AI Adoption and Occupational Exposure in Korea한국의 AI 도입과 직업별 AI 노출

Korea uses Claude intensively and its young workers are losing ground, but the two facts have not been connected at the level of occupations. In August 2025, Korea's per-capita Claude usage was 3.73 times what its working-age population predicts, fifth among all countries (Anthropic, 2025). Between June 2022 and June 2026, employment of Koreans aged 15–29 fell by 285,000.

The Bank of Korea places 94% of that fall in AI-exposed industries, while employment of workers in their 50s rose by 230,000 (Bank of Korea, 2026). Its evidence is industry-level, so it cannot say which jobs inside those industries changed. Its authors also read AI as accelerating hiring shifts already under way, such as the correction to post-pandemic overhiring (Bank of Korea, 2026).

Anthropic's observed exposure measure offers an occupation-level lens built from what people actually do with Claude. Anthropic's own application so far covers only the United States (Massenkoff and McCrory, 2026). This paper brings that lens to Korea in three steps:

1. It profiles how Koreans use Claude, using the Anthropic Economic Index country releases from August 2025 to May 2026: adoption level and trend, position relative to income, and the tasks and collaboration modes that set Korea apart from peer economies.
2. It maps observed exposure onto Korea's workforce through a SOC → ISCO-08 → KSCO crosswalk, showing which occupations and age groups sit in exposed jobs.
3. It estimates a first triple difference on public Local Area Labour Force Survey tables, then sets out a microdata design that sharpens the test and reconciles occupation-level estimates with the Bank of Korea's industry results.

Section 2 describes the data and the crosswalk. Sections 3–5 present the usage profile, the exposure map and a first look at youth employment. Section 6 sets out the causal design, and Section 7 draws out implications for Korean policy and research partnerships.

The analysis joins three public sources: Anthropic Economic Index country aggregates, Anthropic's occupation-level observed exposure, and Korean labour force tables. Every number in the paper can be regenerated from the released files.

Claude usage by country. The Economic Index releases privacy-preserving aggregates from samples of about one million Claude.ai conversations (Anthropic, 2025). I use the one-week samples from August 2025, November 2025 and February 2026, and the calendar-month aggregates for April and May 2026 in the 2026-06-26 data release (Anthropic Economic Index data). The AI Usage Index (AUI) is a country's share of conversations divided by its share of the working-age population, over countries with at least 200 sampled conversations. Recomputed from the raw August 2025 counts, it matches Anthropic's published index for all 115 countries to four decimal places. The three earlier samples are restated on the basis of the monthly index (Appendix).

Observed exposure. Massenkoff and McCrory (2026) score 756 US occupations (SOC 2018) by the share of their tasks that Claude already performs at work. A task counts when an LLM could theoretically speed it up (Eloundou et al., 2024) and it appears in enough work-related traffic in the August and November 2025 samples. Automated use counts fully, augmentative use at half weight, and tasks are weighted by their share of time in the job.

From US to Korean occupations. The scores reach the 8th Korean Standard Classification of Occupations (KSCO-8) through three official crosswalks:

1. SOC 2018 to SOC 2010 (BLS).
2. SOC 2010 to ISCO-08 (BLS).
3. ISCO-08 to KSCO-8 unit groups (Statistics Korea).

Each US occupation carries its May 2024 OEWS employment and splits it equally across many-to-many links. A Korean unit group's exposure is therefore the employment-weighted mean of the US occupations mapped to it. Only two US occupations, 0.03% of US employment, find no Korean match, and all 163 KSCO-8 minor groups receive a score.

Sub-major scores weight minor groups by Korean employment; US weights give almost the same values (correlation 0.98). Because exposure is fixed within a group, comparisons with the US hold each group's internal mix at US weights and reflect differences in the mix across groups.

Korean employment. Employment comes from Local Area Labour Force Survey (LALFS) tables on KOSIS. They give the 56 KSCO-8 sub-major groups by five age bands for every April and October round from 2013 to 2025, plus minor groups by sex for weights. Statistics Korea documents the re-coding of 2017–2024H1 to KSCO-8 (Statistics Korea, 2025); KOSIS also publishes KSCO-8 tables for 2013–16, whose coding is not documented. The KSCO-7 tables KOSIS published until 2024H1 give event-study coefficients within 0.010 of the KSCO-8 ones in every round, with no seam between 2016 and 2017 (Appendix).

Public tables do not cross minor groups with age, and their youngest band is 15–29. The microdata design in Section 6 removes both limits.

On the April–May 2026 average, Korea uses Claude about 1.6 times as intensively as advanced economies at its income level, second only to Australia. Its use leans toward work, toward collaboration rather than delegation, and toward finance and language tasks.

Adoption. On the basis of the monthly index, which divides by conversations attributed to a country, Korea's AUI was 4.42 in August 2025, fifth of 115 countries (3.73 as published under the earlier method). It eased to 3.82 in November 2025 and 3.69 in February 2026, then stood at 3.92 and 3.78 in April and May 2026, ranking between 7th and 14th after August. Korea accounts for about 3% of all Claude.ai traffic.

Relative to income. Usage rises with income. Across all countries, the elasticity of the AUI with respect to GDP per working-age person is 0.69 in August 2025, close to the 0.7 Anthropic (2025) reports, 0.70 in November 2025 and 0.75–0.78 in the 2026 samples. Among the IMF's advanced economies the line is flatter: 0.41–0.59 by sample and 0.58 on the April–May 2026 average.

Korea's GDP per working-age person, about $51,500, is in the lower third of the advanced economies, yet its usage sits well above their income line (Figure 1). On the April–May 2026 average, Korea's ratio of actual to predicted AUI is 1.59, second to Australia's 1.70; Japan and Taiwan, at similar incomes, sit below the line. Across the five samples the ratio runs from 1.47 to 1.92 and ranks between first and fourth among the 36–37 advanced economies with data: second, behind Israel, in both 2025 samples, fourth in February, first in April and third in May 2026. Against the line fitted across all countries, the ratio is 1.8–2.3.

What it is used for. Table 1 compares Korea with the global average and three peers in April–May 2026.

Table 1. Usage profile, April–May 2026. Source: Anthropic Economic Index, 2026-06-26 release.

Korea's work share is similar to Japan's and Taiwan's, and was higher still in November 2025 (51.1% against 46.0% globally) and February 2026 (50.2% against 45.2%). What sets Korea apart is how it works with Claude. Its augmentation share is the highest of the twelve economies compared (Korea, Japan, Taiwan, Singapore, India, Israel, Australia, the United States, Canada, the United Kingdom, Germany and France) and third of 114 countries, behind Norway and Denmark. The lead comes from task iteration, 33.2% of conversations against 30.1% globally and narrowly ahead of Australia's 33.0%. Koreans also use Claude to learn more often (18.0% against 16.7%) and issue one-shot directives less often (29.7% against 31.6%) than the global average, but on both they sit mid-range among the peers.

Whose tasks. Mapping conversations to the occupations whose tasks they resemble shows a tilt toward media, business and science (Figure 2). Computer and mathematical tasks are still the largest category at 23.1%, slightly below their global share of 24.7%. Arts, design and media tasks run 3.0 points above their global share, while sales tasks run 2.0 points below.

Two finer clusters stand out. Buying-and-investing requests take 4.0% of Korean conversations against 1.9% globally, and securities sales agents and financial analysts are among the most over-represented occupations (2.2 and 1.7 times their global shares). Translation requests run at twice the global rate, and interpreters and translators account for 1.8% of Korean conversations, 1.9 times their global share. Web front-end coding (0.74 times) and lifestyle requests such as cooking (0.48 times) are under-represented.

The average Korean worker in October 2025 held a job with observed exposure of 0.098, three-quarters of the US average of 0.132. The gap comes from Korea's occupational mix: 42% of Korean employment sits in groups with exposure below 0.02, such as farming, machine operation, food service, care work and elementary occupations.

Where exposure sits. Table 2 lists the six most exposed sub-major groups. With six others they form the top exposure quartile used in Section 5: the twelve most exposed groups, which held 22% of Korean employment in October 2022 (store sales, which straddles the cut-off, falls in the third quartile). Several are traditional entry points for graduates, and general office support, customer service and ICT each employ young workers well above the economy-wide rate of 12.1%.

Table 2. The six most exposed occupation groups. Sources: Massenkoff and McCrory (2026) via the crosswalk in Section 2; LALFS via KOSIS.

Exposure by age. Because exposure is fixed within each group, its age profile reflects only which groups each age works in (Figure 3). Exposure peaks in the 30s and 40s, with workers aged 15–29 close behind, then dips in the 50s and falls to less than half the peak after 60.

Young workers are modestly over-represented in the top quartile, which in October 2025 held 24.3% of employed 15–29-year-olds against 21.7% of older workers. Korea's more distinctive feature is the very low exposure of its 60+ workforce, which grew from 14% of employment in 2013 to 24% in 2025.

Young workers' employment in the most exposed occupations has been shrinking relative to older workers' since at least 2013. After ChatGPT it stayed close to that path: the public tables show no detectable break, though they cannot exclude a moderate one.

Where the decline fell. Employment of 15–29-year-olds fell from 3.95 million in October 2022 to 3.52 million in October 2025, a drop of 427,000 or 10.8%, while employment at 30 and over rose by 1.05 million. The fall hit every exposure quartile (Table 3). The most exposed quartile accounts for a third of it, somewhat more than its quarter share of youth employment, so two-thirds came from other occupations; older workers in the top quartile gained 7.1%.

Table 3. Employment of 15–29-year-olds by exposure quartile. Source: LALFS via KOSIS, October rounds.

With the Bank of Korea's cut, the top two quartiles, exposed occupations account for 58% of the youth decline, close to their 60% share of youth employment in October 2022, against 94% for exposed industries in its analysis. The data and windows differ, so the comparison is indicative: survey employment from October 2022 to October 2025 here, June 2022 to June 2026 in the Bank of Korea's note.

The longer view. Figure 4 indexes employment to October 2022. Youth employment in the top quartile fell from 115 in 2013 to 100 in 2022, about 1.6% a year, and then to 86 in 2025, about 4.9% a year. In other occupations youth employment rose until 2022 and then fell too. Relative to young workers elsewhere, the comparison Brynjolfsson, Chandar and Chen (2026) draw, the top quartile lost about 5% between 2022 and 2025, after losing about 3% a year before 2022. For workers aged 30 and over both groups grew.

Triple-difference estimates. I estimate a triple difference on 56 occupation groups, 5 age bands and 26 survey rounds from April 2013 to October 2025:

\ln E_{oat} = \sum_{k \neq \text{2022H2}} \beta_k \, \left( X_o \times \text{Young}_a \times \mathbf{1}[t = k] \right) + \mu_{ot} + \lambda_{at} + \alpha_{oa} + \varepsilon_{oat}

X is observed exposure standardised over October 2022 employment, and Young marks ages 15–29. The fixed effects absorb occupation-specific cycles such as tech hiring, age-specific shocks including shrinking cohorts, and each occupation's age profile. Cells are weighted by 2019–22 employment, and standard errors are clustered by occupation (56 clusters).

The event study shows a steady pre-trend (Figure 5). The coefficient falls from 0.167 in April 2013 to zero by construction in October 2022: young workers' employment relative to older colleagues was already shrinking faster in more exposed occupations in 2013, a decade before generative AI. After 2022 the coefficients turn negative, reaching −0.034 (s.e. 0.018) per standard deviation of exposure by October 2025, on or slightly above the extrapolated 2013–22 trend.

The comparison group matters after 2022. Against workers aged 30–49 alone, the pre-2022 decline is similar (0.133 in April 2013), but the post-2022 coefficients stay near zero: −0.008 in October 2025, and −0.002 on average against −0.020 for all workers 30 and over. Most of the post-2022 relative decline therefore reflects relative employment gains among workers 50 and over in exposed occupations, in line with the Bank of Korea's finding for workers in their 50s.

Testing for a break. A simple pre/post contrast would mistake this trend for an effect: the relative decline from October 2019 to October 2022 is as large as the one after 2022 (Table 4). The break test instead fits a linear trend on the pre-2022 rounds only, measures each post-2022 round's deviation from its extrapolation, and averages the six. Because the answer depends on the window used for the trend, Table 4 reports three: 2013–22, 2017–22 and 2019H2–22.

Table 4. Triple-difference estimates on log employment, weighted OLS; comparison group ages 30 and over unless stated. Standard errors clustered by occupation in parentheses; randomization-inference 95% intervals in brackets, from 2,000 reassignments of exposure scores across the 56 groups. Deviation rows average the six post-2022 rounds' deviations from a linear trend fitted on the window's pre-2022 rounds. Poisson PML estimates for the rows that compare with ages 30 and over are within 0.007 of these.

None of the deviations is distinguishable from zero, and their sign depends on the window. The fitted pre-trend flattens as the window shortens (for the top quartile, from −0.044 a year over 2013–22 to −0.022 over 2019H2–22), so the top-quartile deviation moves from +0.051 to −0.037. The intervals are wide. With 12 treated groups out of 56, clustered errors can mislead (MacKinnon and Webb, 2018), so Table 4 also reports randomization inference (MacKinnon and Webb, 2020); its intervals are similar, slightly wider for the top quartile. Their lower bounds reach −0.126 log points, a decline of about 12%, and the minimum detectable effect at 80% power is 0.05–0.06 log points per standard deviation of exposure, or 0.11–0.12 for the top quartile. The public tables rule out a collapse in exposed youth employment after ChatGPT, but not a moderate decline.

Robustness. Poisson pseudo-maximum likelihood on levels gives estimates within 0.007 of the Table 4 estimates that compare with ages 30 and over. Re-mapping exposure to KSCO-7 and re-running the event study on the KSCO-7 tables gives coefficients within 0.010 of the KSCO-8 ones in every round from 2013 to 2024H1 (correlation 0.996), with no seam between 2016 and 2017. The 2013–16 rounds, whose KSCO-8 coding is undocumented, carry about 40% of the pre-2022 decline; the 2017–22 window, which leaves them out, still shows a declining, if imprecise, trend (Appendix).

Reading the evidence. Through October 2025, Korean survey data show no detectable break in young workers' employment in Claude-exposed occupations beyond a relative decline already under way in 2013. This parallels the US finding of no systematic increase in unemployment among highly exposed workers (Massenkoff and McCrory, 2026). US payroll data show more: by June 2026, employment of 22–25-year-olds in highly exposed occupations was about 19% below where it would be had it kept pace with same-age workers in less-exposed occupations (Brynjolfsson, Chandar and Chen, 2026). The closest Korean measure, the 5% relative fall of young workers in the top quartile over 2022–25, is smaller and continues an older trend. Three features of the Korean data would make a US-sized effect harder to see:

1. Age. An effect confined to ages 22–25 would be diluted across the 15–29 band and could sit inside the intervals in Table 4.
2. Margin. In the United States the adjustment runs mainly through hiring: the job-finding rate of 22–25-year-olds into exposed occupations fell about 14% from its 2022 level, an estimate that is only just significant (Massenkoff and McCrory, 2026). Public LALFS tables show only employment stocks; the design in Section 6 adds entry.
3. Granularity. 56 groups blur exposure contrasts within groups.

Brynjolfsson, Chandar and Chen also see some differential trends before generative AI, but in the United States the relative position of exposed young workers had returned to its pre-pandemic level by November 2022. In Korea the pre-existing trend dominates the data.

Section 5 finds no detectable break on public tables, but 56 occupation groups and a 15–29 age band could hide an effect concentrated among new graduates. The next stage estimates the design on survey microdata, which add minor occupation groups, single-year ages, industry and job-start dates, alongside the all-age design used for the United States.

Data. The main source is the public-use microdata of the Local Area Labour Force Survey (LALFS), which covers about 232,000 households each April and October (Dailian, 2026). Since 2024H2 its results use the 8th Korean Standard Classification of Occupations (KSCO), with 2017–2024H1 back-coded, so the panel runs from 2017H1 to the latest wave on one code (Statistics Korea, 2025). The public file with the finest occupation detail (minor groups) also codes industry but omits region (MDIS).

Arms. Each KSCO minor group takes the employment-weighted observed exposure of the US occupations mapped to it (Section 2). Treated occupations are the top exposure quartile and controls are occupations with zero observed exposure, as in the US analysis (Massenkoff and McCrory, 2026); the bottom quartile serves as an alternative control and the middle quartiles are dropped. The estimation will report occupations and workers per arm, because a broad group's average is rarely exactly zero.

Outcomes.

• Employment counts by occupation, industry, age group and wave, from survey weights.
• Entry: workers who started their current job in the six months before the survey, dated by start month. A twelve-month window overlaps across waves and counts only hires who stayed.
• Slack outside unemployment: young people "just resting" (쉬었음) or preparing for jobs. In September 2026 youth unemployment was 5.4% but the expanded underutilization rate was 15.6% (inews24, 2026).
• Unemployment by previous occupation, for recent job leavers who report one.

Estimating equation. Cells are occupation o × industry j × age group a × wave t, and the model is estimated by Poisson pseudo-maximum likelihood (PPML; Santos Silva and Tenreyro, 2006) on weighted counts:

\mathbb{E}[Y_{ojat}] = \exp\Big( \alpha_{oja} + \delta_{jat} + \mu_{ot} + s_{oa,h(t)} + \sum_{k \neq \text{2022H2}} \beta_k \, \text{High}_o \times \text{Young}_a \times \mathbf{1}[t = k] \Big)

PPML lets the fixed effects absorb proportional shocks, such as the tech downturn or rising enrolment, that would load on β in a levels model. Industry-by-age-by-wave effects δ absorb youth-specific shocks within an industry: hiring freezes, rate hikes and the shift from annual group hiring to rolling hiring. Occupation-by-wave effects μ absorb shocks common to all ages, and s absorbs April–October seasonality.

Each β is then the change in the ratio of youth to comparison-age employment inside exposed occupations, relative to unexposed occupations in the same industry. Young means 15–29, with 25–29 and 15–34 as alternatives; the comparison group is 35–49, because employment of people in their 50s rose inside exposed industries (Bank of Korea, 2026) and, in Section 5, inside exposed occupations. This estimand nets out effects common to all ages, so the design also includes the US-comparable all-age difference-in-differences and a young-only version.

Timing. The reference wave is October 2022, the last before ChatGPT's release on 30 November 2022. October 2022 sits near the peak of the tech hiring boom, so the baseline is 2017–19, with the waves from April 2020 to April 2022 shown as leads but kept out of it (Figure 6). Rambachan and Roth (2023) bounds report how large a violation of parallel trends would overturn each estimate.

Exposure is measured from Claude usage in August and November 2025, almost three years after treatment. That gap is why the checks below also rank occupations by theoretical exposure, which reflects capabilities as of early 2023 rather than later usage.

Inference. Few treated clusters make the wild cluster bootstrap unreliable, especially at the sub-major level (MacKinnon and Webb, 2018). P-values therefore come from randomization inference that reassigns treatment across occupations, as in Table 4, and the minimum detectable effect is set from 2017–19 placebo estimates before estimation. For comparison, the US design detects differential unemployment changes of about 1 percentage point (Massenkoff and McCrory, 2026).

Threats and checks.

Table 5. Threats to identification and the checks built into the design.

Reconciling with the Bank of Korea. The Bank of Korea's high-exposure group is industries in exposure quartiles 3 and 4 (Bank of Korea, 2026), and its 2025 note counted jobs in National Pension Service records (Bank of Korea, 2025). Restricting LALFS to wage workers and re-aggregating to that cut shows whether the youth decline sits in exposed occupations or in other jobs inside exposed industries.

The evidence so far casts AI as, at most, an accelerant of a long-running squeeze on entry-level white-collar work in Korea, not its origin. That has four implications for policy and for measurement.

Aim at the career ladder, not at exposure alone. Two-thirds of the 2022–25 fall in youth employment came from outside the most exposed quartile, and youth employment in exposed occupations was already falling relative to older workers' by 2013, so policy aimed only at AI-exposed jobs would miss most of the problem. The Bank of Korea reads AI as accelerating hiring shifts already under way and proposes supporting the training and mentoring costs that make entry-level hiring worthwhile (Bank of Korea, 2026). The occupation-level evidence here supports that emphasis.

Test whether how Koreans use AI matters. Korean users augment more than any of the eleven peers compared, mainly by iterating on tasks. US payroll evidence finds early-career declines concentrated in occupations where AI usage tends to automate tasks; where AI complements workers, employment is flat or rising, particularly among more experienced workers (Brynjolfsson, Chandar and Chen, 2026). That split is between occupations, not between ways of using AI, so it does not show that collaborative use protects young workers. Korea's unusually collaborative usage makes it a good place to test that before building training programmes around it.

Measure at the right grain, with Korean partners. Three upgrades would move this first look toward causal evidence:

1. LALFS microdata from Statistics Korea's MDIS, with minor occupations, single-year ages, industry and job-start dates, for the design in Section 6.
2. Employment Insurance acquisitions by occupation and age from the Ministry of Employment and Labor, which capture the hiring margin where US effects appear.
3. A Korea-specific exposure measure built from Korea's own task-level Claude usage, replacing US occupational weights.

Statistics Korea, the Ministry of Employment and Labor, the Bank of Korea and the Korea Labor Institute hold the data and expertise each step needs.

Track it as the data arrive. A short Korea update at each Economic Index release and each April and October survey round would show, on one page, usage relative to income, the usage mix and exposure-weighted youth employment. That is the monitoring a fast-adopting economy needs while any effects are still small.

• Anthropic (2025). Anthropic Economic Index report: Uneven geographic and enterprise AI adoption. September 2025.
• Anthropic Economic Index data. Anthropic/EconomicIndex, Hugging Face; releases 2025-09-15, 2026-01-15, 2026-03-24, 2026-06-26 and labor_market_impacts. CC-BY.
• Bank of Korea (2025). AI 확산과 청년고용 위축 [AI diffusion and the contraction of youth employment]. BOK Issue Note, October 2025. Reported in Herald Economy, 29 October 2025.
• Bank of Korea (2026). 청년고용 위축, AI 탓인가? 변화하는 경력 사다리와 대응 과제 [Is AI to blame for the youth employment slump? The changing career ladder and policy responses]. BOK Issue Note, 18 August 2026. Reported in The Fact and Etoday.
• Brynjolfsson, E., Chandar, B., and Chen, R. (2026). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab, revised August 2026.
• Bureau of Labor Statistics. SOC crosswalks: SOC 2010 to 2018 and SOC 2010 to ISCO-08; Occupational Employment and Wage Statistics, May 2024.
• Callaway, B., Goodman-Bacon, A., and Sant'Anna, P. H. C. (2024). Difference-in-differences with a continuous treatment. arXiv:2107.02637.
• Dailian (2026). 국가데이터처, 2026년 상반기 지역별고용조사 [The 2026 H1 Local Area Labour Force Survey]. 17 April 2026.
• Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308.
• IMF (2025). World Economic Outlook database, April 2025: Groups and aggregates. International Monetary Fund.
• inews24 (2026). Coverage of Statistics Korea's September 2026 employment trends.
• MacKinnon, J. G., and Webb, M. D. (2018). The wild bootstrap for few (treated) clusters. Econometrics Journal, 21(2), 114–135.
• MacKinnon, J. G., and Webb, M. D. (2020). Randomization inference for difference-in-differences with few treated clusters. Journal of Econometrics, 218(2), 435–450.
• Massenkoff, M., and McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic, 5 March 2026.
• MDIS (2026). Local Area Labour Force Survey public-use microdata: file types. Microdata Integrated Service.
• Rambachan, A., and Roth, J. (2023). A more credible approach to parallel trends. Review of Economic Studies, 90(5), 2555–2591.
• Santos Silva, J. M. C., and Tenreyro, S. (2006). The log of gravity. Review of Economics and Statistics, 88(4), 641–658.
• Song, H., and Lee, B. (2026). 주요국의 AI 사용 실태와 한국의 현황 [AI use in major economies and the state of Korea]. KIEP World Economy Focus, Vol. 9 No. 4. Korea Institute for International Economic Policy, 29 January 2026.
• Statistics Korea (2025). 2024년 하반기 지역별고용조사 시군구 주요고용지표 [2024 H2 LALFS key employment indicators by district]. Korea Policy Briefing, 20 February 2025.
• Statistics Korea. KOSIS tables DT_1ES3B28J and DT_1ES3B28S (LALFS employed persons by occupation and age, on KSCO-8 and KSCO-7) and DT_1ES3B32J and DT_1ES3B32S (by occupation minor group and sex); KSCO-8 to ISCO-08 and KSCO-8 to KSCO-7 concordances, Korean Standard Statistical Classification portal. Retrieved 24 October 2026.

The crosswalk loses almost no employment. The youth estimates hold across estimators and occupation codes; after 2022 they depend on the comparison group and the pre-trend window.

Table A1. Crosswalk diagnostics.

AUI basis. The recomputed August 2025 index matches Anthropic's published values for all 115 countries, with a maximum absolute difference below 0.0001. That method counts conversations without a country in the usage denominator but not in the population denominator, which scales every country by the attributed share of traffic: 0.84 in August and November 2025 and 0.82 in February 2026. The 2026-06-26 monthly index divides by attributed traffic only, so Section 3 restates the three earlier samples on that basis. Within a sample the restatement is one common factor, so ranks and income ratios are unchanged. One observation is dropped as misattributed traffic: the November 2025 sample assigns Seychelles 24,715 conversations, close to Korea's 30,618, for about 87,000 working-age people. Kept, it would rank first with an AUI above 1,000 and lower every other country's index by about 3%.

Sensitivity of the triple difference.

• Comparison group of ages 30–49: the pre-2022 decline is similar (0.133 in April 2013, against 0.167), but the post-2022 coefficients average −0.002 against −0.020 (−0.008 against −0.034 in October 2025). Break-test deviations are less negative in every window: for the top quartile +0.074 with the 2013–22 trend (clustered 95% interval 0.001 to 0.148, randomization p = 0.08), and +0.002 and −0.028 with the 2017–22 and 2019H2–22 trends.
• PPML on levels, comparison group 30 and over: pooled and break-test estimates within 0.007 of weighted log OLS.
• KSCO-7 tables, 2013H1–2024H1: event-study coefficients within 0.010 of the KSCO-8 estimates in every round (correlation 0.996), with no seam between 2016H2 and 2017H1. The 2013–16 rounds, whose KSCO-8 coding is undocumented, carry about 40% of the pre-2022 decline (0.167 to 0.096).
• Pre-trend window, exposure per s.d.: the fitted trend flattens from −0.016 a year (s.e. 0.005) over 2013–22 to −0.012 (0.007) over 2017–22 and −0.008 (0.007) over 2019H2–22, and the average post-2022 deviation moves from +0.011 to −0.004 and −0.019 (Table 4).
• Randomization inference: exposure scores are reassigned across the 56 groups 2,000 times, recomputing the z-score and the top quartile each time. Two-sided p-values for the six deviations against ages 30 and over range from 0.26 to 0.83; the 95% intervals invert the test under a constant additive effect.

Replication. Code and intermediate tables are available from the author. Every table and figure is generated from the public files listed in the references.