the high-exposure index minus the zero-exposure index for 2025, the gap was minus 5.1 for people in their 20s and minus 2.3 for their 30s, but plus 3.2 for their 40s, plus 15.6 for their 50s and plus 22.0 for those 60 and older. [Fact] Exposed work in Korea is not disappearing. It is getting older.

"Youth population decline" is half the story

The press release's second headline says the slowdown owes more to a shrinking youth population than to AI. That claim comes from Yoon's paper, which splits the growth rate of insured employees into population, labour-force participation, employment rate and insurance coverage.

For workers aged 15 to 29, insured employment fell 5.0% in 2024 and 4.6% in 2025. Population contributed minus 2.8 and minus 2.4 points. Participation contributed minus 0.9 and minus 2.2. [Fact] By our arithmetic, population accounts for a bit over half of the 2025 drop, and population plus falling participation accounts for all of it. [Estimate]

The paper itself frames this as supply-side factors, meaning population and participation together, not population alone. And there is a limit the decomposition cannot get around: it is an accounting identity with no term for AI. If weaker hiring discourages young people from looking for work, that shows up as falling participation and gets filed under "supply." So the table shows that AI is not needed to explain the decline. It does not show that AI played no part in it.

The same paper flags one occupational signal that deserves a name. Insured employment among accounting, tax and appraisal professionals and among accounting and bookkeeping clerks has been falling steadily, which the author reads as possible adjustment in structured quantitative work. [Claim] Our pages for accountants and bookkeeping clerks are the ones to watch here.

Inside firms: early, and mostly generative

Jung's establishment survey puts AI adoption at 28.6%, or 656 of 2,297 firms. Leave out plain automation systems and count only predictive and generative AI, and it falls to 23.7%. [Fact] Information and communications led at 62.9%, followed by precision instruments (41.3%) and food manufacturing (39.3%). Transport was lowest at 11.4%.

Among the 656 adopters, 75.6% used generative AI, 27.6% automation systems and 13.3% predictive AI. [Fact] The paper describes firms' main reasons as productivity and efficiency, and says they see AI more as a way to cut repetitive work than to cut staff. That is a qualitative finding. The survey does not attach a percentage to it.

A short correction, since some summaries have merged these figures: the 75.6% is the share of adopters using generative AI. It is not a share of firms saying they use AI for productivity rather than headcount cuts.

On hiring, firms in machinery, R&D, precision instruments and ICT expected demand for both AI staff and new hires to rise. Firms in business support services, health and social work, and culture and arts expected both to fall. [Claim]

What workers say

Kim Dong-kyu's 305 respondents rated how much AI had improved five capabilities on a five-point scale. Efficiency scored 3.40, problem-solving 3.15, productivity 3.14, creativity 2.79 and collaboration 2.37, for an average of 2.97. [Fact] (The press release rounds collaboration to 2.3; the paper reports 2.37.)

The paper's warning is about the bottom rung. When the routine tasks juniors learned on are automated, the "skill ladder" breaks, and it calls for on-the-job training programs built around onboarding people into AI-assisted work. It also finds that adoption is mostly bottom-up, driven by individual employees, while HR and pay systems have not caught up. [Claim]

The limits worth keeping

The Kim and Lee paper has only two pre-ChatGPT years, so pre-existing trends are hard to pin down. Mapping a US measure through the international classification into Korea's lowered the absolute exposure levels. And because exposed jobs cluster in growing industries, industry growth may be hiding displacement. The authors say their results should be read as relative employment changes, not causal effects of AI. Each paper also carries the standard KEIS note that it reflects the authors' views, not the institution's. [Fact]

That is a more honest position than the headline. For a young Korean looking at a planning, software, data or customer-facing career, the data does not say the door is closing. It does say that the door has been narrowing for people your age for at least four years, and that the jobs are still there but held by older workers. The practical response is the one the papers point to: get onto the parts of the work that AI does not yet do well, like collaboration and judgement, which also scored lowest in the worker survey. For more on how young workers fare under observed exposure in the US data, see our write-up of Anthropic's observed exposure study.

Sources

• Ministry of Employment and Labor (2026). Press release, 31 August 2026: 「"인공지능(AI)는 일자리 파괴자 아닌 직무 재구성의 파트너… '인간 중심 인공지능(AI) 발전전략' 세워야" (한국고용정보원, 『고용이슈』 여름호 발간)」. https://www.moel.go.kr/news/enews/report/enewsView.do?news_seq=19850
• Korea Employment Information Service (2026). 『고용이슈』 2026년 여름호 (quarterly Employment Issues, Summer 2026), special section 「AI가 초래하는 노동시장의 변화」 — papers by Yoon Jung-hye; Kim Su-hyun and Lee Jung-a; Jung Soon-ki; Kim Dong-kyu; Kong Jung-seung; issue letter by Kwon Woo-hyun. https://www.keis.or.kr/keis/ko/proj/117/pblc/detail.do?categoryIdx=130&pubIdx=11364

Figures were read from the press release and the paper files attached to it. The annual rates, the pre- and post-2022 gaps, and the population share of the youth decline are this site's arithmetic on the published tables.

This article was produced with AI assistance and reviewed before publication. Data tags: [Fact] = reported in the source; [Claim] = the authors' interpretation; [Estimate] = this site's calculation on source figures.