Buffer or bottleneck? Employment exposure to generative AI and the digital divide in Latin America

Gmyrek, Paweł; Winkler, Hernan; Garganta, Santiago Working Paper Buffer or bottleneck? Employment exposure to generative AI and the digital divide in Latin America ILO Working Paper, No. 121 Provided in Cooperation with: International Labour Organization (ILO), Geneva Suggested Citation: Gmyrek, Paweł; Winkler, Hernan; Garganta, Santiago (2024) : Buffer or bottleneck? Employment exposure to generative AI and the digital divide in Latin America, ILO Working Paper, No. 121, ISBN 978-92-2-041003-5, International Labour Organization (ILO), Geneva, https://doi.org/10.54394/TFZY7681 This Version is available at: https://hdl.handle.net/10419/302848 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/3.0/igo/ XBuffer or Bottleneck? Employment Exposure to Generative AI and the Digital Divide in Latin America Authors / Paweł Gmyrek, Hernan Winkler, Santiago Garganta July / 2024 ILO Working Paper 121 Copyright © International Labour Organization and the World Bank 2024 This is an open access work distributed under the Creative Commons Attribution 3.0 IGO License (http://creativecommons.org/licenses/by/3.0/igo). Users can reuse, share, adapt and build upon the original work, as detailed in the License. The ILO and The World Bank must be clearly credited as the owners of the original work. The use of the emblem of the ILO and The World Bank is not permitted in connection with users’ work. Attribution – The work must be cited as follows: Gmyrek, P., Winkler, H., Garganta, S. Buffer or Bottleneck? Employment Exposure to Generative AI and the Digital Divide in Latin America. ILO Working Paper 121. Geneva: International Labour Office and The World Bank, 2024. Translations – In case of a translation of this work, the following disclaimer must be added along with the attribution: This translation was not created by the International Labour Organization (ILO) or The World Bank and should not be considered an official ILO or World Bank translation. The ILO and The World Bank are not responsible for the content or accuracy of this translation. Adaptations – In case of an adaptation of this work, the following disclaimer must be added along with the attribution: This is an adaptation of an original work by the International Labour Organization (ILO) and The World Bank. Responsibility for the views and opinions expressed in the adaptation rests solely with the author or authors of the adaptation and are not endorsed by the ILO or The World Bank. This CC license does not apply to non-ILO or World Bank copyright materials included in this publication. If the material is attributed to a third party, the user of such material is solely responsible for clearing the rights with the right holder. Any dispute arising under this license that cannot be settled amicably shall be referred to arbitra tion in accordance with the Arbitration Rules of the United Nations Commission on International Trade Law (UNCITRAL). The parties shall be bound by any arbitration award rendered as a result of such arbitration as the final adjudication of such a dispute. All queries on rights and licensing should be addressed to the ILO Publishing Unit (Rights and Licensing), 1211 Geneva 22, Switzerland, or by email to [email protected]. ISBN 9789220410028 (print), ISBN 9789220410035 (web PDF), ISBN 9789220410042 (epub), ISBN 9789220410066 (mobi), ISBN 9789220410059 (html). ISSN 2708-3438 (print), ISSN 2708-3446 (digital) https://doi.org/10.54394/TFZY7681 The designations employed in ILO and World Bank publications and the presentation of material therein do not imply the expression of any opinion whatsoever on the part of the ILO and The World Bank concerning the legal status of any country, area or territory or of its authorities, or concerning the delimitation of its frontiers or boundaries. Details at www.ilo.org/disclaimer This article is a product of the staff of the World Bank and the ILO. It has been released both in The World Bank Policy Research Working Paper Series and the ILO Working Paper Series. The responsibility for opinions expressed in signed articles, studies and other contributions rests solely with their authors, and publication does not constitute an endorsement by the ILO or The World Bank of the opinions expressed in them. Reference to names of firms and commercial products and processes does not imply their endorsement by the ILO or The World Bank, and any failure to mention a particular firm, commercial product or process is not a sign of disapproval. Information on ILO publications and digital products can be found at: https://www.ilo.org/ research-and-publications ILO Working Papers summarize the results of ILO research in progress, and seek to stimulate discussion of a range of issues related to the world of work. Comments on this ILO Working Paper are welcome and can be sent to [email protected]. Authorization for publication: Richard Samans, Director, Research Department ILO Working Papers can be found at: https://www.ilo.org/research-and-publications#workingpapers Suggested citation: Gmyrek, P., Winkler, H., Garganta, S. 2024. Buffer or Bottleneck? Employment Exposure to Generative AI and the Digital Divide in Latin America, ILO Working Paper 121 (Geneva, ILO and The World Bank, 2024). https://doi.org/10.54394/TFZY7681 01 ILO Working Paper 121 Abstract Empirical evidence on the potential impacts of generative artificial intelligence (GenAI) is mostly focused on high-income countries. In contrast, little is known about the role of this technology on the future economic pathways of developing economies. This paper contributes to fill this gap by estimating the exposure of the Latin American labour market to GenAI. It provides detailed statistics of GenAI exposure between and within countries by leveraging a rich set of harmonized household and labour force surveys. To account for the slower pace of technology adoption in developing economies, it adjusts the measures of exposure to GenAI by using the likelihood of accessing digital technologies at work. This is then used to assess the extent to which the digital divide across and within countries will be a barrier to maximize the productivity gains among occupations that could otherwise be augmented by GenAI tools. The findings show that certain characteristics are consistently correlated with higher exposure. Specifically, urban-based jobs that require higher education, are situated in the formal sector, and are held by individuals with higher incomes are more likely to come into interaction with this technology. Moreover, there is a pronounced tilt toward younger workers facing greater exposure, including the risk of job automation, particularly in the finance, insurance, and public administration sectors. When adjusting for access to digital technologies, the findings show that the digital divide is a major barrier to realizing the positive effects of GenAI on jobs in the region. In particular, nearly half of the positions that could potentially benefit from augmentation are hampered by lack of use of digital technologies. This negative effect of the digital divide is more pronounced in poorer countries. About the authors Paweł Gmyrek is a Senior Researcher at the Research Department of the ILO. Hernan Winkler is a Senior Economist at the World Bank Poverty and Equity Global Practice for Latin America and the Caribbean. Santiago Garganta is a Senior Researcher at the Center for Distributive, Labor and Social Studies (CEDLAS) of the National University of La Plata (UNLP). 02 ILO Working Paper 121 Abstract 01 About the authors 01 Acronyms 05 XIntroduction 06 X1 LAC region and the theoretical effects of GenAI 08 X2 Methods 15 Occupational exposure to GenAI 15 Use of a computer at work 19 X3 Findings 22 Cross-country comparisons of the levels of exposure 22 Impact of digital infrastructure on the potential of transformation 26 Within-country patterns 29 Which occupations drive the effects? 30 Differential exposure across earnings levels 32 XFinal discussion 35 Appendix 38 References 45 Acknowledgements 50 Table of contents 03 ILO Working Paper 121 List of Figures Figure 1. GDP per capita, population and income status of LAC countries in the sample 08 Figure 2. Automation and augmentation potential: LAC vs other regions 09 Figure 3. Internet coverage vs per capita income: global and LAC 11 Figure 4. Occupations in the LAC region, by ISCO 1-digit and gender 13 Figure 5. Coverage of ISCO-08 4-digit microdata in SEDLAC (WB) and ILO harmonized microdata collection 17 Figure 6. Hierarchical clustering based on ISCO 2-digit shares, GDP(PPP) and total population 18 Figure 7. Total exposure to GenAI by country 23 Figure 8. Automation potential - detailed breakdown of socio-economic characteristics 24 Figure 9. Augmentation potential - detailed breakdown of socio-economic characteristics 25 Figure 10. Jobs with augmentation potential and access to computer at work, based on PIAAC data 27 Figure 11. Exposure by country, exposure type and access to digital infrastructure 28 Figure 12. Exposure by country, type and detailed country-level characteristics 30 Figure 13. ISCO 2-digit occupations by type of exposure and country (share of exposure > 25%) 31 Figure 14. Earnings of occupations exposed to GenAI, by employment status (exposure above 25%) 33 Figure A 1. Comparison of TechXposure scores vs GBB scores (mean by occupation, z-scores) 38 Figure A 2. Comparison of Felten et al. (2023) ML scores vs GBB scores (z-scores) 38 Figure A 3. Labour market distribution in LAC countries by ISCO-08 2-digit occupations and sex 39 Figure A 4. Ranking of countries by the type of GenAI exposure 40 Figure A 5. Comparison of results on computer use between PIAAC (at work) and SEDLAC (at home) - augmentation category 40 Figure A 6. Jobs in augmentation category that do not use a computed at work: totals by country 40 04 ILO Working Paper 121 List of Tables Table 1. Distribution of AI Exposure by Demographic and Socioeconomic Categories in SEDLAC Data 19 Table A 1. Individual SEDLAC observations by country and year 41 Table A 2. Estimated coefficients of computer use at work from PIAAC 41 Table A 3. Results of the pooled OLS with all individual observations, with country-level normalized population weights 43 05 ILO Working Paper 121 Acronyms EM Emerging Markets GDP Gross Domestic Product GBB Gmyrek, Berg and Bescond (as used in your study for citation) GenAI Generative AI GPT-4 Generative Pre-trained Transformer 4 HIC High Income Countries ILO International Labour Organization IMF International Monetary Fund ISCO International standard Classification of Occupations ISCO-08 International Standard Classification of Occupations, 2008 version LAC Latin America and the Caribbean LLM Large Language Models OECD Organization for Economic Cooperation and Development PIAAC Programme for the International Assessment of Adult Competencies PPP Purchasing Power Parity SEDLAC Socio-Economic Database for Latin America and the Caribbean TFP Total Factor Productivity US United States WB World Bank WEF World Economic Forum 12 ILO Working Paper 121 Fifth, the results of these recent experiments and macroeconomic models do not consider general equilibrium or second order effects on employment. For example, while increased productivity may bring employment and wage gains in sectors facing a consumer demand that is growing rapidly, that may not be the case for sectors facing a more stable consumer demand (Autor, 2024). The nature of these second order effects is likely to be different across countries. In developing economies with a large fraction of the workforce in the informal sector, and where technology adoption and private sector investment are typically concentrated among a small share of formal firms (Cirera and Cruz, 2022), workers displaced from formal sector jobs may face more challenges finding high quality jobs than their counterparts in high-income countries. While detailed macroeconomic modelling of such effects is beyond the scope of our study, the estimates of jobs’ exposure to GenAI presented in this paper provide a profile of the socio-economic groups more likely to experience the first-order impacts. Historically, together with Sub-Saharan Africa, LAC is one of the most unequal regions in the world (World Bank, 2016a), with levels of income inequality strongly influenced by the changes in the structure of the labour market (Azevedo et al., 2013). Concerns about the impacts of new technologies on inequality in LAC are consistent with broader empirical evidence about the effects of recent waves of technological change on labour demand, which tended to be skill-biased and to widen the gap between lowand high-income workers (Acemoglu and Restrepo, 2022; Autor et al., 2008). Acemoglu's (2024) most recent modelling of GenAI outcomes on wages and inequality also suggests that in nearly all theoretical scenarios, the deployment of this technology at the workplace is likely to increase the inequality between capital and labour, and result in higher income inequality between different demographic groups, with particularly negative consequences for the incomes of low-education women in the US.5 In the case of LAC countries, Dutz et al. (2018) conduct a comprehensive discussion of the several challenges for the region in terms of digital technologies and how inclusive they might be, looking at diverse case studies of technology adoption in Latin America. In this regard, the distributive impact of the AI adoption depends strongly on how the effects of increased productivity and output could overcome labour displacement conducted by substitution of technology for workers. Although the use of AI technologies in Latin America remains still very low, recent empirical evidence shows LAC labour patterns more consistent with the skill-biased technological change hypothesis than the job polarization model6 (Brambilla et al., 2023; Messina et al., 2016; Messina and Silva, 2018). Similar conclusions arise from the literature on employment and automation in the rest of the developing world (Das and Hilgenstock, 2022), although Maloney and Molina (2016) find some evidence of incipient polarization in Brazil and Mexico. To further theorize the potential effects of GenAI diffusion on inequality in the region, Figure 4 presents the most recent breakdown of LAC occupations by the highest, 1-digit level of ISCO-08,7 revealing visible differences in the employment structures across genders. 5Albeit with smaller wage effects than the previous waves of automation (see Acemoglu and Restrepo, 2022). 6The skill-biased technological change (SBTC) hypothesis suggests that technology benefits skilled workers, increasing demand for high-skill jobs and widening wage inequality. In contrast, the job polarization model posits that technology creates more high-skill and low-skill jobs, reducing middle-skill job opportunities and hollowing out the middle class. 7Elementary occupations are grouped together with agricultural, fishery and forestry work (96). 13 ILO Working Paper 121 XFigure 4. Occupations in the LAC region, by ISCO 1-digit and gender Note: The breakdowns are presented as a share of male and female employment separately and calculated as a mean share of employment across the countries in each income bracket, based on ILO modelled estimates (ILO, 2023a). For men, the largest share of employment is in the elementary, agricultural, forestry and fishery work, followed by craft and related trade workers. For women, the largest employment categories concern service and sales work, followed by elementary jobs. Among the “Service and sales workers”, the pattern is very similar across country groups, with male employment dominant only in protective services, and female employment having much higher shares in personal care, sales and personal service work. A more detailed analysis at ISCO-08 2-digit level8 reveals that – excluding IT, science and engineering professions – women are significantly more represented across all professional categories, with particular prominence in teaching, health, business administration and legal, social and cultural occupations. This trend extends into clerical work and amplifies in line with countries’ income status. This warrants attention, since recent research has identified clerical and professional job categories as being more exposed to the risks of automation with GenAI (Cazzaniga et al., 2024; Gmyrek et al., 2023; Ozdeneron, Hakki, 2023; WEF, 2023), with pre-GenAI regional assessments also classifying female-held jobs in LAC as being at a higher risk of automation from digital technologies (Egana-delSol et al., 2022). In accordance with the technical documentation of ISCO-08, such differences in the occupational structures also correspond to varying levels of skills and educational attainment, as clerical support workers, technicians and professionals are typically classified in the midto high-skill level brackets (ILO, 2023b). Given that educational attainment and earnings gaps across skills groups have been important drivers of income inequality in LAC (Azevedo et al. 2013), the impact of GenAI that follows the existing labour market structures would likely also have an effect on the overall income inequality. In the best-case scenario, GenAI would boost the productivity of lower-skilled workers in the exposed occupations, allowing them to access higher incomes and therefore leading to a more broad-based income distribution. In the worst-case scenario, the technological transition could result in the automation of largely female-held jobs in the clerical, technical and professional occupations, while the opportunities for new GenAI-augmented jobs could be limited, given the high concentration of current employment in elementary occupations 8Plot A1 in the appendix shows a full breakdown of occupations at a 2-digit level of ISCO-08, by country and gender. 14 ILO Working Paper 121 and in the informal sector, where technology adoption and private sector investment are low. To better understand how the first-order effects of GenAI may affect inequality, this study provides a detailed profile of the socio-economic groups most exposed to this technology. Finally, we acknowledge that the final outcomes of the technological transition process will also be largely dependent on the existing and future policy frameworks in the region. While the analysis of country-level polices and legal frameworks is beyond the scope of this regional study, the detailed country-level statistics that we make publicly available alongside this publication can serve as useful inputs to the discussions underpinning such policy responses.9 9Access to detailed data at: https://pgmyrek.shinyapps.io/AI_Data_Portal_Research/. 15 ILO Working Paper 121 X2 Methods Occupational exposure to GenAI We combine multiple datasets to estimate occupational exposure to AI, leveraging the distinct advantages inherent in each dataset to ensure a comprehensive analysis. We use the AI exposure scores at the 4-digit ISCO-08 level from AI scores from GBB (2023) as the principal indicator of occupational exposure to GenAI. We also consider alternative scores that could be used for this purpose, in particular the ability-based scores developed by Felten et al. (2021, 2023a, 2023b, 2018), based on the US O’NET classifications, recently linked to ISCO-08 4-digit occupations by (Cazzaniga et al., 2024), as well as the patent-based scores of exposure to digital technologies by Prytkova et al. (2024), from which a set of technological groups relevant to GenAI could be isolated. Having compared these alternatives, we find that Felten et al. scores are quite aligned with GBB in broad terms, except for catching a much wider group of Managers and Professionals as highly exposed to AI technologies.10 Since such broad coverage seems somewhat unrealistic in the context of many developing countries, we opt for the scores of GBB, which provide a direct link to ISCO-08 documentation and focus exclusively on GenAI. This choice is further reinforced by the arguments recently advanced by (Nurski and Ruer, 2024), who find that task- (GBB) and ability-based (Felten, 2023) scores render similar general results in the European context. The authors suggest the task-based analysis using GBB scores is particularly advantageous for evaluating employment impacts, since task bundles are better at representing the daily reality of occupations (see Autor, 2015), and considering the share of affected tasks offers more scope to separate the potential for job transformation or displacement due to technological advancements. Indeed, task-based approaches have been widely used in the literature for this type of analysis (Restrepo, 2023, for concrete examples see Acemoglu and Restrepo, 2022, 2018; Frey and Osborne, 2017;), including the recently increasing use of AI-generated task-level scores (Acemoglu, 2024; Eloundou et al., 2023), used as a blueprint for the GBB scoring method. In step 1, we tag occupations at 4-digit level in ISCO-08 into three categories established by GBB: “automation potential”, “augmentation potential” and “the big unknown”. We then rely on the ILO harmonized microdata collection11 to obtain the shares of employment at 4-digit level occupations for 18 countries for each of these three AI exposure categories (Figure 7). We also calculate the shares that such exposed occupations make up in the higher, 2-digit level of occupational classification. From this step, we switch to the harmonized household surveys from the SocioEconomic Database for Latin America and the Caribbean (SEDLAC) to calculate AI exposure across 10 See Figures A2 and A3 in the Appendix for a quick visual comparison of GBB scores to Felten et al. (2023) and Prytkova et al. (2024). Felten et al.’s scores cover a wider range of AI than GenAI covered by GBB. Prytkova et al. (2024) focus on tech abilities in patents, which can be far from readiness for market-level deployment. A detailed analysis of these scores at the 4-digit occupational level is available upon request. 11 Calculations of 2-digit employment from ILO Micro data repository by David Bescond, ILO STATISTICS. 16 ILO Working Paper 121 and within 16 Latin American countries.12 Our sample of SEDLAC data consists of about 900,000 individual survey observations, with details by country provided in the Appendix (Table A1). XBox 1. GBB Scores of occupational exposure to GenAI (Gmyrek et al., 2023) GBB scores were developed based on the technical documentation of ISCO-08, which contains a list of typical tasks for each of the 436 detailed occupational groups at the most detailed, 4-digit level, and which forms the basis on which national statistical labour survey reports are linked to the internationally comparably ISCO-08 standard at the ILO. GBB build on the findings of Eloundou et al. (2023), who demonstrate a close alignment of GPT4 predictions with a survey of 70 AI experts on the potential of automating occupational tasks with LLMs, and more broadly on Bubeck et al. (2023), who provide extensive tests of the model’s capabilities and demonstrate its capacity for elaborating logical links between items, resolving complex tasks and providing justifications for its decisions. Using the Application Programming Interface (API) of GPT-4, the authors designed a sequential call that loops over each of 3,123 tasks in that documentation and requested the model to assess the technical feasibility of performing a given task with GPT-4 or LLM technology of similar capabilities. The model is asked to rate tasks on a scale of 0 to 1, with 1 representing the possibility of performing a given task by the LLM in full autonomy form a human operator, and to elaborate a written justification of each score (no task received a score of 1). The scores and justifications are then reviewed for consistency and stability of predictions over time, with the written justifications reviewed by humans. Tasks with scores above 0.8 (high possibility of automation) are transformed into embeddings, with a semantic clustering algorithm applied to identify the major groups of such tasks, which are subsequently reviewed by humans. Task-level scores for each occupation are used to calculate the mean score and the standard deviation (SD) for each occupation. These two moments of distribution are subsequently used to elaborate a theoretical framework for further classification of scores. Occupations (i) with a high mean (µi > 0.6) and a high difference between the mean and SD (µi - σi > 0.5) are classified as jobs with a high automation potential. Occupations with a low mean score (µi < 0.4) and a high sum of the mean and SD (µi + σi > 0.6) are considered to have a high potential for augmentation, meaning that while some of their tasks could be automated, the human role remains crucial for the majority of their tasks. Occupations between these two categories are classified as “the big unknown”, since, depending on the progress of technology and the use of adjacent technological applications (e.g. LLM-based agents), they could fall closer to automation or augmentation. Remaining occupations are classified as not affected, with the understanding that GenAI in its current form would have minimal or no impact on their tasks. The scores and individual task distributions are visualized by the authors through a publicly available interactive app: https://pgmyrek.shinyapps.io/AI_Data_Portal_Research/. While GBB calculate separate scores for highand low-income countries, the results are very similar and thereby only the high-income ones are used for all countries regardless 12 SEDLAC is produced by the University of La Plata’s center for Distributional, Labor and Social Studies (CEDLAS) and The World Bank’s Equitable Growth, Finance and Institutions LCR-POV-Poverty and Equity Group (ELCPV). This project aims to improve the comparability of social and economic statistics across 25 countries in the Latin American and the Caribbean (LAC) region. This involves the harmonization of household survey variables in eight categories: income, demographics, education, employment, infrastructure, durable goods, services, and aggregate welfare. 17 ILO Working Paper 121 of their income levels. See Appendix for a comparison of GBB scores to Felten et al. (2023) and Prytkova et al. (2024). See Gmyrek et al. (2023) for a detailed description of the score generating process. The advantage of SEDLAC database is that it contains a host of harmonized variables at the individual level, including the income aggregates used to measure poverty, as well as demographic characteristics and labour market outcomes. In step 2, we impute the AI exposure scores from GBB to individual respondent data in SEDLAC, using the ISCO-08 occupation reported in the household survey. Such imputation is straightforward for the 8 countries with 4-digit ISCO-08 occupations in SEDLAC, and for which we can directly compare the calculations to the estimates from the ILO as an additional validation measure (Figure 5). In contrast, there are 8 countries in SEDLAC with 2-digit ISCO-08 scores where the imputation is less obvious and depends on other circumstances. We have two types of such cases. XFigure 5. Coverage of ISCO-08 4-digit microdata in SEDLAC (WB) and ILO harmonized microdata collection When we have the 4-digit ISCO-08 employment structure from ILO for the same country, we use the estimates of the shares of exposure at the 2-digit level calculated in step 2 above.13 When we do not have the 4-digit employment structure from a different source for the same country, we use that of a “similar” country.14 These country similarities are defined in step 3, by applying a hierarchical clustering algorithm to several country-level characteristics including the full breakdown of 2-digit ISCO-08 employment shares, GDP per capita (PPP) and total population (Figure 6). 13 This concerns Brazil, Colombia, Costa Rica and Mexico. 14 This concerns Argentina, Bolivia, Guatemala and Nicaragua. 18 ILO Working Paper 121 XFigure 6. Hierarchical clustering based on ISCO 2-digit shares, GDP(PPP) and total population In step 4, we impute the estimated shares of automation, augmentation and the big unknown to individual responses at the 2-digit ISCO-08 occupation level in SEDLAC. Having a data frame with 2-digit level shares enables aggregation of individual responses by main categories of interest captured in SEDLAC microdata. We focus on gender (male, female), area (rural, urban), age (15-14, 25-34, 35-44, 45-54, 55-64), education (low, medium, high), poverty status (non-poor, poor), income quintiles (Q1 through Q5), formality (legal, productive), labour relationships (employer, salaried employee, self-employed, family worker without salary) and sector of economic activity. The shares of exposure are calculated in such a way that automation, augmentation, big unknown and other occupations add to 100 percent within each category. This means that we can interpret such results as a share of employment in each type of AI exposure within each grouping category (for example, shares of automation, augmentation, big unknown and other occupations among people with low education or among those belonging to the age bracket of 35-44). Table 1 describes these variables in more detail. 19 ILO Working Paper 121 XTable 1. Distribution of AI Exposure by Demographic and Socioeconomic Categories in SEDLAC Data15 Variable name Description Education Low: fewer than 9 years of education Middle: 9 to 13 years of education High: 14 or more years of education Poverty An individual is considered poor (non-poor) if they live in a household whose income per capita is below (above) the poverty line for upper middle-income countries (US$ 6.85-a-day in purchasing power parity terms) Income quintiles Q1 through Q5, by whether the individual’s household income per capita is in said quintile. Formality (legal) A salaried worker is informal if they do not have the right to a pension linked to employment when retired Formality (productive) An individual is considered an informal worker if they belong to any of the following categories: (i) unskilled self-employed, (ii) salaried worker in a small private firm, (iii) zero-income worker. Unskilled workers are all individuals without a tertiary or superior education degree. Small firms are those with 5 or fewer employees. These criteria and definitions refer to individuals’ main job. Sector of economic activity Primary sector Low-tech manufacturing (food, beverages, tobacco, textiles and clothing) Other manufacturing Construction Retail, restaurants, hotels and repairs Utilities, transport and communications Banking, finance, insurance, professional services Public administration Education, health and personals services Domestic service Use of a computer at work The method applied so far enables detailed insights into country-level data on exposure of occupations to GenAI, with further breakdowns by demographic and socioeconomic characteristics of the affected groups (Table 1). At the same time, the variation in AI exposure across and within countries is only driven by the variation in the occupational structures, because the same occupation in different countries uses the same score of GenAI exposure. To address this limitation, in the next step we introduce cross-country variation of occupation-level scores, by accounting for the variability in the use of computer equipment in the same occupation located in different national contexts. 15 For more details see https://www.cedlas.econo.unlp.edu.ar/wp/wp-content/uploads/Methodological_Guide_v201404.pdf. 20 ILO Working Paper 121 We first proceed by imputing the GenAI exposure measures at the 4-digit ISCO08 to the microdata from the Programme for the International Assessment of Adult Competencies (PIAAC) collected by the OECD. These surveys include rich information on detailed tasks carried out by people at work, such as whether workers use a computer (and internet)16 at work. Using this binary indicator, we split each group of GenAI exposure into those who use a computer at work, and those who do not. Not using a computer at work means that even if the worker is in an occupation that is exposed to GenAI augmentation, such potential productivity gains are unlikely to realize given the lack of access to digital infrastructure. We first implement this exercise for the four countries in the LAC region (Chile, Ecuador, Mexico and Peru) and two developed economies (Slovenia and New Zealand) included in the PIAAC dataset.17 Since there are only four Latin American countries in PIAAC, we extrapolate the measures of computer use at work from PIAAC to the full set of countries in the SEDLAC database. In particular, we estimate a predictive model for the probability of computer use at the individual level using the full set of countries in PIAAC18 and independent variables that are available both in the PIAAC and SEDLAC databases.19 We then use the estimated model and the set of independent variables to predict the probability of computer use in the SEDLAC database. More specifically, we first estimate the following Logit model in PIAAC: ()() computer fISCOage female eduinternetbroadband Pr =1 =,,,,GDP ,, ci ci o ci a ci ci cc c , , , ,, (1) Where computerci , is a binary variable equal to 1 if individual i in country c uses a computer at work; IS CO ci o , is a vector of 39 dummy variables for each 2-digit ISCO08 occupation20; ageci a , is a vector of 4 dummy variables indicating age groups; educi , is a dummy variable equal to one for High School graduates; GDP c is the log of GDP per capita in 2017 US$ PPP; internetc is the rate of internet users per 100 people, and; broadbandc is the number of fixed broadband subscriptions per 100 people. Since the reference year of the PIAAC surveys varies by country, we use the corresponding year of the country-level variables (i.e. GDP, internet and broadband). These country-level variables are helpful to capture the link between the economy-wide level of digital and economic development with the level of computer use at work. In the next step, we use the estimated equation (1) to predict the probability of using a computer at work at the individual level in the SEDLAC database.21 The probability of not using a computer at work is simply 1-Pr(computer=1). When choosing the reference years of the country-level variables of the model, we use the reference year of the SEDLAC surveys. Then, the probability 16 For the main results presented in the paper, we use variable “g_q04”, which contains the response to the following question: “Do you use a computer in your job?/Did you use a computer in your last job?”. We also do robustness checks by creating a binary variable equal to 1 when the worker uses both a computer and internet at work, by using the variables “g_q05a”, “g_q05c”, “g_q05d” and “g_q05h”, which contain information about the frequency (i.e. “never”, “less than once a month”, “less than once a week but at least once a month”, “at least once a week but not every day”, or “every day”) of internet use for mail, work related information, conduct transactions and participate in video calls, respectively. More specifically, when the worker responds that he or she “never” uses the internet for any of those four reasons at work, we consider that the worker does not use internet at work. If the worker responds that he or she uses the internet at work for any of those purposes with any frequency other than “never”, then we consider that the worker uses internet at work. Simultaneous use of computer and internet presents a more restrictive condition, which results in greater digital gaps. These calculations at the country level are shown in Figure A5. More detailed statistics are available upon request. 17 While PIAAC includes several developed economies, we did not choose them for this initial step of the analysis because their surveys were either collected several years earlier (2011-2012) and the adoption of digital technologies increased dramatically since then. In addition, some developed countries with more recent surveys do not have ISCO 08 information at the 4-digit level (e.g. United States). In contrast, New Zealand and Slovenia’s surveys were collected during the second round (2014-2015), which is the same timeframe of Chile’s survey, and closer to the timeframe of Ecuador, Mexico and Peru (2017). 18 There are 38 countries with publicly available PIAAC microdata, see https://www.oecd.org/skills/piaac/data/. 19 A similar exercise was implemented by Garrote et al., (2021) to adjust working-from-home measures by internet access rates. 20 While there are 43 ISCO08 2-digit level categories, we drop from the sample the categories 01 (Commissioned Armed Forces Officers), 02 (Non-Commissioned Armed Forces Officers) y 03 (Armed Forces Occupations, Other Ranks) since they are not available in PIAAC. 21 Table A2 in the appendix contains the estimated coefficients. 21 ILO Working Paper 121 of an individual being exposed to AI and of using a computer at work will be the multiplication of both individual probabilities. For example, take the case of a group of workers (e.g. workers with high education) that, on average, has a 0.23 GenAI augmentation exposure probability (i.e. the average of the binary exposure measure at the 4-digit level). Then let’s assume that, based on the individual predictions from the Logit model, on average, individuals in such group have a 0.7 (0.3) likelihood of using (not using) a computer at work. As a result, we conclude that workers in such group have a 0.161 (=0.7 x 0.23) probability of being exposed to GenAI augmentation and of using a computer at work, while they have a 0.069 (0.3 x 0.23) probability of being exposed to GenAI augmentation and of not using a computer at work.22 In the final step, we calculate measures of AI exposure across the same socio-demographic char - acteristics as summarized in Table 1, this time with a simultaneous breakdown by computer use at the workplace for each of these characteristics. 22 The logic is the same when the measures of GenAI exposure are imputed at the 2-digit ISCO08 level, which are not binary but continuous measures between 0 and 1. That is, the exposure to GenAI augmentation within a group of workers would be the average of the continuous exposure measure. 28 ILO Working Paper 121 XFigure 11. Exposure by country, exposure type and access to digital infrastructure We can observe that, in the case of jobs exposed to potential automation, the shares of such jobs that do not use a computer are generally very low across countries. In other words, most of such occupations are already digitized. A notable exception applies to countries with lower income, such as Nicaragua, Guatemala and Honduras, where around half of the jobs exposed to potential automation are predicted to not use a computer. One way to think about this pattern is that, in poorer countries, the lack of digital infrastructure might offer a temporary buffer from the risk of imminent automation to some occupations in this category – a trend that likely extends to countries with relatively lower incomes outside Latin America. The plot also re-confirms that such automation-exposed jobs are disproportionately held by women. The situation is visibly different in the case of occupations with augmentation potential. First, the distribution of such jobs is more equal among women and men. In addition, the shares of jobs that do not use a computer at work are also more evenly distributed across men and women. An intuitive way of thinking about this part of Figure 11 is that the light blue zones (no computer) represent the transformation potential that cannot be attained due to digital infrastructure limitations. If one was to assume that such a transformation could translate into productivity 29 ILO Working Paper 121 gains in those occupations, the zones without a computer can be seen as an unattainable potential productivity gains.34 Taking this analysis one step further, we can quantify these effects. For example, if we calculate the unattained augmentation potential as an average value across the LAC countries (weighted by total employment of each country), it corresponds to 6.24 percent of female employment and 6.22 percent of male employment. If we think of this gap as an average share of jobs within all jobs with augmentation potential, it corresponds to a non-negligible 44 percent of such jobs held by women and 50 percent of such jobs held by men. Applying these calculations to the 2023 ILO modelled estimates of total employment in each country, we estimate that there are some 17 million jobs among the 16 LAC countries in our SEDLAC sample that could, in theory, experience additional productivity from the technological transformation with GenAI, but which will not be in position to do so due to the lack of digital infrastructure. Some 7 million such jobs are held by women and nearly 10 million are held by men (see Figure A6 in Appendix for country-level breakdowns). Within-country patterns In this section, we present a sample of within-country analysis, using the example of Colombia and Costa Rica (Figure 12). Detailed breakdowns by country are provided in the Appendix and made available in a dynamic way on our online portal.35 Within each country, the plot can be read as a detailed breakdown of a given category on the vertical axis into the shares of employment represented in horizonal separations. For example, in the case of Colombia, among women, 5.5 percent of employment is exposed to the risk of automation, 11.3 percent of employment is in the augmentation category, and 22.5 percent in the big unknown. The remainder, not presented in the graph, contains all other female-held occupations that do not fall into any of those three exposure categories. The dark shaded areas represent jobs with a computer at work, while the light shading represent the jobs in each exposure type that do not use a computer. We can observe that among the female-held jobs exposed to potential automation, very few jobs do not use a computer, whereas among those with a transformation potential, this digital constraint applies to about half of such jobs. 34 We also implement a robustness check where we use the variable of ”computer ownership at home”, as self-reported by households in SEDLAC, to adjust for digitalization. In particular, we replace the imputed ”computer use at work” variable for this measure, to check if the patterns prevail. As seen in figure A.5 in the appendix, while the absolute values of exposure adjusted by digitalization are different across methods, the relative figures are very similar. In particular, among jobs exposed to GenAI augmentation, the share of workers with a computer at home tends to be higher in richer countries. 35 https://pgmyrek.shinyapps.io/AI_Data_Portal_Research/ 30 ILO Working Paper 121 XFigure 12. Exposure by country, type and detailed country-level characteristics Figure 12 can also be read vertically, allowing for within-country analysis across characteristics of both the worker and the job. Continuing with the example of Colombia, we can observe that the share of jobs with high automation potential is higher among women (5.5 percent) than men (1.6 percent), and among urban (3.9 percent) than rural jobs (0.8 percent). The share of such occupations is also highest among young people and decreases with age brackets, while it increases with education levels and household income brackets. While these general trends have already been discussed in the preceding section, the detailed breakdowns enable country-specific insights, necessary for a broader reflection on adequate policy responses. In order to support such processes, in addition to our paper, we make this detailed country-level data publicly available online. Which occupations drive the effects? To understand what drives these effects, in Figure 13 we focus only on occupations where at least a quarter of all individual observations under a given ISCO 2-digit category falls into one of the exposure groups. For example, taking “augmentation & computer”, we can observe that, in many countries, over 50 percent of teaching professionals can be found in this category. The exceptions concern lower income countries, such as Guatemala, Honduras and Nicaragua, where the majority of teaching professionals are in the category of potential augmentation, but without access 31 ILO Working Paper 121 to a computer at work. This suggests that the differences in digital development might impose important limitations on the benefits of GenAI for the education systems in poorer countries. XFigure 13. ISCO 2-digit occupations by type of exposure and country (share of exposure > 25%) Other occupations that could benefit from augmentation and already use a computer at work concern legal, social and cultural professionals, and numerical and material recording clerks in most countries, as well as health professionals, information and communication technicians and some of the assemblers in countries with relatively higher incomes. In elaborating on these categories, examples could include legal professionals like tax lawyers using GenAI for case analysis, social workers employing case management software, and cultural professionals such as digital archivists. Assemblers with a computer can include skilled workers like electricians who may utilize GenAI-based diagnostic or instructional tools in higher-income countries. 32 ILO Working Paper 121 Nevertheless, a larger share of the assemblers’ jobs falls into the category with augmentation potential but no computer at work, alongside personal service and refuse workers. While such jobs retain a central human component, in settings with a computer and internet access, some of their tasks could benefit from the AI transformation. For example, personal service workers, such as home health aides, could use a scheduling and client management software to enhance service coordination, while companies hiring refuse workers could implement waste tracking sys - tems and route optimization software to improve efficiency and reduce environmental impact.36 However, due to the lack of digital infrastructure in LAC countries, such potential benefits would simply remain out of reach to those job categories. Figure 13 also demonstrates a variety of occupations falling into the “big unknown” category, with a vast majority of such jobs already using a computer at work. This concerns business and administration professionals and associate professionals, information and communication technicians, customer service clerks, general and keyboard clerks as well as hospitality, retail and other services managers. Decomposing these groups into more detailed occupations, it is easy to illustrate why this category represents the zone between the potential of full job automation and augmentation with generative AI. For example, business and administration professionals could see GenAI software streamline complex data analysis, associate professionals might use it to manage logistics more efficiently, while customer service clerks might rely on GenAI for support with query resolution. As the technology evolves, the balance between augmentation and automation of tasks might shift, potentially redefining some of the jobs in customer query roles significantly faster and exposing them to a higher risk of full automation than other occupations in this category. It should also be noted that a large share of customer query clerks is already found in the category of high automation potential, alongside other groups of clerical occupations. The fact that, in most cases, such jobs already use a computer at work further shortens the distance to the potential full automation of these occupations, making potential shifts between the “big unknown” and full automation more fluid. Differential exposure across earnings levels As the final step of the analysis, we look deeper in the relationship between labour earnings and the degree of occupational exposure to GenAI, focusing exclusively on the occupations that report already using a computer. To do that, we use the latest available micro data for each country and focus separately on income of employees and self-employed. To ensure comparability across countries, we show the median income of each ISCO-08 2-digit level occupation as a percentage of the median income of wage employees across all observations in each survey sample.37 36 This clearly assumes that such technologies would be adopted in a way that supports workers’ tasks, rather than imposed to increase the level of algorithmic control and limit worker agency, which can have negative effects on working conditions (Adams-Prassl et al., 2023; ILO, 2023c). 37 The median is favored in wage analysis as it resists skew from outliers. It represents the central tendency of a data set and provides a clearer indication of typical income levels, especially in instances where income distribution is not symmetrical. 33 ILO Working Paper 121 XFigure 14. Earnings of occupations exposed to GenAI, by employment status (exposure above 25%) Note: Incomes were calculated as the median income in local currency of each ISCO-08 occupation, based on the latest available survey data for each country. They were then recalculated as a distance from each group’s (employees/self-employed) median income and normalized as a share of the median income of employees in each country, which provides a common reference point. Grey dots represent all occupations for which the sample size and existing data allow for calculation of the median income. Coloured dots represent only these ISCO-08 2-digit occupations where at least 25 percent of occupations in a country within a given 2-digit category are estimated as exposed to automation, augmentation or the big unknown and have a computer at work (theoretical readiness for GenAI-driven transformation). In Figure 14, we first plot all country-level data points for LAC (grey dots), with the size of each dot representing the employment shares. The dots below the horizontal reference line at zero represent occupations with an income below the median income of wage employees in each country. Conversely, dots above that line refer to jobs with income above the median income of wage employees, with the lack dotted line representing the overall trend of higher incomes accruing to jobs with a lower number in the ISCO-08 structure, that is, professional and managerial positions. 34 ILO Working Paper 121 Subsequently, we mark the jobs that are exposed to an immediate interaction with GenAI, that is, occupations within the categories of automation, augmentation and the big unknown, which report already using a computer at work. We highlight only these occupations in each country where at least a quarter of jobs in a 2-digit category falls into one type of exposure, with exposure types marked in colours and the share of employment reflected in the size of each marker. We can observe that, in the case of wage employees, nearly all exposed occupations are either around or above the median income, whereas for self-employed we see more occupations with below-median incomes. The exposure to automation is generally grouped around occupations with an income between the median and an equivalent of two median incomes, with only some of the information and communication jobs exceeding that threshold. Occupations with a high augmentation potential are generally grouped around the higher income bracket of somewhere between 1.5 to 3 values of the median income of wage employees. The category of big unknown is more spread across the income distribution, with some self-employed sales workers below the median level and the top earners around three median incomes. In the bigger picture, Figure 14 shows that most of the exposed categories concern jobs around what could be defined as middleand upper-middle income jobs, with hardly any occupations showing significant exposure among the low-income occupations. In other words, the main thrust of the first order effects of GenAI technologies can be expected among the people who already have high incomes and who are in jobs requiring relatively higher skill levels, while the jobs of the poor are quite likely to remain outside the immediate effects of this technological transition. 35 ILO Working Paper 121 XFinal discussion This study examines the exposure to GenAI within the labour markets of the LAC region, revealing both widespread potential impacts and significant variability across different demographics and sectors. Our findings indicate that a substantial proportion – between 30 and 40 percent of employment in LAC – is exposed in some way to GenAI. This exposure is linked with the economic status of countries, suggesting that income levels are a strong correlate of GenAI’s impact on labour markets. However, it is crucial to note that such exposure does not imply automation, and that for the vast majority of these jobs, the potential lies in transforming the tasks that these occupations perform. Our estimates for the potential effects of automation in LAC amount to 2 to 5 percent of employment depending on the country. These figures, while seemingly modest, should not be trivialized as they represent individuals’ livelihoods that are at stake. In addition, some of the jobs from the large category of “the big unknown” might move closer to automation over time, as the technology and its applications to workplace tasks develop further. Comparisons of our results to other studies are complicated, due to the significant differences in the concepts applied, occasional lack of detailed data that would enable a more precise assessment, diverging methods of presenting the findings, and the general scarcity of studies that cover non-HIC countries (Comunale and Manera, 2024). For example, Eloundou et al. (2023) state that up to 80% of the US workforce could have at least 10% of their tasks replaced, while 19% of workers could lose at least 50% of their tasks to LLMs – a finding that is hard to directly relate to our framework, except for the similarity of a much stronger augmentation effect over automation. McKinsey (2023) points to a similar group of “knowledge work” as being most exposed but focus the analytical work on additional value generation through productivity increases, rather than on direct effects on employment. WEF Future of Jobs (2023), even though global in scope, focuses exclusively on large enterprises, pointing to clerical and administrative jobs among occupations with the fastest expected declines. Goldman Sachs (2023), based on extrapolation of O*NET occupations to emerging economies, suggests that “most jobs and industries are only partially exposed to automation and are thus more likely to be complemented rather than substituted by AI”.38 According to our best knowledge, there are no prior studies with detailed insights to GenAI exposure in the LAC region and our estimates of total potential exposure are generally lower than the 40 percent estimated by Cazzaniga et al. (2024) for emerging economies. However, wealthier LAC countries show exposure levels closer to this estimate. Within this context, irrespective of country-specific differences, our estimations show that certain characteristics consistently correlate with higher GenAI exposure. Specifically, urban-based jobs that require higher education, are situated in the formal sector, and are held by individuals with higher relative incomes are more likely to come into interaction with this technology. Moreover, there is a pronounced tilt towards younger workers facing greater exposure, including the risk of job automation, in particular in the finance, insurance, and public administration sectors. While these groups might also be better positioned to reap the benefits of new technologies (Ananian et al., 2006; Aubert et al., 200639; Cazzaniga et al., 2024), shedding well-paid, formal, skilled and female-dominated jobs can hardly be a positive scenario for the already highly informal and unequal economies of the LAC region. Our findings suggest that – at least as the first-order effect 38 The larger automation scores in that study are hard to compare to our findings, since the underlying data is not public. 39 Studies of earlier waves of technological change, such as the introduction of the internet and digital innovations the workplace, demonstrate that such changes tend to put younger populations at an advantage, in comparison to older workers. 36 ILO Working Paper 121 – the middle class is the group whose jobs and earnings have the highest levels of overall exposure to GenAI, with many possible directions that this transformation can take. These findings suggest an important role for government interventions, aimed at minimizing disruptions resulting from sudden job losses through job protection measures, and maximizing the productive benefits of the transition, for example through equipping workers with foundational skills that can help them keep up with the changing character of jobs and drive the productive character of such changes, rather than see their skills become obsolete. The fact that some vulnerable groups, such as women and youth, face greater exposure to automation highlights the importance of life-long learning so that that workers have the skills to adapt to changes in the world of work. In the short-term, as shown in numerous ILO studies, social protection systems can play an important role as macroeconomic shock stabilizers, and reduce the impact of transitions for the affected workers and their households at the microeconomic level (ILO, 2023), especially when their use is combined with skills development programs (ILO, 2023e). In the medium and long-term, reducing gender gaps in the exposure to automation would require addressing factors that perpetuate occupational segregation by gender, such as gender-based social norms (Carranza et al., 2023). At the same time, our findings show that the shares of jobs that could benefit from a productive transformation with GenAI are consistently higher than those with automation risks across all LAC countries, ranging between 8 and 12 percent of employment across countries. This is particularly the case for the jobs in education, health and personal services. In addition, the sectors oriented towards customer service (retail, trade, hotels, restaurants, etc.) face an elevated exposure to "the big unknown", which means that a productive augmentation could also be sought in these jobs with the right policies and incentives in place. Therefore, a tempting narrative that can be constructed based on these statistics is that, in the big picture, more can be gained than lost as a net job and economic effect of the transformation. This is where our analysis provides new information to assess whether the lack of digital infrastructure could be a buffer or a bottleneck to reap the economic benefits of GenAI. On the one hand, our findings show that most workers exposed to GenAI automation are using digital technologies, which suggests that the potential negative effects may not take long to materialize. On the other hand, we find that inadequate digital infrastructure is a major bottleneck to realizing the positive effects of augmentation, thereby impacting a significant segment of the labour force in the LAC region. Nearly half of the occupations that could potentially benefit from augmentation are hampered by digital shortcomings that will prevent them from realizing that potential. Specifically, 6.24 percent of jobs held by women and 6.22 percent of those held by men are affected due to these gaps. Similar limitations apply to the jobs in the “big unknown” category: even though some of them could potentially pivot towards augmentation through increasing complementarity between GenAI and the worker in these occupations, the digital gaps will prevent large shares of these jobs from benefitting from such a scenario. The extension of our finding is that the influence of digital divides would likely be even more stark in regions of lower economic development than LAC, which underlines the need for equalizing digital access in developing countries. Policies to achieve such goal should include not only those related to digital infrastructure, but also those that aim to strengthen the incentives to adopt digital technologies for a productive use (World Bank, 2016b)40 to ensure that the transformative 40 World Bank (2016) provides several examples of policies not directly related to expanding the digital infrastructure that promote the adoption of digital technologies by firms and people, such as fostering more competition in product markets (both domestic and through international trade), improving the quality of the educational systems, etc. 37 ILO Working Paper 121 promise of GenAI does not bypass those who are most in need of its advantages. Focus on informality in the transition will also be of key importance. We find that workers with formal jobs are more exposed to GenAI automation than their counterparts in the informal sector. Even though formal jobs typically offer some coverage of the social protection systems, eventual automation of these occupations does not guarantee that the same workers would easily find new formal employment. Since workers in the informal sector from developing countries rarely move to better paying and formal jobs (Donovan et al., 2023), allowing automation to slowly erode the existing formal sector would simply expand informality. Policies aimed at reducing the segmentation of the labour market along the formal-informal lines would help improve the chances of displaced workers’ transitioning back to the formal sector. In this context, it is important to recognize that technological transformation can actually offer opportunities for job e-formalization through innovative government services (Chacaltana Janampa et al., 2024). Finally, while this study provides a detailed overview of the LAC region, it is not without limitations, some of which can be considered as open avenues for future inquiry. The first of those concerns data on the use of computers and internet at work, for which the imputation from PIACC was the best available strategy in our case. Obtaining new survey data for at least the most exposed occupations would surely offer more precision to future estimates in that regard. In the absence of such data, imputation based on the second cycle of OECD’s PIAAC, covering 2022-23, might be a viable option, as soon as such data become publicly available. Second, more could be done to understand how the task content of occupations varies across countries and in systematizing such differences according to income groups and possibly regional characteristics beyond LAC. Third, future studies could try to obtain more fine-grained characteristics of individuals’ internet access, since network latency, reliability or type of devices used may affect the adoption and impacts of GenAI by workers.41 Accordingly, we will continue our efforts to collect data on firms’ and workers’ adoption of GenAI in developing countries, in order to validate the exposure measures developed in this study and to adapt our policy responses to more detailed findings. In that regard, we would welcome collaboration with institutions, including national administrations, that might be able to assist us in the collection of such data for future research. 41 Figure A5 shows a robustness check where exposure to GenAI is split by whether workers have access to internet and a computer (instead of computer only). 44 ILO Working Paper 121 Dependent Variable: AI Exposure Augmentation Automation Big Unknown (0.00327) (0.00146) (0.00341) Construction -0.00372** -0.00328*** 0.0222*** (0.00159) (0.000812) (0.00192) Retail, wholesale trade, restaurants, hotels, etc. 0.111*** -0.00104 0.329*** (0.00187) (0.000840) (0.00257) Elect., gas, water, transp., communication 0.147*** 0.0227*** 0.0721*** (0.00288) (0.00153) (0.00274) Banks, finance, insurance, professional ss. 0.0693*** 0.0856*** 0.191*** (0.00260) (0.00250) (0.00353) Public Adm., defence 0.0571*** 0.0339*** 0.145*** (0.00301) (0.00257) (0.00424) Education, Health, personal services 0.295*** -0.0161*** 0.0155*** (0.00298) (0.00127) (0.00259) Domestic ss. 0.0467*** -0.0422*** -0.0632*** (0.00217) (0.00116) (0.00230) Constant 0.0193*** -0.0118*** 0.0297*** (0.00304) (0.00150) (0.00431) Country and year fixed effects Yes Yes Yes Mean of dependent variable 0.1203 0.0323 0.1821 Observations 888,685 888,685 888,685 R2 0.139 0.076 0.228 45 ILO Working Paper 121 References Acemoglu, D., 2024. The Simple Macroeconomics of AI. Mass. Inst. Technol. Acemoglu, D., Anderson, G., Beede, D., Buffington, C., Childress, E., Dinlersoz, E., Foster, L., Goldschlag, N., Haltiwanger, J., Kroff, Z., Restrepo, P., Zolas, N., 2023. Advanced Technology Adoption: Selection or Causal Effects? AEA Pap. Proc. 113, 210–214. https://doi.org/10.1257/ pandp.20231037 Acemoglu, D., Anderson, G.W., Beede, D.N., Buffington, C., Childress, E.E., Dinlersoz, E., Foster, L.S., Goldschlag, N., Haltiwanger, J.C., Kroff, Z., Restrepo, P., Zolas, N., 2024. Automation and the Workforce: A Firm-Level View from the 2019 Annual Business Survey. NBER Chapters. Acemoglu, D., Restrepo, P., 2022. Tasks, Automation, and the Rise in U.S. Wage Inequality. Econometrica 90, 1973–2016. https://doi.org/10.3982/ECTA19815 Acemoglu, D., Restrepo, P., 2018. Artificial Intelligence, Automation and Work (No. w24196). National Bureau of Economic Research. https://doi.org/10.3386/w24196 Adams-Prassl, J., Abraha, H., Kelly-Lyth, A., Silberman, M. ‘Six’, Rakshita, S., 2023. Regulating algorithmic management: A blueprint. Eur. Labour Law J. 14, 124–151. https://doi. org/10.1177/20319525231167299 Ananian, S., Aubert, P., Behaghel, L., 2006. Travailleurs âgés, nouvelles technologies et changements organisationnels : un réexamen à partir de l’enquête « Reponse ». Suivi d’un commentaire de Luc Behaghel : emploi des seniors - Des effets du changement technologique aux recommandations. Econ. Stat. 397, 21–49. https://doi.org/10.3406/estat.2006.7125 Aubert, P., Caroli, E., Roger, M., 2006. New technologies, organisation and age: firm-level evidence*. Econ. J. 116, F73–F93. https://doi.org/10.1111/j.1468-0297.2006.01065.x Autor, D., 2024. Applying AI to Rebuild Middle Class Jobs. Working Paper Series. https://doi. org/10.3386/w32140 Autor, D., Dorn, D., Katz, L., Patterson, C., van Reenen, J., 2020. The Fall of the Labor Share and the Rise of Superstar Firms*. Q. J. Econ. 135, 645–709. Autor, D.H., 2015. Why Are There Still So Many Jobs? The History and Future of Workplace Automation. J. Econ. Perspect. 29, 3–30. https://doi.org/10.1257/jep.29.3.3 Autor, D.H., Katz, L.F., Kearney, M.S., 2008. Trends in U.S. Wage Inequality: Revising the Revisionists. Rev. Econ. Stat. 90, 300–323. Azevedo, J.P.W.D., Inchauste Comboni, M.G., Sanfelice, V., 2013. Decomposing the recent inequality decline in Latin America. World Bank. Babina, T., Fedyk, A., He, A., Hodson, J., 2024. Artificial intelligence, firm growth, and product innovation. J. Financ. Econ. 151. 46 ILO Working Paper 121 Bakker, B.B., Ghazanchyan, M., Emmerling, J., Vibha, V., 2020. The Lack of Convergence of LatinAmerica Compared with CESEE: Is Low Investment to Blame? [WWW Document]. IMF. URL https:// www.imf.org/en/Publications/WP/Issues/2020/06/19/The-Lack-of-Convergence-of-Latin-AmericaCompared-with-CESEE-Is-Low-Investment-to-Blame-49519 (accessed 3.26.24). Brambilla, I., César, A., Falcone, G., Gasparini, L., Lombardo, C., 2023. Routinization and Employment: Evidence for Latin America. Rev. Desarro. Soc. 131–176. https://doi.org/10.13043/DYS.95.4 Brynjolfsson, E., Li, D., Raymond, L.R., 2023. Generative AI at Work. Working Paper Series. https:// doi.org/10.3386/w31161 Brynjolfsson, E., Rock, D., Syverson, C., 2021. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. Am. Econ. J. Macroecon. 13, 333–372. https://doi.org/10.1257/ mac.20180386 Carranza, E., Das, S., Kotikula, A., 2023. Gender-Based Employment Segregation: Understanding Causes And Policy Interventions. Caunedo, J., Keller, E., Shin, Y., 2023. Technology and the Task Content of Jobs across the Development Spectrum. World Bank Econ. Rev. 37, 479–493. https://doi.org/10.1093/wber/lhad015 Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Augustus J., P., Carlo, P., Emma J., R., Mendes, M., 2024. Gen-AI: Artificial Intelligence and the Future of Work [WWW Document]. IMF. URL https://www. imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligenceand-the-Future-of-Work-542379 (accessed 2.19.24). Chacaltana Janampa, J., de Mattos, F.B., García, J.M., 2024. New technologies, e-government and informality | International Labour Organization. ILO Work. Pap. 112, ILO Working Papers, Geneva. Comunale, M., Manera, A., 2024. The Economic Impacts and the Regulation of AI: A Review of the Academic Literature and Policy Actions [WWW Document]. IMF. URL https://www.imf.org/en/ Publications/WP/Issues/2024/03/22/The-Economic-Impacts-and-the-Regulation-of-AI-A-Reviewof-the-Academic-Literature-and-546645 (accessed 3.26.24). Das, M., Hilgenstock, B., 2022. The exposure to routinization: Labor market implications for developed and developing economies. Struct. Change Econ. Dyn. 60, 99–113. Dieppe, A., 2021. Global Productivity: Trends, Drivers, and Policies. Doellgast, V., O’Brady, S., Kim, J., Walters, D., 2023. AI in Contact Centers: Artificial Intelligence and Algorithmic Management in Frontline Service Workplaces. Donovan, K., Lu, W.J., Schoellman, T., 2023. Labor Market Dynamics and Development*. Q. J. Econ. 138, 2287–2325. https://doi.org/10.1093/qje/qjad019 Egana-delSol, P., Bustelo, M., Ripani, L., Soler, N., Viollaz, M., 2022. Automation in Latin America: Are Women at Higher Risk of Losing Their Jobs? Technol. Forecast. Soc. Change 175, 121333. https://doi.org/10.1016/j.techfore.2021.121333 Eloundou, T., Manning, S., Mishkin, P., Rock, D., 2023. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. 47 ILO Working Paper 121 Erumban, A., Samaan, D., Van Ark, B., 2024. Latin America’s Productivity Puzzle: Insights into diversification and structural change. Preseatation at the CEPAL conference: https://www.cepal. org/sites/default/files/events/files/daniel_samaan_presentation.pdf. Faverio, M., Tyson, A., 2022. What the data says about Americans’ views of artificial intelligence. Pew Res. Cent. URL https://www.pewresearch.org/short-reads/2023/11/21/what-the-data-saysabout-americans-views-of-artificial-intelligence/ (accessed 3.14.24). Felten, E., Raj, M., Seamans, R., 2021. Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strateg. Manag. J. 42, 2195–2217. https:// doi.org/10.1002/smj.3286 Felten, E.W., Raj, M., Seamans, R., 2023a. Occupational Heterogeneity in Exposure to Generative AI. https://doi.org/10.2139/ssrn.4414065 Felten, E.W., Raj, M., Seamans, R., 2023b. How will Language Modelers like ChatGPT Affect Occupations and Industries? SSRN Electron. J. https://doi.org/10.2139/ssrn.4375268 Felten, E.W., Raj, M., Seamans, R., 2018. A Method to Link Advances in Artificial Intelligence to Occupational Abilities. AEA Pap. Proc. 108, 54–57. https://doi.org/10.1257/pandp.20181021 Frey, C.B., Osborne, M.A., 2017. The future of employment: How susceptible are jobs to computerisation? Technol. Forecast. Soc. Change 114, 254–280. https://doi.org/10.1016/j.techfore.2016.08.019 Garrote, S.D., Gomez, P.N., Ozden, C., Rijkers, B., Viollaz, M., Winkler, H., 2021. Who on Earth Can Work from Home? World Bank Res. Obs. Gmyrek, P., Berg, J., Bescond, D., 2023. Generative AI and Jobs: A global analysis of potential effects on job quantity and quality (Working paper). Goldman Sachs, 2023. Global Economics Analyst The Potentially Large Effects of Artificial Intelligence on Economic Growth (BriggsKodnani), Global Economics Analyst. Grampp, M., Brandes, D., Laude, D., 2023. Generative AI’s fast and furious entry into Switzerland (Delloite). Gutiérrez, G., Philippon, T., 2017. Declining Competition and Investment in the U.S. NBER Work. Pap., NBER Working Papers. Hui, X., Reshef, O., Zhou, L., 2023. The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market. https://doi.org/10.2139/ssrn.4527336 ILO, 2023a. ILO Modelled Estimates (ILOEST database) [WWW Document]. ILOSTAT. URL https:// ilostat.ilo.org/resources/concepts-and-definitions/ilo-modelled-estimates/ (accessed 6.20.23). ILO, 2023b. ISCO documentation: part III. Definitions of major groups, sub-major groups, minor groups and unit groups. [WWW Document]. URL https://www.ilo.org/public/english/bureau/stat/ isco/docs/groupdefn08.pdf (accessed 6.20.23). ILO, 2023c. World Employment and Social Outlook 2023: The value of essential work [WWW Document]. URL https://www.ilo.org/digitalguides/en-gb/story/weso2023-key-workers (accessed 7.31.23). 48 ILO Working Paper 121 ILO, 2023d. World Social Protection Report 2020–22: Social protection at the crossroads – in pursuit of a better future. ILO Flaghip Report. ILO, 2023e. The ILO strategy on skills and lifelong learning 2030 | International Labour Organization. IMF, 2022. Productivity in Latin America and the Caribbean: Recent Trends and the COVID-19 Shock. Lane, M., Williams, M., Broecke, S., 2023. The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers. OECD, Paris. https://doi.org/10.1787/ea0a0fe1-en Lewandowski, P., Park, A., Hardy, W., Du, Y., 2019. Technology, Skills, and Globalization: Explaining International Differences in Routine and Nonroutine Work Using Survey Data. Lo Bello, S., Sanchez Puerta, M.L., Winkler, H.J., 2019. From Ghana to America : The Skill Content of Jobs and Economic Development. Policy Res. Work. Pap. Ser., Policy Research Working Paper Series. Maloney, W.F., Molina, C.A., 2016. Are automation and trade polarizing developing country labor markets, too ? Policy Res. Work. Pap. Ser., Policy Research Working Paper Series. McKinsey, 2023. Economic potential of generative AI | McKinsey. Messina, J., Pica, G., Oviedo, A., 2016. Job Polarization in Latin America. Inter-American Development Bank, Washington, DC. Messina, J., Silva, J., 2018. Wage Inequality in Latin America. World Bank Publ. - Books. Noy, S., Zhang, W., 2023. Experimental evidence on the productivity effects of generative artificial intelligence. Science 381, 187–192. https://doi.org/10.1126/science.adh2586 Nurski, L., Ruer, N., 2024. Exposure to generative AI in the European labour market: distributional implications. OECD, 2023. The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers (No. 288), OECD Artificial Intelligence Papers. https://doi.org/10.1787/ ea0a0fe1-en OECD/PIAAC, 2019. Skills Matter: Additional Results from the Survey of Adult Skills. Organisation for Economic Co-operation and Development, Paris. Ozdeneron, Hakki, 2023. AI-Exposed Jobs Employ More Women. URL https://www.reveliolabs. com/news/macro/ai-exposed-jobs-employ-more-women/ (accessed 6.26.23). Peng, S., Kalliamvakou, E., Cihon, P., Demirer, M., 2023. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. https://doi.org/10.48550/arXiv.2302.06590 Pew Research Center, 2023. Americans explain in their own words what makes them either more concerned or more excited about the increased presence of AI in daily life. Internet Sci. Tech. URL https://www.pewresearch.org/internet/2022/03/17/ai-and-human-enhancement-americansopenness-is-tempered-by-a-range-of-concerns/ps_2022-03-17_ai-he_01-02/ (accessed 3.14.24). 49 ILO Working Paper 121 Prytkova, E., Petit, F., Deyu, L., Sugat, C., Tommaso, C., 2024. The Employment Impact of Emerging Digital Technologies. Restrepo, P., 2023. Automation: Theory, Evidence, and Outlook. Working Paper Series. https:// doi.org/10.3386/w31910 Rutgers, 2024. U.S. Workers Assess the Impacts of Artificial Intelligence on Jobs: Topline Survey Results, Heldrich Center for Workforce Development. Schady, N., Holla, A., Sabarwal, S., Silva, J., Yi, C., 2023. Collapse and Recovery: How the COVID-19 Pandemic Eroded Human Capital and What to Do about It. Svanberg, M., Li, W., Fleming, M., Goehring, B., Thompson, N., 2024. Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? https://doi.org/10.2139/ssrn.4700751 Webb, M., 2020. The Impact of Artificial Intelligence on the Labor Market. WEF, 2023. The Future of Jobs Report 2023. World Bank, 2016a. Poverty and Shared Prosperity 2016: Taking on Inequality, Poverty and Shared Prosperity. The World Bank. https://doi.org/10.1596/978-1-4648-0958-3 World Bank, 2016b. World Development Report 2016: Digital Dividends (Text/HTML). 50 ILO Working Paper 121 Acknowledgements We would like to thank Richard Samans, Sergei Suarez Dillon Soares, Janine Berg, Leonardo Iacovone, Paulo Bastos, Carlos Rodriguez Castelan, William F. Maloney and staff from ILO Research and from the World Bank Poverty team for valuable comments. We also thank ILO STATISTICS team for data contribution, particularly David Bescond, who produced several inputs to our estimates based on ILO micro data repository, as acknowledged in detail in the paper. Research Department (RESEARCH) International Labour Organization Route des Morillons 4 1211 Geneva 22 Switzerland T +41 22 799 6530 resear[email protected] www.ilo.org/research XAdvancing social justice, promoting decent work The International Labour Organization is the United Nations agency for the world of work. We bring together governments, employers and workers to improve the working lives of all people, driving a human-centred approach to the future of work through employment creation, rights at work, social protection and social dialogue. I S B N 9789220410028 9HSTCMA*ebaaci+ 9 789220 410028 Contact details