Korea’s AI-exposed industries account for 94% of its youth job decline
The first rung of a career ladder is usually built from unremarkable work. A new employee learns the systems, watches how experienced colleagues handle exceptions, and gradually earns responsibility. Automating those early tasks can make a team more efficient this quarter. It can also remove the place where future experts are trained.
A new Bank of Korea issue note gives that tension unusually sharp numbers. From June 2022 to June 2026, youth employment in Korea fell by 285,000. Industries with high AI exposure accounted for 268,000 of the decline, or 94%.
The figure is striking. It is not a count of jobs eliminated by AI, and the researchers explicitly say their findings do not prove causation. The study shows where the damage is concentrated. It also suggests that the purpose of AI adoption, whether automation or augmentation, matters more than a simple yes-or-no measure of access.
A concentrated decline with several possible causes
The researchers used National Pension subscriber records and the Economically Active Population Survey. Between June 2022 and June 2026, youth employment fell 31.4% in information services, 27.4% in publishing, 16.6% in computer programming, and 11.6% in professional services. Employment among people in their 50s kept rising in those same industries.
The issue note describes this as “seniority-biased technological change.” Models can handle codified work that fits into a prompt, checklist, or template. Junior workers often perform that work. Experienced employees are more likely to contribute judgment, relationships, and company-specific knowledge that are harder to reproduce.
AI is not the only change in the period. Pandemic-era over-hiring reversed. Companies increasingly favored experienced recruits. In-house training weakened, while remote work made informal observation and supervision harder. Industry exposure cannot tell us why a particular person lost or kept a job, and the study cannot cleanly separate these forces.
The evidence supports a more careful conclusion than “AI took 268,000 jobs.” It is also more consequential than “nothing has happened.” AI may be speeding up a hiring model that already treated training as somebody else’s expense.
Automation and augmentation produce different signals
The most instructive finding sits behind the headline. Youth employment declined more in industries where AI use leaned toward automation. The researchers found no comparable pattern in industries where AI mainly augmented human work.
Consider a customer-support team. It can use a model to remove junior positions. It can also let junior staff prepare responses while experienced colleagues review difficult cases and explain corrections. A software team can eliminate starter assignments, or use an agent to give new developers faster feedback on richer assignments. Both can claim to have adopted AI. Only the second approach preserves a route for gaining experience.
An earlier Bank of Korea study of productivity found that AI adoption reduced average work time by 3.8%, roughly 1.5 hours a week. Yet the saved time had essentially no relationship with actual output growth. Productivity gains were more visible among self-employed workers, professionals, and intensive users who had greater autonomy or stronger performance incentives. Providing a tool is not the same as redesigning work.
That result weakens the familiar assurance that workers will simply move to higher-value tasks. Remove a low-value task without creating a supervised higher-value one, and a new employee does not advance. The entry point disappears before its replacement exists.
A degree does not supply workplace experience
Since November 2022, unemployment among young Koreans with at least a bachelor’s degree averaged 7.0%. The rate for those without a degree was 5.4%. The two groups had similar unemployment rates before that point, according to the issue note.
Education still matters, but it cannot provide experience inside a particular workplace. The study also attributes the employment decline to two movements: fewer young people were hired, and more young workers left employment. The problem is not confined to landing a first job. It includes staying employed long enough to develop a record and learn what formal education cannot teach.
Korea’s administrative data makes the pattern especially visible, but related warnings are appearing elsewhere. The ILO’s 2026 review of empirical evidence finds limited large-scale displacement so far, along with risks of weaker opportunities for young workers and uneven productivity gains. Its Global Employment Trends for Youth 2026 estimates that 6.1% of jobs held by people aged 15 to 29 are in occupations highly exposed to AI-related change.
Those percentages describe different populations and methods, so they are not directly comparable. Together they identify a problem that an unemployment headline can miss: entry routes may deteriorate before total employment collapses.
Replace obsolete tasks, not the path to competence
Keeping every junior task in its pre-AI form would be wasteful. Some routine work should disappear. Employers still need to decide how a beginner becomes capable when those repetitions are gone.
A workable redesign has four parts:
- Measure progression, not only time saved. Track how many junior employees reach independent responsibility, how long it takes, and where they stall.
- Keep reviewed work in the loop. Let new workers produce real outputs with AI, then require experienced review and an explanation of consequential corrections.
- Fund mentoring as work. Coaching fails when it is invisible labor added to a senior employee’s full workload. Give it time, budget, and recognition.
- Audit automation by career stage. Before removing a task, identify the skill it taught, decide where that skill will now be learned, and define evidence that the replacement works.
Teams testing human-agent work should also borrow a warning from CollabSkill: a score from one interface and task setting is not a hiring score. Aggregate results can improve training. Early unfamiliarity with a tool should not become a permanent label attached to a worker.
For controlled trials, api.ish.chat can route the same assisted task across models, while ish.chat makes it easier to compare how assistants explain and revise the work. The useful measure is not which model completes a template fastest. It is whether the workflow leaves a junior employee better able to handle the next case.
The Korean study does not prove that AI caused the youth employment decline. It shows where the pressure has landed and why productivity alone is an incomplete measure. Before removing the next entry-level task, an employer should be able to name the skill it taught and show where a new worker will learn that skill now.
