Global Automation Atlas

Jasmin Baier; Prashant Garg; Tommaso Crosta

arxiv: 2605.17086 · v1 · pith:YVI2MFA5new · submitted 2026-05-16 · 💰 econ.GN · cs.AI· cs.CY· q-fin.EC· stat.AP

Global Automation Atlas

Prashant Garg , Tommaso Crosta , Jasmin Baier This is my paper

Pith reviewed 2026-05-20 14:59 UTC · model grok-4.3

classification 💰 econ.GN cs.AIcs.CYq-fin.ECstat.AP

keywords automation exposurelabor substitutionlabor augmentationtask-based measurementcross-country comparisonAI role in automationtechnology channelsgender differences

0 comments

The pith

A country-specific task atlas shows automation exposure rising with income but skewed toward labor substitution especially in poorer nations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper creates a new way to measure how much automation affects different jobs that adjusts for each country's specific economy instead of using the same scores everywhere. It separates cases where machines replace workers from cases where they help workers do more, while also tracking the type of technology and how much AI is involved. This matters for understanding automation's impact because most prior measures cannot compare rich and poor countries fairly or isolate these different effects. The approach produces detailed labels for over two million task-country combinations across 124 nations that together represent nearly all global population and output. Five patterns stand out in the results, including wide differences in overall exposure, a tilt toward replacement over help, simpler technologies dominating in low-income places, and greater female exposure to replacement effects.

Core claim

We develop a task-based and country-specific approach to classify automation exposure across the world to disentangle labor-substituting from labor-augmenting automation, the relevant technology channel, and the material role of AI. Our measure spans 124 countries, generating an atlas of 2.33 million task-country labels for economies covering 99% of world population and GDP. We present five descriptive results on uneven exposure levels, the balance between substitution and augmentation, the prevalence of different technology channels by income, the varying role of AI, and gender differences in substitution exposure.

What carries the argument

A task-based and country-specific classification system that generates separate labels for each task in each country's economic context to separate substitution from augmentation and to identify technology channels including AI.

If this is right

Exposure levels vary sharply from about 3 percent of tasks in the lowest-income settings to over 60 percent in the highest, and overall exposure tends to increase with national income while still differing within income groups.
Most exposed tasks across countries lean toward labor substitution rather than augmentation, with low-income countries showing stronger substitution skew and middle-income countries showing more mixed patterns.
Less advanced automation forms account for more than half of exposed tasks in low-income countries but only about a quarter in high-income countries, while more complex channels become more common as income rises.
AI appears less often in simpler automation channels overall, yet it shows up more in labor-substituting roles in lower-income settings and more in labor-augmenting roles in higher-income settings.
Female workers appear more exposed to labor-substituting automation than male workers across the countries studied.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Linking the atlas labels to existing cross-country employment or productivity datasets could allow researchers to test how exposure levels relate to actual job changes or wage shifts over time.
Development policies might need to differentiate responses by income group, focusing on retraining for substitution-heavy exposures in poorer countries while addressing augmentation opportunities in richer ones.
Repeating the classification exercise with updated task data or new technology classifications could track how exposure patterns evolve as automation technologies advance.
The separation of exposure level, labor margin, technology channel, and AI involvement into distinct dimensions opens the door to more targeted studies of how each factor affects labor markets independently.

Load-bearing premise

The task descriptions and fixed classification rules can reliably distinguish labor substitution from augmentation and identify specific technology channels including AI in a consistent way across countries at very different income levels without major measurement bias or context errors.

What would settle it

Independent firm or worker surveys in several low-income and high-income countries that record materially different shares of tasks being automated or augmented, or different technology channels, than the atlas classifications predict for those same tasks.

Figures

Figures reproduced from arXiv: 2605.17086 by Jasmin Baier, Prashant Garg, Tommaso Crosta.

**Figure 2.** Figure 2: Cross-country patterns in automation exposure. (a) Economically exposed task share. (b) Exposed task share and GDP per capita. 6 7 8 9 10 11 log GDP per capita 0.1 0.2 0.3 0.4 0.5 0.6 Economically exposed task share Income group Low income Lower middle income Upper middle income High income India China U.S. Indonesia Pakistan Nigeria Brazil Bangladesh Mexico DR Congo Germany South Africa Korea Angola Polan… view at source ↗

**Figure 4.** Figure 4: Automation channels and AI materiality. 6 7 8 9 10 11 Log GDP per capita (PPP) 10% 20% 30% 40% 50% 60% Share of exposed task mass in channel Low Lower-middle Upper-middle High (a) Channel mix across development Rule-based workflow Information transformation Physical execution Planning & control Inference / scoring 0% 20% 40% 60% 80% 100% AI-material share within exposed tasks Physical execution Rule-based … view at source ↗

**Figure 6.** Figure 6: ranks these covariates using random forests and mean absolute TreeSHAP values. The bars measure average predictive contribution in percentage points of the outcome share; colours show the accumulated-local-effect direction. Because the covariates are highly correlated, the ranking is descriptive evidence, and should be interpreted as motivating evidence for future research, rather than causal claims.7 [PI… view at source ↗

read the original abstract

Automation affects the labour content of work differently across different contexts. Yet, most existing exposure measures assign fixed scores to tasks or occupations, limiting comparisons of automation exposure across countries. We develop a task-based and country-specific approach to classify automation exposure across the world to disentangle labor-substituting from labor-augmenting automation, the relevant technology channel, and the material role of AI. Our measure spans 124 countries, generating an atlas of 2.33 million task-country labels for economies covering 99% of world population and GDP. We present five descriptive results. First, exposure is highly uneven, ranging from 3.3% of tasks in South Sudan to 61.6% in China, and rises strongly with income, although substantial variation remains within income groups. Second, across countries, exposed tasks are skewed towards substitution rather than augmentation, but low-income countries are disproportionately exposed to substitution, whereas middle-income countries are more heterogeneous. Third, less technologically advanced forms of automation account for more than half of exposed tasks in low-income countries but about one quarter in high-income countries; while other more complex channels generally rise with income levels. Fourth, AI tends to be less prevalent in simpler channels of automation, but also more prevalent in labour-substituting margins in lower income settings and to augment labour in higher income settings. Fifth, we find that females seem to be disproportionately more exposed to labour-substituting automation than males. Our methodology provides a basis for comparing automation exposure across development stages, linking it with cross-country data and allowing us to treat exposure levels, labour margins, technological channels and AI involvement as separate dimensions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A country-specific automation exposure atlas that splits substitution, augmentation, channels and AI is a practical extension, though the cross-country labeling rules need close checking for consistency.

read the letter

The paper's main advance is building a task-based atlas that varies automation exposure by country instead of using fixed scores, while separating labor-substituting from labor-augmenting effects, different technology channels, and AI involvement. It produces 2.33 million task-country labels for 124 countries that cover 99 percent of world population and GDP, then reports five descriptive patterns: exposure rises with income but varies within groups, low-income places see more substitution skew, simpler automation dominates there while complex channels grow with income, AI shows distinct patterns by margin and income level, and females appear more exposed to substitution than males. This setup lets researchers treat exposure levels, margins, channels, and AI as separate dimensions and link them to other cross-country data, which is a clear improvement over prior one-size-fits-all measures for development and labor work. The approach is straightforward in principle and the scale is ambitious. The main soft spot is whether the classification rules hold up across income levels. If the task inventory starts from high-income sources like O*NET and the substitution-versus-augmentation or AI tags are not adjusted enough for differences in skills, infrastructure, and complementary capital, then the reported gradients could partly reflect how the rules were applied rather than real economic differences. The abstract gives little detail on validation or robustness steps, so it is hard to judge how much this matters. This work is aimed at labor economists and development researchers who need comparable automation metrics across stages of development. It deserves peer review so the labeling methodology and any robustness checks can be examined directly.

Referee Report

3 major / 2 minor

Summary. The paper develops a task-based, country-specific classification of automation exposure for 124 countries, producing an atlas of 2.33 million task-country labels covering 99% of world population and GDP. It separates labor-substituting from labor-augmenting automation, identifies technology channels, and assesses AI involvement. Five descriptive findings are presented: exposure ranges from 3.3% (South Sudan) to 61.6% (China) and rises with income; exposed tasks skew toward substitution, especially in low-income countries; less advanced automation dominates in low-income settings while complex channels rise with income; AI patterns vary by channel and income; and females are more exposed to substituting automation.

Significance. If the classifications prove robust, the atlas would supply a valuable new granular dataset for cross-country research on automation, enabling separate analysis of exposure levels, substitution/augmentation margins, technology channels, and AI. The global coverage and explicit separation of dimensions are strengths for linking to other economic data. The work is primarily descriptive and measurement-oriented rather than deriving parameter-free results or machine-checked proofs.

major comments (3)

[Methods] Methods section: The operationalization of country-specific adjustments to the underlying task inventory (likely O*NET-derived) for distinguishing substitution from augmentation is insufficiently detailed. It is unclear how rules are adapted when infrastructure, skills, and complementary capital differ sharply (e.g., South Sudan vs. China), raising the risk that reported gradients—such as the substitution skew in low-income countries—reflect measurement artifacts rather than genuine differences.
[Results] Results, second and third findings: The claims that low-income countries are disproportionately exposed to substitution and that less advanced automation accounts for >50% of exposed tasks there (vs. ~25% in high-income countries) are load-bearing for the central contribution. These require explicit validation steps, such as sensitivity analyses to alternative classification thresholds or cross-checks against local task data, to rule out high-income calibration bias.
[Results] Section on AI involvement (fourth finding): The reported patterns—that AI is less prevalent in simpler channels but more prevalent in labor-substituting margins in lower-income settings—depend on how AI involvement is identified in the classification. The manuscript should supply concrete examples or decision rules for this identification across income levels, as the distinction is central to disentangling technology channels.

minor comments (2)

[Abstract] Abstract: Could briefly note the primary data sources for the task inventory and any external validation performed, to give readers immediate context for the scale of 2.33 million labels.
[Figures/Tables] Figures and tables: Ensure legends explicitly define income-group cutoffs and how task exposure shares are aggregated from the 2.33 million labels.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment below, providing clarifications on our approach and indicating where revisions will strengthen the presentation and robustness of the results.

read point-by-point responses

Referee: [Methods] Methods section: The operationalization of country-specific adjustments to the underlying task inventory (likely O*NET-derived) for distinguishing substitution from augmentation is insufficiently detailed. It is unclear how rules are adapted when infrastructure, skills, and complementary capital differ sharply (e.g., South Sudan vs. China), raising the risk that reported gradients—such as the substitution skew in low-income countries—reflect measurement artifacts rather than genuine differences.

Authors: We appreciate the referee drawing attention to the need for greater transparency in the country-specific adjustments. Our approach adapts the task inventory by incorporating observable country-level proxies for infrastructure (e.g., electricity access and digital connectivity from World Bank data), skills (average years of schooling), and complementary capital (capital stock per worker from Penn World Table). These factors adjust the substitution versus augmentation classification: for example, tasks requiring stable power or advanced digital infrastructure receive a higher substitution probability in high-income settings like China than in low-infrastructure contexts like South Sudan. We agree the current Methods section could be more explicit about these rules. In the revised version we will add a detailed decision tree with concrete examples contrasting low- and high-income countries to demonstrate that the reported gradients reflect differences in complementary factors rather than artifacts. revision: yes
Referee: [Results] Results, second and third findings: The claims that low-income countries are disproportionately exposed to substitution and that less advanced automation accounts for >50% of exposed tasks there (vs. ~25% in high-income countries) are load-bearing for the central contribution. These require explicit validation steps, such as sensitivity analyses to alternative classification thresholds or cross-checks against local task data, to rule out high-income calibration bias.

Authors: We concur that these two findings are central and that additional validation would be valuable. The thresholds for substitution/augmentation and technology-channel complexity are derived from task content combined with the country proxies described above. In the revision we will add sensitivity analyses that perturb the classification thresholds by ±10–20 percent and show that the key income gradients remain qualitatively unchanged. We have also conducted limited cross-checks with available national occupational surveys for a small number of middle- and high-income countries; these support the patterns. However, comparable granular local task data do not exist for the majority of low-income countries in our sample, which limits the scope of such checks. We will document the available cross-checks in a new appendix while noting this data constraint. revision: partial
Referee: [Results] Section on AI involvement (fourth finding): The reported patterns—that AI is less prevalent in simpler channels but more prevalent in labour-substituting margins in lower-income settings—depend on how AI involvement is identified in the classification. The manuscript should supply concrete examples or decision rules for this identification across income levels, as the distinction is central to disentangling technology channels.

Authors: We thank the referee for underscoring the importance of transparent AI identification. AI involvement is assigned when a task’s automation relies on adaptive learning, pattern recognition, or predictive modeling, following distinctions in the recent literature between rule-based and machine-learning-based technologies. Examples already in the manuscript include non-AI robotic process automation for routine data entry versus AI-augmented diagnostic support in higher-income contexts. To make the rules fully explicit, we will insert a new table in the revised manuscript listing decision criteria and multiple concrete examples differentiated by channel, substitution/augmentation margin, and income group. This addition will allow readers to evaluate the fourth finding more directly. revision: yes

standing simulated objections not resolved

Comprehensive cross-checks against local task data for all 124 countries, especially the lowest-income ones, because uniformly available granular national task inventories do not exist for the full sample.

Circularity Check

0 steps flagged

No circularity: new classification atlas is self-contained measurement construction

full rationale

The paper develops and applies a task-based, country-specific classification procedure to generate 2.33 million task-country labels across 124 economies. No equations, fitted parameters, or predictive derivations are described that reduce by construction to the authors' own prior inputs or self-citations. The five descriptive results follow directly from applying the stated classification rules to external task inventories and country data; the work is therefore a measurement exercise whose outputs are not forced by internal redefinitions or load-bearing self-references.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the exact free parameters, axioms, and invented entities used in the classification rules cannot be identified. The work implicitly relies on the assumption that tasks can be consistently labeled by substitution/augmentation status and technology channel across countries.

axioms (1)

domain assumption Tasks can be classified into labor-substituting versus labor-augmenting and into distinct technology channels including AI involvement using available task descriptions and country context.
This classification rule is required to generate the 2.33 million labels and the five descriptive results.

pith-pipeline@v0.9.0 · 5830 in / 1249 out tokens · 73046 ms · 2026-05-20T14:59:54.129470+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We develop a task-based and country-specific approach to classify automation exposure across the world to disentangle labor-substituting from labor-augmenting automation, the relevant technology channel, and the material role of AI. Our measure spans 124 countries, generating an atlas of 2.33 million task-country labels

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

Abebe, G., Caria, S., Fafchamps, M., Falco, P., Franklin, S., and Quinn, S. (2021). Anonymity or distance? job search and labour market exclusion in a growing african city.The Review of Economic Studies, 88(3):1279–1310. Acemoglu, D., Aghion, P., and Zilibotti, F. (2006). Distance to frontier, selection, and economic growth. Journal of the European Econom...

work page 2021
[2]

and Van Reenen, J

Bloom, N. and Van Reenen, J. (2007). Measuring and explaining management practices across firms and countries.The Quarterly Journal of Economics, 122(4):1351–1408. BLS (2024a). Employment projections, national industry–occupation matrix. Industry–occupation staffing patterns used for occupation-mediated industry weights. BLS (2024b). Occupational employme...

work page 2007
[3]

and Rodrik, D

Mazzucato, M. and Rodrik, D. (2026). Industrial policy with conditionalities: A taxonomy and sample cases. Industrial and Corporate Change. Article dtaf063. McElheran, K., Li, J. F., Brynjolfsson, E., Kroff, Z., Dinlersoz, E., Foster, L. S., and Zolas, N. (2024). AI adoption in america: Who, what, and where.Journal of Economics & Management Strategy, 33(2...

work page 2026
[4]

Vaccaro, M., Almaatouq, A., and Malone, T

United Nations, Department of Economic and Social Affairs, Statistics Division, New York. Vaccaro, M., Almaatouq, A., and Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis.Nature Human Behaviour, 8:2293–2303. Webb, M. (2020). The impact of artificial intelligence on the labor market. Working paper; SSR...

work page 2024
[5]

World development indicators

World Bank (2024b). World development indicators. Data retrieved 2026 via the World Bank API. World Bank (2024c). Worldwide governance indicators. Yin, M., Vu, H., and Persico, C. (2026). How (un)stable are LLM occupational exposure scores? evidence from multi-model replication. Working Paper 35110, National Bureau of Economic Research. 25 Supplementary M...

work page 2026
[6]

26 The ISCO layer is then built by mapping SOC occupations to ISCO groups under bridge shares mok: ˆE(w) k = ∑ o m(w) ok ˆEo, ˆE(m) k = ∑ o m(m) ok ˆEo

The SOC-level occupation summary is ˆEo = ∑ t∈T(o) wtoEt. 26 The ISCO layer is then built by mapping SOC occupations to ISCO groups under bridge shares mok: ˆE(w) k = ∑ o m(w) ok ˆEo, ˆE(m) k = ∑ o m(m) ok ˆEo. We report the task-weighted variant as the baseline and retain the modal variant as a sensitivity check. Analogous summaries are produced for chan...

work page 2026
[7]

The context-free benchmark uses BENCHMARK CONTEXT: Generic ordinary workplace 28 benchmark and removes geography and income from the prompt

The income-group benchmark replaces that line withINCOME GROUP CONTEXT: {income_group} and asks for a typical country in the named World Bank income group. The context-free benchmark uses BENCHMARK CONTEXT: Generic ordinary workplace 28 benchmark and removes geography and income from the prompt. In all three variants, only the context changes. Schema summ...

work page 2017
[8]

Manufacturing

High income Upper middle income Lower middle income Low income Notes: Points are countries with coverage in both datasets. The x-axis is the IMF AIPI composite for 2023; the y-axis is country-level AI-material share. The dotted line is a linear fit. Pearson correlation is0.90and Spearman correlation is0.93. Colours mark World Bank income groups; Table A.6...

work page 2023
[9]

not widely deployed

Level exact Level ±1 Channel Margin (exposed) P1 vs P2 P1 vs P3 P2 vs P3 0 20 40 60 80 100Agreement (%) 84% 100% 83% 100% 82% 100% (b) Pairwise paraphrase agreement Exact level match Adjacent (±1) All 3 agree (exact level) All 3 within 1 level of each other 0 20 40 60 80 100Share of tasks (%) 74.7% 99.8% (c) Joint within-task stability Notes: Three paraph...

work page 2017
[10]

Panel (b) reports each country’s Spearman rank correlation with the US-conditioned occupation ranking

n=19 =4.6/10 n=35 =5.5/10 n=27 =6.1/10 n=41 =7.0/10 (c) Same-construct country shift in top-10 exposed occupations Notes: Panel (a) decomposes variance in the124country×923occupation exposure matrix into occupation, country, and country–occupation components. Panel (b) reports each country’s Spearman rank correlation with the US-conditioned occupation ran...

work page 2010
[11]

For each income group, the tables report the top entries selected by exposed share multiplied by the relevant labour-margin share among exposed tasks

57 Supplementary Note B.5 Top substitution and augmentation occupations and industries by income group This subsection lists the occupation and industry rankings behind the main-text discussion of substitution-only and augmentation-only exposure. For each income group, the tables report the top entries selected by exposed share multiplied by the relevant ...

work page 2024
[12]

Barro–Lee educational-attainment data (Barro and Lee,

2019 Years of schooling Average years of schooling among adults aged 15–64. Barro–Lee educational-attainment data (Barro and Lee,

work page 2019
[13]

Penn World Table 10.01 (Feenstra et al.,

2015 Capital intensity Natural log of real capital stock per worker,log(rkna/emp), whereempis persons engaged. Penn World Table 10.01 (Feenstra et al.,

work page 2015
[14]

World Development Indicators (World Bank, 2024b) Latest non-missing, 2018–2024 Government effectiveness Government effectiveness percentile rank

2019 Investment (% GDP) Gross fixed capital formation as a percentage of GDP. World Development Indicators (World Bank, 2024b) Latest non-missing, 2018–2024 Government effectiveness Government effectiveness percentile rank. Worldwide Governance Indicators (World Bank, 2024c) 2024 Regulatory quality Regulatory quality percentile rank. Worldwide Governance ...

work page 2019
[15]

The main 68-country random-forest specification uses all rows in the table

and World Bank GDP (World Bank, 2024b) 2023 trade; latest GDP Notes: The labels match the country-covariate figures. The main 68-country random-forest specification uses all rows in the table. The wider-coverage 90-country specification drops human capital, capital intensity, and investment to reduce complete-case sample loss. PWT variables are constructe...

work page 2023
[16]

Colours indicate the sign of the one-dimensional accumulated local effect

Outcomes are exposed share among all tasks and substitution-only or augmentation-only shares within exposed tasks. Colours indicate the sign of the one-dimensional accumulated local effect. This companion is reported because unconditional permutation importance can be sensitive when predictors are correlated. Figure B.11:Linear Shapley companion for the c...

work page 2015

[1] [1]

Abebe, G., Caria, S., Fafchamps, M., Falco, P., Franklin, S., and Quinn, S. (2021). Anonymity or distance? job search and labour market exclusion in a growing african city.The Review of Economic Studies, 88(3):1279–1310. Acemoglu, D., Aghion, P., and Zilibotti, F. (2006). Distance to frontier, selection, and economic growth. Journal of the European Econom...

work page 2021

[2] [2]

and Van Reenen, J

Bloom, N. and Van Reenen, J. (2007). Measuring and explaining management practices across firms and countries.The Quarterly Journal of Economics, 122(4):1351–1408. BLS (2024a). Employment projections, national industry–occupation matrix. Industry–occupation staffing patterns used for occupation-mediated industry weights. BLS (2024b). Occupational employme...

work page 2007

[3] [3]

and Rodrik, D

Mazzucato, M. and Rodrik, D. (2026). Industrial policy with conditionalities: A taxonomy and sample cases. Industrial and Corporate Change. Article dtaf063. McElheran, K., Li, J. F., Brynjolfsson, E., Kroff, Z., Dinlersoz, E., Foster, L. S., and Zolas, N. (2024). AI adoption in america: Who, what, and where.Journal of Economics & Management Strategy, 33(2...

work page 2026

[4] [4]

Vaccaro, M., Almaatouq, A., and Malone, T

United Nations, Department of Economic and Social Affairs, Statistics Division, New York. Vaccaro, M., Almaatouq, A., and Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis.Nature Human Behaviour, 8:2293–2303. Webb, M. (2020). The impact of artificial intelligence on the labor market. Working paper; SSR...

work page 2024

[5] [5]

World development indicators

World Bank (2024b). World development indicators. Data retrieved 2026 via the World Bank API. World Bank (2024c). Worldwide governance indicators. Yin, M., Vu, H., and Persico, C. (2026). How (un)stable are LLM occupational exposure scores? evidence from multi-model replication. Working Paper 35110, National Bureau of Economic Research. 25 Supplementary M...

work page 2026

[6] [6]

26 The ISCO layer is then built by mapping SOC occupations to ISCO groups under bridge shares mok: ˆE(w) k = ∑ o m(w) ok ˆEo, ˆE(m) k = ∑ o m(m) ok ˆEo

The SOC-level occupation summary is ˆEo = ∑ t∈T(o) wtoEt. 26 The ISCO layer is then built by mapping SOC occupations to ISCO groups under bridge shares mok: ˆE(w) k = ∑ o m(w) ok ˆEo, ˆE(m) k = ∑ o m(m) ok ˆEo. We report the task-weighted variant as the baseline and retain the modal variant as a sensitivity check. Analogous summaries are produced for chan...

work page 2026

[7] [7]

The context-free benchmark uses BENCHMARK CONTEXT: Generic ordinary workplace 28 benchmark and removes geography and income from the prompt

The income-group benchmark replaces that line withINCOME GROUP CONTEXT: {income_group} and asks for a typical country in the named World Bank income group. The context-free benchmark uses BENCHMARK CONTEXT: Generic ordinary workplace 28 benchmark and removes geography and income from the prompt. In all three variants, only the context changes. Schema summ...

work page 2017

[8] [8]

Manufacturing

High income Upper middle income Lower middle income Low income Notes: Points are countries with coverage in both datasets. The x-axis is the IMF AIPI composite for 2023; the y-axis is country-level AI-material share. The dotted line is a linear fit. Pearson correlation is0.90and Spearman correlation is0.93. Colours mark World Bank income groups; Table A.6...

work page 2023

[9] [9]

not widely deployed

Level exact Level ±1 Channel Margin (exposed) P1 vs P2 P1 vs P3 P2 vs P3 0 20 40 60 80 100Agreement (%) 84% 100% 83% 100% 82% 100% (b) Pairwise paraphrase agreement Exact level match Adjacent (±1) All 3 agree (exact level) All 3 within 1 level of each other 0 20 40 60 80 100Share of tasks (%) 74.7% 99.8% (c) Joint within-task stability Notes: Three paraph...

work page 2017

[10] [10]

Panel (b) reports each country’s Spearman rank correlation with the US-conditioned occupation ranking

n=19 =4.6/10 n=35 =5.5/10 n=27 =6.1/10 n=41 =7.0/10 (c) Same-construct country shift in top-10 exposed occupations Notes: Panel (a) decomposes variance in the124country×923occupation exposure matrix into occupation, country, and country–occupation components. Panel (b) reports each country’s Spearman rank correlation with the US-conditioned occupation ran...

work page 2010

[11] [11]

For each income group, the tables report the top entries selected by exposed share multiplied by the relevant labour-margin share among exposed tasks

57 Supplementary Note B.5 Top substitution and augmentation occupations and industries by income group This subsection lists the occupation and industry rankings behind the main-text discussion of substitution-only and augmentation-only exposure. For each income group, the tables report the top entries selected by exposed share multiplied by the relevant ...

work page 2024

[12] [12]

Barro–Lee educational-attainment data (Barro and Lee,

2019 Years of schooling Average years of schooling among adults aged 15–64. Barro–Lee educational-attainment data (Barro and Lee,

work page 2019

[13] [13]

Penn World Table 10.01 (Feenstra et al.,

2015 Capital intensity Natural log of real capital stock per worker,log(rkna/emp), whereempis persons engaged. Penn World Table 10.01 (Feenstra et al.,

work page 2015

[14] [14]

World Development Indicators (World Bank, 2024b) Latest non-missing, 2018–2024 Government effectiveness Government effectiveness percentile rank

2019 Investment (% GDP) Gross fixed capital formation as a percentage of GDP. World Development Indicators (World Bank, 2024b) Latest non-missing, 2018–2024 Government effectiveness Government effectiveness percentile rank. Worldwide Governance Indicators (World Bank, 2024c) 2024 Regulatory quality Regulatory quality percentile rank. Worldwide Governance ...

work page 2019

[15] [15]

The main 68-country random-forest specification uses all rows in the table

and World Bank GDP (World Bank, 2024b) 2023 trade; latest GDP Notes: The labels match the country-covariate figures. The main 68-country random-forest specification uses all rows in the table. The wider-coverage 90-country specification drops human capital, capital intensity, and investment to reduce complete-case sample loss. PWT variables are constructe...

work page 2023

[16] [16]

Colours indicate the sign of the one-dimensional accumulated local effect

Outcomes are exposed share among all tasks and substitution-only or augmentation-only shares within exposed tasks. Colours indicate the sign of the one-dimensional accumulated local effect. This companion is reported because unconditional permutation importance can be sensitive when predictors are correlated. Figure B.11:Linear Shapley companion for the c...

work page 2015