Korea's AI-Exposed Industries Had 268,000 Fewer Youth Jobs
At noon on August 18, the Bank of Korea put two age curves on the same labor-market page.
From June 2022 through June 2026, employment among Koreans aged 15 to 29 fell by 285,000. Of that decline, 268,000 jobs, or 94.0%, sat in industries the bank classified as highly exposed to artificial intelligence. Employment among people in their 50s moved the other way. It rose by 230,000, with 173,000 of those additional jobs in the same high-exposure group.
The figures came from a new BOK issue note reported by Yonhap News Agency. The 268,000 is not a count of people dismissed by an AI system, and the dataset does not match an older worker entering a company to a younger worker leaving it. Its industry comparison also cannot reconstruct employment without ChatGPT, a weak economy, post-pandemic adjustment, remote work, or employers’ preference for experienced hires.
They show a composition change large enough to demand a better company-level explanation.
Several facts that look contradictory can coexist. A company that adopts AI can increase total employment while taking in fewer young workers. Senior employees can retain context-heavy work as software absorbs part of the routine work once given to beginners. The juniors who do get hired can use AI more often than executives, even as their cohort gets smaller.
The newest evidence from Korea and the United States points less toward a single wave of mass layoffs than toward a quieter choke point: the first paid opportunity to acquire judgment. The flow is showing up in hiring before it consistently appears in separations or base pay.
For a manager, this changes the decision attached to an entry-level requisition. The relevant comparison is not a junior salary against a software license. It is the current saving against the future cost of experienced hiring, thinner succession, lost mentoring capacity, and a cohort that never learned the company’s customers, systems, or exceptions.
A chief financial officer can see the salary that disappears this quarter. A chief human resources officer has to make the missing cohort visible before it becomes an expensive search for experience three years later.
A noon release with two age curves
The August BOK report extended a line of research the bank began before the latest data arrived. Its 2025 issue note used National Pension Service subscriber records to examine employment among people aged 15 to 29 across industries with different levels and types of AI exposure.
In the earlier window, youth employment fell by 211,000 over three years, and 208,000 of that decline was in highly exposed industries. Jobs held by people in their 50s rose by 209,000, including 146,000 in those industries. The August 2026 update covers a later endpoint and reports 285,000, 268,000, 230,000, and 173,000. Those are updated stocks across a longer period, not a second loss that should be added to the older figures.
The current breakdown makes the concentration easier to see. Youth employment fell 31.4% in information services, 27.4% in publishing, 16.6% in computer programming, systems integration and management, and 11.6% in professional services. A new worker in those fields may once have begun with research, coding, document production, classification, or checking, although the industry data cannot show which task disappeared inside a particular company.
Stocks alone can hide how the door changed. The bank also compared monthly flows in highly exposed industries before the pandemic and after the release of ChatGPT. Average monthly youth inflow fell from 32,600 during January 2016 through December 2019 to 29,100 during July 2022 through June 2026, a decline of about 11%. Average monthly outflow rose from 3,700 to 4,900, about 32%.
A missing hire leaves no layoff notice. A company can leave a requisition unopened, require three years of experience, replace a graduate intake with one lateral hire, or distribute basic work among software and incumbents. Repeated across employers, those quiet choices change who gets inside.
The outflow increase keeps retention in the picture. Young employees may be leaving because their employer restructures, because the work no longer offers a credible path, because another sector pays more, or for reasons unrelated to AI. The flow data does not identify one motive for every move. It does reject a story in which reduced entry is the only thing happening.
Education adds another clue and another caution. Since November 2022, the report put average unemployment among young university graduates at 7.0%, compared with 5.4% for people with a junior-college education or less. Before ChatGPT, from January 2019 through October 2022, the corresponding rates were 8.2% and 8.0%.
The bank used education as a proxy because highly educated workers are more common in AI-exposed occupations. That does not turn a degree into a measure of AI use. Graduate unemployment can move with hiring cycles, applicant preferences, field of study, and the jobs people are willing to accept. The comparison is evidence of a changed gap, not a count of graduates replaced by a model.
Age needs similar care. Growth among workers in their 50s does not prove that companies chose an older person instead of a named young applicant. Korea is aging, labor-force participation is changing, and older workers can enter roles that never belonged to graduates. The contrast becomes more useful when it is treated as a workforce mix to investigate, rather than a generational contest to declare.
Samil Oh, head of the BOK employment research team, centered the report’s explanation on the content of early work. Junior employees often begin with codified tasks and textbook knowledge. Experienced colleagues carry client history, organizational memory, exception patterns, and relationships that are harder to specify. A tool that handles the codified portion can therefore change the expected value of an entry hire before it changes the value of an incumbent with context.
Use mode altered the pattern in the bank’s analysis. Youth employment declined more in industries where AI use leaned toward automation, meaning the tool performed a task. The same pattern did not appear in industries where use leaned toward augmentation, such as drafting or checking that assisted a person.
That difference is operational. An exposure score says a technology could touch a task. It does not say whether a company removes the task, gives the employee more cases, improves quality, shortens a queue, or creates a new service. The choice is made in workflow design, staffing, and budget.
Exposure is not a layoff reason
AI labor claims often collapse three different measures.
Exposure estimates how much of an occupation or industry could be affected, based on its tasks. Adoption identifies a firm that reports using AI. Usage records an employee or system actually invoking a product. None of the three, by itself, measures accepted output, revenue, productivity, headcount, hiring, separation, promotion, or pay.
The BOK report starts with exposure and observes employment by age and industry. It cannot see the exact AI product, the date a manager changed a job, or the counterfactual applicant who would have been hired. Its authors also identified post-pandemic adjustment, remote work, a shift toward experienced hiring, and broader changes in industry demand as possible contributors.
The strongest alternative explanation begins there. Information services, publishing, programming, and professional services passed through a technology correction, higher interest rates, weaker demand, and changed workplace geography during the same period. Employers across Korea were already favoring experienced candidates. An exposure index can sort the industries where those forces happened to be strongest without proving AI supplied the force.
Under that account, the 268,000 figure is partly a label placed on a hiring slowdown that would have occurred anyway. The objection deserves weight because industry exposure is not randomly assigned. The automation-versus-augmentation split and similar U.S. hiring patterns make a purely unrelated slowdown less satisfying, but they do not close the causal gap. A board should treat the evidence as a reason to inspect its own decisions, not permission to attribute every cancelled requisition to a model.
A separate Korean dataset supplies a useful counter-signal. On August 6, the Korea Research Institute for Vocational Education and Training published an analysis of 106,389 firm-year observations from 2017 through 2024. It compared employment within the same companies in years when they used AI with years when they did not, controlling for firm and year effects and asset size.
Total employment was estimated to be 2.9% higher in an AI-use year. Regular employment was 2.4% higher. AI used for product or service development was associated with a 4.0% increase. The 1.5% estimate for production-process use was not statistically significant, which means the study did not establish an increase or decrease for that purpose.
This is not evidence that AI creates 2.9% more jobs across Korea. The sample covered corporations with at least 50 permanent employees and capital of at least 300 million won. It did not measure industry-wide net employment, occupation or age composition, use intensity, investment, or jobs at competitors. A growing adopter may be gaining share from another business. A firm can add engineers, salespeople, and senior operators while reducing its intake of graduates.
The firm panel asks what happened inside adopters. The BOK series asks how age composition moved across exposed industries.
| Source | Population and unit | Reported change | What the figure cannot prove |
|---|---|---|---|
| BOK August 2026 update | Koreans aged 15 to 29, by industry, June 2022 to June 2026 | 285,000 fewer youth jobs; 268,000 in highly exposed industries | That AI caused 268,000 dismissals |
| BOK 2025 issue note | National Pension Service subscribers, by age and industry | 208,000 of a 211,000 youth decline in highly exposed industries | A firm-level adoption effect or a permanent forecast |
| KRIVET firm panel | Larger Korean corporations, 106,389 firm-year observations | 2.9% higher total employment in AI-use years | Industry net jobs, junior hiring, or a role-level effect |
| Stanford and ADP | U.S. payroll sample, workers aged 22 to 25, by occupation | A 19% relative employment gap in highly exposed occupations | Economy-wide displacement or a causal AI estimate |
| U.S. Census working paper | U.S. industry-state cells, workers aged 22 to 24 | 12% adjusted decline in the most exposed quintile after 10 quarters | A randomized result or the same measure used by BOK |
Consider a software company with 1,000 employees. This is a hypothetical example, not an estimate from any of the studies. It adopts AI, adds 60 people in enterprise sales, security, infrastructure, and customer implementation, and loses 20 people elsewhere. Total employment grows 4%. If it once hired 40 graduates but now hires 10, the company can be an expanding AI adopter and a weaker entry point at the same time.
That arithmetic is easy. The harder issue is timing. Revenue and total headcount can rise now. The shortage of people who have accumulated two or three years of internal context arrives later. The company then hires experienced workers from another firm, promotes from a smaller pool, asks senior employees to carry more review work, or accepts slower expansion.
A public employment series will see those choices only after they are repeated across many companies. A quarterly planning meeting can see them before the requisition disappears.
This is why an AI business case needs denominators by role and level. Total employees, total AI users, and total output can all improve while the number of first hires, paid learning repetitions, and internal promotions falls. A chief executive can truthfully report expansion while the company quietly stops producing its next cohort of experienced workers.
Hiring slowed before separations explained the gap
Six days before the BOK update, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at Stanford’s Digital Economy Lab released a revised working paper using ADP payroll records for millions of U.S. workers through June 2026. Employment among workers aged 22 to 25 in highly AI-exposed occupations was 19% below where it would have been if it had kept pace with less-exposed occupations for the same age group.
The Korean and U.S. datasets should not be merged into one global rate.
The 19% is relative to less-exposed peers. It is neither a 19% fall in all employment among young Americans nor a claim that 19% of young people lost a job. The authors found no widespread, economy-wide displacement in their sample.
Their flow analysis points toward the entry door. Reduced hiring of young workers explained more of the gap than increased separations. Experienced workers did not show a comparable divergence. Base compensation also did not carry the main adjustment, which makes a pay-only monitor poorly suited to early detection.
The revised paper separates occupations according to how AI is used. Employment declined in occupations where usage mainly substituted for human tasks. In occupations where usage complemented workers, employment was flat or rising, particularly among experienced employees. This resembles the Korean automation and augmentation split without making the two methods identical.
The authors ran several checks. The pattern persisted when technology firms and computer occupations were excluded, and when they controlled for interest-rate exposure and remote work. It weakened when education was controlled for. Some divergence began before generative AI, and the result was more pronounced in the ADP analysis sample than in national survey benchmarks.
Those qualifications belong beside the 19% number. The paper calls its results descriptive indicators, not causal estimates. ADP observes a large payroll network, not every employer. A balanced sample reduces composition changes from firms entering or leaving the provider, but it also limits what the analysis can see about jobs moving between firms.
An April 2026 U.S. Census Bureau working paper by Lee Tucker approached the issue through matched employer-employee administrative tabulations. Among workers aged 22 to 24, regression-adjusted employment in the most AI-exposed fifth of industry-state cells declined 12% over the 10 quarters after ChatGPT’s introduction. Employment in less-exposed industries remained stable, and reduced hiring was the main source of the decline.
The Census paper uses industries and states. Stanford uses occupations inside a payroll sample. BOK uses Korean industries and a broader youth age range. The exposure measures, time windows, controls, and labor institutions differ. Their percentages cannot be averaged.
Across the studies, early-career hiring changes before a broad layoff wave appears. Separate administrative sources point to that mechanism, with confounding and measurement limits still attached.
This distinction changes what a board should request. A monthly workforce report built around total headcount and voluntary attrition can look stable while entry hiring falls. Even a diversity report can miss the issue if it shows company-wide age bands without openings, applicants, offers, starts, and internal movement by job family.
The earliest useful signal is a flow. Count requisitions opened at each level, applications that meet the stated requirements, offers, accepted starts, transfers, promotions, separations, and the time people remain at each rung. Then separate roles where AI automates a task from roles where it helps an employee perform the task.
A decline in entry hiring may be rational in one workflow. A company should not preserve a job whose only purpose is copying data between systems. It may redesign that position around customer investigation, exception handling, field observation, or quality control. It may move the learning route into an apprenticeship or rotation.
Trouble begins when the company removes the old work without naming the new route to competence. The cost is delayed, shared across departments, and easy to omit from the AI project budget.
Company growth can hide a missing cohort
A first job pays for output. It also gives a worker repeated contact with ordinary cases, mistakes, review, customers, and exceptions. Over time, those repetitions produce someone who can handle unusual work without constant supervision.
When AI absorbs the ordinary cases, the worker may learn faster if the job is redesigned well. A junior analyst can ask a model to draft a comparison, then verify source quality, trace a conflicting number, explain uncertainty, and see how a decision-maker responds. The tool compresses mechanical work while the manager keeps the learning loop.
The opposite design removes the employee and sends all routine cases to a tool. Senior staff take the exceptions. The short-term ratio looks attractive because experienced people spend less time on basic review. Yet the organization has removed the place where a beginner would have learned to recognize an exception.
At a July 9 policy forum, the Korea Development Institute and Ministry of Finance and Economy put the problem in career-ladder terms. Korea Labor Institute researcher Jiyeun Chang described the early pattern as blocked entry rather than layoffs. Hanshin University professor Byung You Cheon argued that entry tasks have acted as a training ground for practical judgment, and that losing them can push the cost of skill transmission beyond the firm that made the saving.
That external cost has an internal version. A company can recruit a trained worker later, but another employer, customer, university, government program, or the worker paid for the earlier learning. If many firms wait to buy experience, the price rises and the pool narrows.
Delayed entry also has consequences outside an AI study. A January 2026 Bank of Korea issue note estimated that a young person with one year of non-employment had a 66.1% probability of regular employment five years later. With three years of non-employment, the estimate fell to 56.2%. Each additional past year without employment was associated with 6.7% lower current real wages.
Those estimates cover delayed labor-market entry and its scarring effects. They do not assign the delay to AI. They show why a temporary reduction in entry opportunities can last longer than the software cycle that helped produce it.
The BOK note compared the pattern with Japan’s employment ice age, when people who struggled to enter stable work from the mid-1990s through the 2000s carried weaker employment and income outcomes forward. The analogy has limits. Korea in 2026 has different demographics, institutions, employers, and technology. Its use is temporal: a missed first rung can outlive the recession, policy, or product release associated with it.
For a company, the delayed bill arrives through different accounts. Recruiting may pay an experienced-hire premium. A manager can spend weeks onboarding someone who knows the profession but not the organization. A smaller internal pool weakens succession, and one resignation can take an unusual amount of context out of the building.
Finance often records these costs under different accounts and dates. The removed junior requisition belongs to this year’s payroll saving. A search fee arrives in a later recruiting budget. Senior review time is spread across projects. Lost capacity after a departure may appear as delayed revenue or customer escalation. No line automatically connects them.
That makes the first-rung decision a joint planning problem. Finance tests the multi-year cost. HR shows cohort flows and skill progression. The operating manager identifies which work produces judgment and who has capacity to review it. The AI owner distinguishes tasks that disappear from tasks that are reassigned.
Headcount reports who is present. A career ladder explains where the next experienced worker will come from.
One blanket retention target will not help. Some entry roles will shrink. Others will become more demanding, with a worker using AI from the first week. A few old tasks should disappear entirely. The aim is not to protect a familiar job description. It is to preserve an observable path from beginner to independently useful colleague.
Young workers use the tools more
Thirteen messages a week complicate the picture.
On August 12, OpenAI published administrative usage findings from its enterprise customer base. Six months after adoption, early-career employees sent 13 more messages per week than executives. Usage declined with seniority, even though the employment studies found a narrowing entry door for the same broad cohort.
This finding runs against surveys in which leaders often report more AI use. Administrative product data can see activity that a survey respondent forgets or interprets differently. It also has a strict boundary: messages measure interaction with one vendor’s products, not completed work, quality, promotions, or representative behavior across the labor force.
OpenAI also reported that firms in the top tenth of usage generated 8.3 times as many output tokens per active user as typical firms by June, up from 2.6 times in January. Since February, weekly active enterprise Codex users grew 108 times in legal, 41 times in sales, 41 times in recruiting, and 26 times in marketing, compared with five times in engineering.
These are platform measures from a vendor with an interest in greater adoption. A token is not a profitable product. A message is not an accepted work item. Rapid growth from a small base can create a large multiple. Still, the role distribution weakens the assumption that agentic use belongs only to senior technologists.
Put next to the employment studies, the data produces a practical contradiction. Employers may see early tasks as easier to automate while early-career workers use the tools more intensively. If the employee never gets hired, the company cannot observe whether that fluency would have produced a better workflow.
A recruiting requirement can make the contradiction worse. An employer asks for five years of experience and advanced AI skill in the same posting. Experienced candidates have the domain history. Younger candidates may have more recent tool practice. The company searches for both attributes in one scarce person rather than building a job in which they can be combined through review.
From an applicant’s side, the requirement creates a closed loop. Personal projects and courses can demonstrate tool use, but they cannot supply access to a real customer’s history, a regulated production system, or the judgment that comes from an accountable decision. The employer asks for experience while withholding the setting in which that experience is made. A paid entry route breaks the loop; an unpaid test or another certificate does not.
A better design pairs access with accountable work. A junior employee receives a real case and the permissions needed to handle it. The manager defines the output and assigns a reviewer. The record then shows what the tool produced, what the employee checked, what changed during review, and whether the final work was accepted.
Managers need enough review capacity for this to work. Removing every basic task does not free all senior time if the remaining exceptions are harder. Giving juniors an AI product without an experienced reviewer can speed up weak work. A software seat and a course completion cannot substitute for a feedback loop.
The most informative metric is not weekly messages. It is the movement from assisted output to independently accepted output. For a customer-support analyst, that might mean a drafted response, correct policy citation, accurate account action, and fewer corrections over 90 days. For a financial analyst, it might mean traceable sources, a reconciled model, an explained variance, and a recommendation that survives review.
Job design can use AI fluency without pretending that fluency equals judgment. The first weeks can include more cases, faster feedback, and visible comparisons between a model’s answer and the organization’s actual standard. Later work can add ambiguous cases and wider decision rights.
This is augmentation with a career consequence. The tool changes how the worker learns, while the employer keeps a paid route into the profession.
Before cancelling the junior requisition
The decision should fit on one page before it reaches a staffing meeting. Its purpose is to expose costs that sit outside the salary line, not to force every manager to keep every opening.
| Field | Question to answer | Evidence before the decision | Revisit signal |
|---|---|---|---|
| Demand | Which customer, operational, or regulatory demand created the work? | Case volume, backlog, revenue dependency, error or delay cost | Demand changes by a stated amount |
| Task mode | Which tasks will be automated, assisted, moved, or stopped? | Workflow sample with tool and human steps | Correction time or exception volume rises |
| Entry route | Where can a person with little experience now enter? | Requisition, apprenticeship, rotation, or paid project owner | No starts in the job family for two quarters |
| Learning repetitions | Which cases build judgment rather than busywork? | Defined case types, volume, feedback, and increasing difficulty | Employee sees only generated output or rare exceptions |
| Review capacity | Who reviews the work, and how many hours are funded? | Named reviewer, weekly capacity, backup, escalation route | Review backlog or senior overtime exceeds the limit |
| Accepted output | What can a new hire complete by day 90? | Quality standard, acceptance owner, correction count | Completion rises but acceptance or accuracy falls |
| Progression | Which additional decisions can the worker own by month six? | Skill evidence, case range, customer exposure, manager sign-off | Training is complete but decision rights do not change |
| Replacement cost | What will an experienced external hire cost later? | Market pay, search time, fee, onboarding time, vacancy cost | Premium or time to fill crosses the approved case |
| Succession | Which senior roles depend on this cohort? | Age and tenure mix, internal candidates, critical context | One departure leaves no ready or near-ready successor |
| Review date | When does the company reopen the choice? | Owner, date, data source, and possible decisions | The date arrives without a new decision record |
Start with demand because automation can hide a volume problem. If customer cases are rising while headcount stays flat, productivity may be real. If demand fell first, assigning the whole saving to AI overstates the technology’s contribution.
Task mode forces the manager to name what changed. A tool may draft an answer while a person still gathers facts, secures permission, checks policy, takes an action, and handles an appeal. Calling the entire workflow automated erases human work and creates a false staffing case.
The entry-route row prevents an informal promise. A company that closes ten graduate roles and mentions a future apprenticeship has removed ten openings today. The apprenticeship counts only after it has a budget, start date, manager, paid participants, defined work, and a conversion decision.
Learning repetitions separate useful experience from course attendance. A worker develops judgment by handling cases and receiving correction. A catalog of videos may explain the tool without creating evidence that the person can use it on company work.
Review capacity puts senior labor inside the budget. If an experienced employee spends six hours each week checking junior work, record it. If AI cuts that time after three months, record the reduction. If exception complexity increases it, the original business case should change.
Accepted output gives the first 90 days a business unit. A manager can define what the job must produce without claiming a lifetime career path in advance. Quality, correction, customer result, and time belong together. Speed alone encourages unreviewed volume.
Progression makes training directional. A person might start by checking sources, then reconcile a case, then explain a recommendation to a customer, then own a bounded decision. The sequence will differ by role. It should end in wider responsibility, not permanent supervision of a machine.
Replacement cost makes the CFO’s time horizon explicit. The comparison includes more than future salary. Search time, agency fees, onboarding, vacancy, context transfer, and the risk of a bad lateral hire belong in the model. An apprenticeship can still be more expensive. At least the two options use the same horizon.
Succession connects one requisition to the workforce plan. A company may decide that a role has no future internal ladder and buy the capability as needed. That is a strategic choice. Letting the pipeline vanish because each manager optimized a local budget is not.
The same fields can improve public monitoring. Korea’s Ministry of Employment and Labor said in July that it was developing a Korean canary dashboard to track AI employment shocks in real time and refine youth measures. The announcement describes a monitoring project, not an intervention result.
A national dashboard should keep stocks and flows separate. Employment by age and exposure is a starting point. Hiring, separation, movement, job level, occupation, industry, and use mode show where the change occurs. Program participation belongs in another column from a paid start, six-month retention, and movement into regular employment.
Companies do not need to expose individual employee records to contribute. Aggregated cohorts with privacy thresholds can show how many entry roles opened, started, progressed, and closed, and whether automation or augmentation changed the workflow. The same data gives management an earlier warning than a national release.
The first Monday without a first hire
Imagine a manager arriving on Monday with approval to remove one junior analyst requisition. The team bought an AI research tool. Drafts arrive faster, and two senior analysts believe they can absorb the remaining work. The annual salary disappears from the forecast.
Nothing fails that morning.
By October, the tool has produced hundreds of drafts. The senior analysts correct the difficult ones and rarely save the reasons in a place another person can study. Basic cases move quickly. Nobody measures how much review time migrated into meetings and late messages.
In January, one senior analyst leaves. The other knows the customer history but cannot cover every account. Recruiting opens a role requiring four years of experience with the company’s market and strong AI skills. Candidates with the market knowledge want more pay. Candidates with recent tool fluency have not handled the same decisions.
The original requisition saved money. The later vacancy does not prove the decision was wrong. Demand may have changed, or the old junior role may have contained work that deserved to disappear. The missing evidence is whether the team considered a redesigned entry route before closing it.
The one-page file would have made the choice testable. It would identify the tasks the tool handled, the review hours that remained, the cases through which a beginner could learn, the accepted output expected after 90 days, the experienced-hire premium, and the date for another decision.
The company might still cancel the job. It could create a six-month rotation shared by several teams, hire an apprentice after demand crosses a threshold, or decide to buy external expertise. Each route has an owner and a future cost.
Without that record, headcount growth can reassure the board while the first rung narrows. Usage can rise while no one gains decision rights. Productivity can improve while senior employees become harder to replace. A national employment release then reveals a cohort gap after thousands of local decisions have accumulated.
Korea’s 268,000 figure does not settle what caused the decline. It makes the composition impossible to ignore. Before removing an entry role, account for how the work will be done and how the next experienced worker will be made.
On the following Monday, the manager should be able to answer one concrete question: if nobody starts here, where will the person qualified to lead this work come from?
This article separates industry exposure, firm adoption, product usage, hiring flows, and causal claims. Published August 20, 2026.