Robots & AI Exposure and Wage Inequality: A Within Occupation Approach

Jaccoud, Florencia Working Paper Robots & AI Exposure and Wage Inequality: A Within Occupation Approach UNU-MERIT Working Papers, No. 2025-013 Provided in Cooperation with: Maastricht Economic and Social Research Institute on Innovation and Technology (UNU-MERIT), United Nations University (UNU) Suggested Citation: Jaccoud, Florencia (2025) : Robots & AI Exposure and Wage Inequality: A Within Occupation Approach, UNU-MERIT Working Papers, No. 2025-013, United Nations University (UNU), Maastricht Economic and Social Research Institute on Innovation and Technology (UNU-MERIT), Maastricht, https://doi.org/10.53330/EAJL3597 This Version is available at: https://hdl.handle.net/10419/326942 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-nc-sa/4.0/ #2025-013 Robots & AI Exposure and Wage Inequality: A Within Occupation Approach Florencia Jaccoud Published 22 April 2025 DOI: https://www.doi.org/10.53330/EAJL3597 Maastricht Economic and social Research institute on Innovation and Technology (UNU-MERIT) email: [email protected] | website: http://www.merit.unu.edu Boschstraat 24, 6211 AX Maastricht, The Netherlands Tel: (31) (43) 388 44 00 UNU-MERIT Working Papers ISSN 1871-9872 Maastricht Economic and social Research Institute on Innovation and Technology UNU-MERIT UNU-MERIT Working Papers intend to disseminate preliminary results of research carried out at UNU-MERIT to stimulate discussion on the issues raised. https://creativecommons.org/licenses/by-nc-sa/4.0/ Robots & AI Exposure and Wage Inequality: A Within Occupation Approach∗ Florencia Jaccoud† Abstract This paper examines the linkages between occupational exposure to recent automation technologies and inequality across 19 European countries. Using data from the European Union Structure of Earnings Survey (EU-SES), a fixed-effects model is employed to assess the association between occupational exposure to artificial intelligence (AI) and to industrial robots—two distinct forms of automation—and withinoccupation wage inequality. The analysis reveals that occupations with higher exposure to robots tend to have lower wage inequality, particularly among workers in the lower half of the wage distribution. In contrast, occupations more exposed to AI exhibit greater wage dispersion, especially at the top of the wage distribution. We argue that this disparity arises from differences in how each technology complements individual worker abilities: robot-related tasks often complement routine physical activities, while AI-related tasks tend to amplify the productivity of high-skilled, cognitively intensive work. Keywords: Inequality; Robots; Artificial Intelligence; Occupations JEL Codes: J31, O33, J24. ∗We are grateful to Filippo Bontadini, Tommaso Ciarli, Rinaldo Evangelista, Neil Foster-McGregor, Önder Nomaler, Fabien Petit, Guido Pialli, Ekaterina Prytkova, Matteo Tubiana, and Bart Verspagen for comments on previous versions of this paper. We are extremely thankful to Stijn Broecke, Sugat Chaturvedi, Tommaso Ciarli, Alexandre Georgieff, Deyu Li, Fabien Petit, Ekaterina Prytkova, and Jacopo Staccioli for kindly sharing the data on AI and/or robots exposure used in this paper. We are also grateful to the participants of the CORA 2024 Conference on Robots and Automation; Social Situation Monitor Research Seminar 2023 ‘The future of work: artificial intelligence and its labor market and social impacts’; the 10th Ph.D. Workshop in Economics of Innovation, Complexity and Knowledge, as well as the participants at the UNU-MERIT 2023 Research Week for the insightful comments that helped improve the paper. The author acknowledge the support by the European Union’s Horizon 2020 research and innovation program under grant agreement No. 101004703 -PILLARS (Pathways to inclusive labor markets). †Camerino University / UNU-MERIT. Email: [email protected]u.edu. 1 Introduction Over the past three decades, income inequality within countries has risen by an average of 10% across OECD nations (OECD,2015). Structural and institutional changes, globalization, and technological advances are among the key factors contributing to this trend. Given the significant job polarization resulting from the widespread adoption of computerization in developed economies, changes in the occupational structure have emerged as a prominent driver of wage inequality (Autor et al.,2003;Goos and Manning,2007). As a result, highly skilled workers in the United States and several European countries have seen a relative wage increase compared to low-skilled workers, widening the wage gap across different occupations (Acemoglu and Autor,2011;Goos et al.,2014;Cortes,2016). Yet, recent literature emphasizes that within-occupation wage disparities also play a significant role in overall inequality. For instance, Kim and Sakamoto (2008) demonstrate that in the U.S., over half of the rise in total inequality from 1983-1985 to 2000-2002 can be attributed to within-occupation inequality. This trend has persisted over time, as shown by Mishel et al. (2013) for the period from 1979 to 2007. In Europe, Akerman et al. (2013) found that within-occupation inequality explained more than 70% of the rise in wage inequality in Sweden from 2001 to 2007, a pattern echoed across other European nations (Fernández-Macías and Arranz-Muñoz,2020;van der Velde,2020). Furthermore, van der Velde (2020) finds that within-occupation wage inequality tends to be higher in jobs with a larger proportion of non-routine tasks. This paper examines the association between the exposure of occupations to robots and AI and within-occupation wage inequality in 19 European countries. In this context, we define exposure as the relationship between technology and the tasks associated with an occupation. Put simply, when there is a greater overlap between the tasks that a particular technology can perform and those involved within an occupation, the exposure of that occupation will be higher. While many studies have focused on automation and its relationship with task content, particularly routine tasks, emerging automation technologies target distinct task types. For example, AI has a greater focus on cognitive and complex tasks and is thus more relevant to high-wage occupations (Webb,2020;Felten et al.,2021;Georgieff and Hyee,2021;Engberg et al.,2024). Robots, in contrast, typically perform manual tasks such as assembly and welding (Squicciarini and Staccioli,2022;Montobbio et al.,2022;Prytkova et al.,2024). This distinction in task content is crucial for understanding within-occupation wage disparities (Jung and Mercenier,2014;van der Velde,2020). Occupations more exposed to AI involve cognitive tasks, in which there is more heterogeneity in individuals’ abilities and performance, resulting in greater wage dispersion. On the other hand, jobs more exposed to robots tend to be more standardized, limiting workers’ autonomy and variation in task performance, reducing wage disparity within those occupations.1Based on this differentiation in task content, we hypothesize a positive association between AI exposure and within-occupation wage dispersion, and a negative association for occupations more exposed to robots. Thus we argue that these technologies serve as proxies for the underlying task structure of occupations. 1It is worth noting that, while robots and AI can be analytically distinguished, the two technologies have become increasingly intertwined–particularly since the 2010s–as AI developments are progressively integrated into robotic systems (Jaccoud et al.,2024). 1 We use data from the European Union Structure of Earnings Survey (EU-SES) provided by Eurostat for 19 European countries. To capture exposure to these technologies, we use the AI and robot occupational scores developed by Webb (2020). By using patent data, the author links the tasks that these technologies perform with the job task descriptions in the Standard Occupational Classification 2010 (SOC-2010). We then map the SOC-2010 classifications to the International Standard Classification of Occupations 2008 (ISCO-08) in concordance with our database’s classification system. Our empirical strategy relies on a fixed-effect model, with the dependent variable being the logarithm of the wage gap at the 2-digit occupational level. The main independent variables are the measures of robot and AI exposure derived from Webb (2020). Additionally, we conduct several robustness checks by incorporating alternative exposure indices to robots and AI, ensuring the robustness of our findings against different specifications. Our main findings support our hypothesis, indicating that occupations that are more exposed to robots are those in which there is a lower wage disparity, and this is mainly driven by the bottom half of the wage distribution. Conversely, the ones more exposed to AI are associated with larger within-occupation wage inequality, particularly in the upper half of the wage distribution. This paper contributes to the extensive body of research on automation and inequality. While studies such as Autor et al. (2003); Goos and Manning (2007); Acemoglu and Autor (2011); Goos et al. (2014); Kaltenberg and Foster-McGregor (2020); van der Velde (2020) have primarily examined ICT-related automation, our study broadens the scope by analyzing the roles of both robots and AI in wage inequality. Our results align with van der Velde (2020), which finds a positive link between non-routine task occupations and wage dispersion and similarly suggests a negative relationship between manual tasks and wage variability, mirroring the effects seen with robot exposure. Although there have been several attempts to study the relationship between robots and inequality, they have mainly focused on betweenoccupation inequality. By focusing on the manufacturing sector and examining the impact of changes in the occupational structure, Brall and Schmid (2023) found that robot adoption primarily increases wage inequality through a compositional effect in Germany. Similarly, Barth et al. (2020) finds that in Norway, while robot penetration generally raises wages, the effect is most pronounced for managers and STEM workers. Our results should be interpreted with caution for two main reasons. First, the measures of AI and robot exposure used in this study do not reflect actual technology adoption, but rather the degree to which occupations are associated with these technologies based on their task content. As such, we treat exposure to AI and robots as proxies for occupational task structure, not as direct drivers of inequality. This means that the observed associations may be influenced by other mediating factors–such as productivity effects–that can shape wage outcomes. Second, due to the limitations of our data, we cannot make any causal claims regarding the relationship between technology exposure and wage inequality. Despite these limitations, the paper contributes valuable empirical insights by systematically linking occupational task characteristics to patterns of wage dispersion—an angle that remains underexplored in the existing literature. The rest of the paper is organized as follows. Section 2reviews relevant literature on AI and industrial robots. Section 3describes our data sources and key variables. Section 4 provides descriptive evidence of these technologies and their relationship with labor market. 2 Section 5, presents the empirical strategy and results. Finally, Section 6concludes. 2 Background literature Task content of recent automation technologies The seminal work by Autor et al. (2003) shifted focus to the task content of occupations, highlighting its importance in understanding how technology reshapes employment.2From this perspective, occupations involving tasks that follow explicit rules can be easily codified and automated. Consequently, routine tasks are the most susceptible to automation. This insight gained traction in both theoretical and empirical literature, sparking the development of various indicators to measure occupational exposure to automation (Autor et al.,2003; Goos et al.,2009;Acemoglu and Autor,2011;Autor,2013;Goos et al.,2014;Nedelkoska and Quintini,2018;Marcolin et al.,2019;Foster-McGregor et al.,2021;Gregory et al., 2022). Nevertheless, different technologies are designed for distinct activities, leading to varying associations with tasks and occupations. More recent automation technologies, such as industrial robots and AI, perform different activities and, therefore, relate differently to workforce.3 In recent years, advancements in AI have been fostered by significant improvements in data mining and computational power, which have directly contributed to the progress of machine learning technologies (Brynjolfsson et al.,2021;Zolas et al.,2021;Bonney et al., 2024). Supervised learning systems form the backbone of AI, where machines are trained to predict outcomes using vast databases. Supervised learning, the core of AI, enables machines to predict outcomes using large datasets. This remarkable progress has led to the development of generative AI based on extensive natural language models. Reflecting this shift, the Organization for Economic Co-operation and Development (OECD) recently updated its definition of AI, describing it as “a machine-based system that, for explicit or implicit objectives, infers from input how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment” (OECD, 2023). Due to its transformative potential and cross-industry applications, AI is often seen as a general-purpose technology (GPT) (Brynjolfsson et al.,2017;Trajtenberg,2018;Filippucci 2It is worth noting that the discussion on technology and its implications for employment and workforce composition is longstanding. For a detailed survey of the literature, see Piva and Vivarelli (2018) and Autor (2022). 3It is important to note that our distinction between AI and robots focuses on the technological characteristics of these tools and their associated occupational exposure. However, recent research has emphasized that digital transformation at work also includes the increasing use of digital platforms, algorithmic management, and monitoring systems, particularly affecting low-skilled and manual occupations (Fernandez Macias et al.,2023;Filippi et al.,2023;Urzì Brancati et al.,2023). These developments often reshape work organization by influencing task allocation, decision-making, and control mechanisms–without necessarily involving physical robots or advanced AI systems. For example, workers in logistics and delivery may be subject to algorithmic control and performance monitoring via digital platforms, reflecting a different but significant dimension of digital exposure. While not explicitly captured in our current framework, these mechanisms are part of the broader landscape of technological change. 3 et al.,2024).4It combines tangible inputs, such as computing power and IT infrastructure, with intangibles like software, data, and skilled labor (e.g., large datasets are essential for training algorithms, along with expertise from programmers, data scientists, and IT professionals) (Corrado et al.,2021;Filippucci et al.,2024). The combination of all these inputs results in a technology that performs cognitive activities, typically associated with highskilled workers, which is an aspect that distinguishes this technology from others (Acemoglu et al.,2022;Czarnitzki et al.,2023). Despite methodological differences, research consistently shows that occupations involving problem-solving, reasoning, and perception–i.e., non-routine cognitive tasks–are more exposed to AI than other types of occupations (Webb,2020;Felten et al.,2021;Georgieff and Hyee,2021;Engberg et al.,2024). Empirical evidence indicates that firms highly exposed to AI tend to post more AI-related job openings while reducing non-AI hiring (Acemoglu et al.,2022). Georgieff and Hyee (2021) further emphasize this trend, finding a positive correlation between increases in AI-related job postings and occupational exposure to AI across 36 sectors, underscoring the demand for AI-specific skills (Frank et al.,2019;Webb, 2020). However, it remains challenging to determine ex-ante whether AI exposure leads to the substitution or augmentation of certain occupations, and most indicators of AI exposure do not yet reflect this distinction (Guarascio et al.,2023). Although AI adoption is still at an early stage, preliminary studies suggest that it has not significantly impacted overall employment in the U.S., though it may positively affect employment at regional or firm levels (Acemoglu et al.,2022;Guarascio et al.,2023;Damioli et al.,2024). Further, it is associated with higher wages for high-skilled workers (Ernst et al., 2019;Felten et al.,2019;Ozgul et al.,2024). In contrast, industrial robots perform a different set of tasks. More recently, traditional mobile functions have been enhanced through improved sensing capacities and informationsharing with other machines. Industrial robotics can execute various operations, including object manipulation, painting, wire welding, and assembly, among others, implying a strong connection to manual tasks. For example, occupations most exposed to this technology typically involve tasks such as moving objects in factories and welding. This is translated into robots performing more physical and manual activities. For instance, Squicciarini and Staccioli (2022) highlight the high exposure to robots in occupations such as ‘Hand and pedal vehicle drivers’ (ISCO-08 9331), and ‘Vehicle cleaners’ (ISCO-08 9122) among others. Empirical studies show that robots have a skilled-biased technological change (SBTC) nature, substituting mostly low-skilled occupations (Graetz and Michaels,2018;Webb,2020; Fernández-Macías and Arranz-Muñoz,2020;Montobbio et al.,2022). In line with the literature presented above, Figures 2.1a to 2.1d corroborate the different associations of these technologies between different tasks. We observe a positive association between non-routine cognitive intensive occupations and AI exposure, while the opposite is observed for robots. Conversely, Figure 2.1c indicates that there is no correlation between non-routine manual tasks and AI exposure, while a positive association is observed for the case of robot exposure. 4Some scholars question the characterization of AI as a GPT, instead arguing that it should be understood more as a system (Vannuccini and Prytkova,2024). 4 Figure 2.1: Correlation between robot and AI exposure and task content of occupations (a) AI Non-routine Cognitive (b) Robot Non-routine Cognitive (c) AI Non-routine Manual (d) Robot Non-routine Manual Notes: The AI and robot exposure are built using the indices provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. The task content measures are derived from O*NET database, categorized at the SOC level. Following Hardy et al. (2018), we select task-related items that capture routine and non-routine activities, distinguishing between cognitive and manual dimensions. Task intensity scores for each SOC occupation are then calculated by averaging the importance scores within each category. Finally, occupations are mapped to ISCO-08 classifications. Figure 2.2 provides further insights into the relationship between technology exposure and wage distribution by depicting the average exposure to robots and AI by wage deciles.5 The figure reveals an inverse relationship between robot exposure and wage deciles, as illustrated by the downward-sloping curve. Lower-wage deciles show higher exposure to robots, while higher-wage occupations display minimal exposure. This pattern aligns with the nature of industrial robots, which are primarily utilized for manual tasks like welding, assembly, and painting—activities generally associated with lower-skilled roles. Consequently, this observation supports findings in the literature indicating that low-skilled occupations are more susceptible to automation through robotics (Chiacchio et al.,2018;Graetz and Michaels, 2018;Acemoglu and Restrepo,2020;Dauth et al.,2021). Conversely, the upward-sloping trend for AI exposure suggests that average exposure rises across higher wage deciles. Occupations at the upper end of the wage distribution demonstrate greater average exposure to AI, which is consistent with the discussion above. 5Refer to Section 3for the details on the computation of the indexes. 5 Table 3.2: Description of variables used in the final dataset Variable name Variable definition Log of the wage gap p90/p10 Log real hourly wage gap between the wage of the 90th percentile and 10th percentile at the occupation-country-year cell Log of the wage gap p90/p50 Log real hourly wage gap between the wage of the 90th percentile and 50th percentile at the occupation-country-year cell Log of the wage gap p50/p10 Log real hourly wage gap between the wage of the 50th percentile and 10th percentile at the occupation-country-year cell AI exposure Webb Exposure to AI for occupation obased on Webb (2020) Robot exposure Webb Exposure to robots for occupation obased on Webb (2020) Robot exposure Montobbio et al Exposure to robots for occupation obased on Montobbio et al. (2022) AI exposure Prytkova et al Exposure to AI for occupation obased on Prytkova et al. (2024) Robot exposure Prytkova et al Exposure to AI for occupation obased on Prytkova et al. (2024) AI exposure Felten et al Exposure to AI for occupation obased on Felten et al. (2019) AI exposure OECD Exposure to AI for occupation obased on Georgieff and Hyee (2021) AI exposure Engberg et al Exposure to AI for occupation obased on Engberg et al. (2024) Nro of 4-digit occupations Number of 4-digits occupations within major ISCO 2-digit group Std of routine cognitive intensity Standard deviation of the intensity of routine cognitive activities at 4-digit level for occupation o Std of routine manual intensity Standard deviation of the intensity of routine manual activities at 4-digit level for occupation o HHI of high-skilled occupations Herfindahl index of high-skilled employment (tertiary education or more) for occupation-country-year cell Std of AI exposure Standard deviation of the AI exposure at ISCO 4-digit level for occupation o Std of robot exposure Standard deviation of the robot exposure at ISCO 4-digit level for occupation o Notes: This Table shows the list of variables used. The data source for constructing the occupation-country-year panel is the EU-SES, while the information on the technology exposure has been elaborated based on the literature cited above. 12 Table 3.3: Descriptive statistics of dependent and explanatory variables mean sd p50 min max obs Dependent variables Log of the wage gap p90/p10 0.98 0.35 0.93 0.26 3.50 2883 Log of the wage gap p90/p50 0.54 0.20 0.51 0.14 2.88 2883 Log of the wage gap p50/p10 0.44 0.20 0.41 0.01 1.31 2883 Independent variables AI exposure Webb 0.41 0.15 0.40 0.06 0.91 2883 Robot exposure Webb 0.56 0.45 0.36 0.11 1.61 2883 AI exposure Felten et al 0.67 0.03 0.66 0.61 0.72 2883 AI exposure OECD 0.65 0.19 0.68 0.00 0.96 2067 AI exposure Prytkova et al 0.00 0.02 0.00 0.00 0.12 2883 AI exposure Engberg et al 0.14 0.28 0.00 0.00 1.06 2883 Robot exposure Montobbio et al 0.31 0.10 0.28 0.13 0.56 2883 Robot exposure Prytkova et al 0.24 0.93 0.01 0.00 7.40 2883 Control variables HHI index 0.33 0.21 0.24 0.10 1.00 2883 Share of manufacturing 0.23 0.25 0.15 0.00 1.00 2883 Blau index (gender) 0.63 0.13 0.58 0.50 0.99 2883 Share of female 0.45 0.25 0.45 0.00 0.96 2883 Share of unionization 0.70 0.32 0.82 0.00 1.00 2883 Nro of 4-digit occupations 11.53 6.76 11.00 2.00 26.00 2883 HHI of high-skilled occupations 0.59 0.15 0.55 0.33 1.00 2883 Std of AI exposure 0.17 0.07 0.19 0.03 0.36 2883 Std of robot exposure 0.29 0.23 0.22 0.04 0.94 2883 Std of routine cognitive intensity 0.24 0.10 0.22 0.03 0.62 2883 Std of routine manual intensity 0.37 0.14 0.40 0.06 0.73 2883 Notes: This Table shows the descriptive statistics of the dependent and independent variables for the whole sample. Own elaboration based on the five waves of the EU-SES. 13 4 Descriptive overview The 2000s saw a significant increase in robot adoption, leading to extensive literature exploring its effects on labor markets at various levels of aggregation (Graetz and Michaels,2018; Aghion et al.,2019;Acemoglu and Restrepo,2020).14 Figure 4.1 illustrates the evolution of robot stock per 1,000 workers-—referred to as robot density—-within the twelve European countries with the highest adoption rates. Germany stands out as a leader in robot penetration over time. However, it is also clear that some Eastern European economies made significant advancements in adoption following the financial crisis. The Czech Republic and Slovakia are particularly noteworthy, with their robot density experiencing average annual growth rates of 18.53% and 13.5%, respectively, from 1995 to 2017.15 As a result, by 2017, these two countries ranked second and third in robot density, following Germany.16 Meanwhile, Italy, Sweden, Finland, Denmark, Belgium, Austria, and the Netherlands have also shown a notable increase in robot adoption, albeit at a slower pace. Figure 4.1: Robot density 1995-2017. Notes: Figure 4.1 shows the evolution of robots stock per 1000 workers in twelve European countries. Source: Own elaboration based on the International Federation of Robotics (IFR) and EU-KLEMS for employment. In contrast to robots, AI is more recent and not widely diffused yet (Corrado et al.,2021; Vannuccini and Prytkova,2024). Consequently, systematic databases enabling time series 14For a comprehensive review, see (Jurkat et al.,2022). 15Refer to Table A.1 for more details on the average growth rate of robot density by country. 16This trend may be linked to the rise of Global Value Chains (GVCs), specifically due to Germany shifting production to Eastern European countries. 14 analysis are currently unavailable. However, there are empirical efforts to assess its pervasiveness. One notable source is the AI Observatory by the OECD, which offers indicators of AI in investment across countries. For instance, Figure A.1 in Appendix Billustrates the increasing development of AI over time, reflecting a sharp rise in venture capital investments in this technology across several European countries. A clear upward trend has been evident since 2016, in line with the findings of Corrado et al. (2021). Additionally, Calvino and Fontanelli (2023) have provided insightful analysis on AI adoption at the firm level in 11 countries, using novel data collected within the framework of the OECD AI Diffusion Project. Noteworthy findings indicate that larger and younger firms have a higher adoption rate of AI, which is strongly correlated with existing complementary capabilities. The authors emphasize that over 20% of large firms in European countries use AI.17 These findings align with recent Eurostat data on AI usage by enterprises, indicating that 8% of EU enterprises employed at least one AI technology in 2021, with the figure rising to 28% for large enterprises.18 In the remainder of this section, we will zoom into the relationship between robots and AI and workforce composition. Although AI is not yet fully widespread, the above-mentioned evidence suggests that its adoption is accelerating quickly. Therefore, the scope of the present analysis is limited to examining the potential channels through which a wider adoption of this technology may affect inequality. 4.1 Inequality and wage variability In this section, we will provide empirical evidence of the relevance of within-occupation wage inequality and its relationship with exposure to AI and robots. Figure 4.2 shows the Theil index decomposition into within and between occupation inequality for 19 European countries.19 This indicates that within-occupation inequality accounts for a significant portion of overall inequality in most of the countries studied. In 15 out of 19 countries, the share of within-occupation inequality exceeds 50%. The figure varies, ranging from 38.35% to 79.60%, with Italy having the lowest proportion and France having the highest. 17The surveyed European countries include Belgium, Denmark, France, Germany, Italy, Portugal, and Switzerland, with survey years spanning over 2017–2020. 18For further details refer to EUROSTAT - Use of AI in enterprises. 19The Theil index serves as a metric to quantify the divergence of a given income distribution from a hypothetical scenario where each individual earns an equal amount. It can be decomposed into inequality within and between groups. Refer to appendix A.3 for details on the computation. 15 Figure 4.2: Theil index decomposition: between and within occupation inequality. Notes: Figure 4.2 shows the decomposition of the Theil index into within and between occupation inequality at ISCO-08 2-digit level for the real hourly wage. It is the average across the five waves of the EU-SES (2002, 2006, 2010, 2014 and 2018). Source: Own elaboration based on EU-SES. The evidence presented above is based on a relatively broad 2-digit level of occupational classification. To provide a more detailed analysis, Figure A.2 in the appendix illustrates the Theil decomposition for occupations at 3-digit level of the ISCO-08 classification for countries with available data. While the within-occupation component shows a slight reduction, it still accounts for half or more of the overall inequality in most countries. Notable exceptions include Cyprus and Luxembourg, where the within-occupation component explains only 32.6% and 41.2% of the inequality, respectively. The importance of within-occupation wage disparity is further confirmed in Tables A.2 and A.3, where we conduct an ANOVA analysis to examine the variance of the p90/p10 real hourly wage gap as an indicator of inequality. It suggests that around 67% of the variance of the p90/p10 wage gap is explained by within-occupation variation. It also derives from Table A.2 that inequality is higher in high-paying occupations (i.e., managers, science and engineering professionals) and shows a larger dispersion of the indicator. Given that occupations involve a combination of tasks, which may relate to specific technologies designed for certain activities, exposure to these technologies may potentially be linked to wage inequality. We aim to explore this further in the remainder of this section. Recalling from Figure 2.2, occupations with higher exposure to robots are predominantly situated at the lower end of the wage distribution, whereas those more exposed to AI are concentrated at the upper end. Consequently, we can expect different patterns in wage variability, as higher-paid occupations exhibit greater dispersion. Hence, Figure 4.3 illustrates the correlation between the logarithm of the p90/p10 wage gap and the occupational exposure scores to robots and AI. It derives from this that greater exposure to AI positively 16 correlates with inequality, while the opposite occurs for robots.20 Figure 4.3: Correlation between wage inequality within the occupation and exposure to AI and robots (a) AI (b) Robot Notes: Figure 4.3 shows the correlation between the logarithm of the real hourly wage gap for the p90/p10 and the AI score (panel 4.3a) and the robot score (panel 4.3b). The robot and AI scores are at ISCO-08 2-digit level. They are built using the indexes provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. We mapped ISCO-88 to ISCO-08 to harmonize occupations across the different waves. The details of the harmonization can be consulted in appendix A. For comparability reasons, we restrict the sample to the euro countries. These are Belgium, Cyprus, Estonia, Finland, France, Italy, Lithuania, Luxemburg, Netherlands, Poland, Portugal, Slovakia, and Spain. Source: Own elaboration based on EU-SES. The observed correlations between exposure to AI or robots and wage dispersion can be 20This is also observed when analyzing the correlation with the standard deviation of the mean of the real hourly wage in the occupation in Figure A.6 in Appendix A. 17 better understood by considering the underlying task characteristics that these technologies proxy. Specifically, occupations with higher exposure to robots typically involve manual and standardized tasks. These tasks tend to be less complex and allow for limited variation in how they are performed. As a result, differences in individual performance are relatively minor, leading to more uniform productivity and, consequently, more compressed wage distributions within these roles. By contrast, occupations with greater exposure to AI are generally characterized by cognitive and complex tasks that are more heterogeneous in nature. These tasks rely more heavily on individual judgment, problem-solving, and other capabilities that vary significantly across workers. This variation contributes to broader differences in performance and output, which in turn results in wider wage dispersion within these occupations. Thus, in line with recent literature, we treat AI and robot exposure not as direct causes of wage inequality, but as indicators of the underlying task structure within occupations– structures that themselves are closely associated with differences in wage dispersion.21 In the remainder of the paper, we further examine these associations and how they reflect the differentiated role of technology across occupational task profiles. 5 Empirical strategy and results In light of the descriptives outlined in 4.1, in this section, we will examine the association between exposure of occupations to robots and AI and within-wage inequality by implementing a fixed-effects model. To do this, we apply the following specification: yoct =βKT o+γot +¯ X+ϵoct (1) Where yoct is the main dependent variable, which is the logarithm of the real hourly wage gap in occupation oin country cat year t. We consider three outcomes for withinoccupation inequality yoct: the wage gap between the p90/p10, the p50/p10 and the p90/p50. KT orepresents the exposure score of each occupation for each technology T(robot and AI). βis the main coefficient and represents the change in wage inequality for a unit change in the robot or AI occupational score. Based on the evidence presented in the previous section, we expect βto have a positive (negative) sign, meaning that increasing exposure to AI (robots) is associated with greater (lower) inequality. The variable γot includes a set of control variables that are occupation-specific.22 First, we control for the degree of collective bargaining coverage. By regulating the wage relationship, trade unions tend to be associated with a reduction in the dispersion of wages within occupations (Card et al.,2004;Biewen and Seckler,2019). Additionally, the gender composition of the labor market can influence wage distribution. Nevertheless, empirical evidence has yielded mixed results regarding the direction of this effect (Kollmeyer,2013). On one hand, there is literature suggesting that an increase in the proportion of women in the labor market can reduce wage inequality. One of the main reasons is the greater propensity 21This conceptual distinction is crucial: our empirical strategy does not aim to demonstrate a causal effect of AI or robots on wage inequality, but rather to explore how their exposure is associated with differing patterns of within-occupation inequality, mediated by the nature of task content. 22Refer to appendix B.1 for the details on the construction of control variables. 18 of low-educated women entering the labor force (Esping-Andersen,2007;Harkness,2010). Conversely, some scholars argue that higher female participation in the labor market can exacerbate wage inequality, particularly if female employment substitutes for high-skilled male employment (Acemoglu et al.,2004) or due to differing labor characteristics, such as lower trade union affiliation (Card,2001). Therefore, despite the unclear direction of the effect, we control for the share of female employment in each occupation to net out its impact on wage inequality.23 Further, there is evidence that wage dispersion is more pronounced in certain industries. In particular, there is a strand of literature that has pointed to a positive implication of servicification on wage inequality (Blum,2008;Boddin and Kroeger,2022), mainly related to their relatively greater share of high-skilled workers. Thus, to account for the industrial composition, we control for the share of manufacturing employment. The fact that the occupational analysis is conducted at the ISCO-08 2-digit level presents certain challenges. As noted in Section 3, this choice stems from the lack of consistent information on occupations at the 3-digit level for most countries. Two important considerations should be highlighted. First, the number of occupations within each major subgroup at the 2-digit level varies significantly, especially for professional categories. Naturally, this implies that variability in the outcome variable can partly be attributed to greater dispersion within occupational groups. To address this, we control for the number of 4-digit occupations within each major 2-digit category. A related issue is the substantial variation in technology exposure within each major category (Foster-McGregor et al.,2021).24 Consequently, we control for the standard deviation in exposure to robots and AI at the 4-digit level for each occupation o. In the same vein, we also include the standard deviation of routine intensity for each occupation oto further control for occupational characteristics. Finally, ¯ Xrepresents the set of fixed effects. We control for country and year-fixed effects. The first one accounts for aspects related to institutional dimensions which are timeinvariant; while the second captures changes related to business cycles–i.e. the financial crisis. We use five waves of the EU-SES: 2002, 2006, 2010, 2014 and 2018 and 19 countries in our analysis.25 Results Table 5.1 shows the association of exposure to robots and AI and inequality within occupations for the logarithm of the p90/10, p90/50 and p50/10 wage gap. When examining AI exposure, the positive coefficient indicates that occupations more exposed to this technology tend to have, on average, higher wage inequality. For the p90/p10 wage gap, a 0.1 increase in the AI exposure score is associated with an increase in inequality of 23As an additional measure, we also control for the Blau index to account for labor market diversity instead of the female share. This indicator captures potential non-linearities that may exist in the relationship between inequality and female participation. 24As occupations are aggregated, variation in technology exposure is reduced. This trend is evident in Tables C.5 to C.7, where we show correlations between the alternative technology indicators used in this paper. These tables reveal that as occupations are aggregated, the correlation between different measures increases, indicating notable heterogeneity at more granular levels. 25The countries included are Belgium, Bulgaria, Cyprus, Czech Republic, Estonia, Finland, France, Hungary, Italy, Lithuania, Luxemburg, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Spain, and Sweden. Table B.2 provides further details on the composition of the sample. 19 approximately 2.8%. This association is primarily driven by the upper half of the wage distribution. Specifically, a 0.1 increase in exposure corresponds to a rise in wage dispersion of around 1.7% for the p90/p50 gap and 1.1% for the p50/p10 gap. Unlike AI, occupations that are more exposed to robots have a lower wage inequality. For the p90/p10 wage gap, a 0.1 increase in the robot score is related to a 3.7% decrease in within-occupation inequality. This effect is pervasive across the two portions of the wage distribution as it is highly significant and negative. Nevertheless, it is slightly higher for the lower part of the wage distribution, where a 0.1 change in the robot score exposure is related to a 1.9% decrease in wage inequality, compared to a 1.8% for the upper half. These findings indicate that occupations most exposed to AI (robots) are those where there is more (less) wage inequality. The contrasting signs observed in the coefficients for robots and AI are likely linked to the different types of tasks characterizing occupations associated with each technology. Specifically, AI exposure tends to be higher in occupations involving non-routine cognitive tasks, while robot exposure is more prevalent in those involving non-routine manual tasks. Importantly, we interpret exposure to AI and robots not as direct causes of wage inequality, but rather as proxies for the underlying task structure within occupations. That is, these technologies tend to be deployed in occupational settings with distinct task characteristics, which themselves are associated with different degrees of heterogeneity in worker performance and compensation. Following Jung and Mercenier (2014), this pattern can be explained by differences in the dispersion of (unobserved) worker abilities across occupations. Workers in non-routine cognitive jobs–often those more exposed to AI–typically possess higher skill levels and exhibit a broader distribution of capabilities. Since these tasks rely heavily on individual judgment and problem-solving (e.g., graphical design), performance outcomes are more variable, resulting in greater wage dispersion. In contrast, occupations more exposed to robots involve standardized manual tasks (e.g., moving objects), where worker autonomy and discretion are limited. In these contexts, individual ability has a smaller impact on task execution, leading to more uniform productivity and narrower wage distributions.26 26It is worth noting that an additional channel that can influence wage inequality is differences within and between firms (Akerman et al.,2013;Card et al.,2018;Card,2022). For instance, Song et al. (2018) find that for the U.S. two-thirds of wage inequality is explained by wage dispersion between firms, whereas the rest is driven by wage variation within firms. In Europe, it is found that recent increase in inequality at the aggregate level is positively associated with growth in large firms (Mueller et al.,2017). Further, when also incorporating the technical change dimension, there are also large disparities observed between firms that innovate and those that do not innovate (Cirillo et al.,2017). Automation adoption can further spur wage differentials between firms that invest in these technologies and the ones that do not by expanding the demand for workers with specific skills that have higher wages (Bessen et al.,2022). 20 Table 5.1: Dependent variable: Log of real hourly wage gap. (1) (2) (3) Wage gap p90/p10 Wage gap p90/p50 Wage gap p50/p10 Robot exposure Webb -0.363∗∗∗ -0.178∗∗∗ -0.185∗∗∗ (0.023) (0.012) (0.015) AI exposure Webb 0.273∗∗∗ 0.166∗∗∗ 0.107∗∗∗ (0.053) (0.036) (0.028) Constant 1.075∗∗∗ 0.635∗∗∗ 0.439∗∗∗ (0.069) (0.042) (0.038) Observations 2883 2883 2883 R20.548 0.485 0.453 Country FE Yes Yes Yes Year FE Yes Yes Yes Manufacturing share Yes Yes Yes Sex Yes Yes Yes Union Yes Yes Yes Nro Ocupp Yes Yes Yes HHI High skill Yes Yes Yes Notes:∗∗∗ p < 0.01;∗∗ p < 0.05;∗p < 0.1. Robust standard errors between parentheses. Table 5.1 displays the results of regressing the log of the real hourly wage gap on the AI and robot exposure score at the occupation-country-year level. The AI and robot occupational scores are at ISCO-08 2-digit level. They are built using the index provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. For years before 2010, we mapped ISCO-88 to ISCO-08. The details of the harmonization can be consulted in appendix A. Control variables for each occupation-country-year cell include: the share of employment in manufacturing, the share of female employment, the share of unionized workers, and the HHI index of high-skill employment. Additional control variables at the occupational level include: the number of 4-digit occupations within each ISCO-08 2-digit occupation, the standard deviation of robot and AI exposure at the 4-digit level within ISCO-08 2-digit occupations, the standard deviation of routine task content at the 4-digit level within ISCO-08 2-digit occupations. Country and year fixed effects are also included. For further details on the construction of the control variables refer to appendix B.1. Own elaboration based on the five waves of the EU-SES. 21 From our analysis, we can envision two scenarios. First, if AI’s development leads to greater standardization of tasks across varying skill levels, it could, in fact, reduce wage inequality—-incipient empirical evidence suggests this as a possible direction (e.g., Engberg et al. (2024) and Georgieff (2024)). However, because AI can both substitute for and complement high-skilled workers, it also has the potential to worsen inequality by increasing the demand for specialized skills. According to the model proposed by Acemoglu and Restrepo (2022), technology not only automates existing tasks but also creates new ones that complement it. Therefore, it is plausible that AI will generate new tasks that are associated with high-skilled workers. In response to the inequality risks posed by AI, policymakers may consider adaptive and timely measures. Training programs to equip workers with skills that align with AI-driven changes will be crucial for a resilient workforce. Additionally, redistributive policies will play a critical role in narrowing the wage gap between highand low-paying occupations. These approaches will be key to ensuring that AI’s economic benefits are broadly shared and that its adoption fosters inclusive growth across the workforce. Our paper is subject to certain limitations. Perhaps the most significant limitation is that we cannot establish causality due to the characteristics of our data. Due to limitations in the available data, we primarily focus on occupations at the 2-digit level, even though a more detailed analysis would have been beneficial. While we attempted to provide estimates using the ISCO-08 3-digit level for some countries, this approach was feasible for only a limited number of them. Lastly, technology continually evolves, engaging in new activities over time. Consequently, exposure to technology may not remain constant, as assumed by the index employed in this paper. We partially addressed this issue by exploring a dynamic indicator developed by Prytkova et al. (2024) and Engberg et al. (2024). However, these indicators are only available from 2012 onward, while our dataset begins in 2002. Therefore, it is important to interpret this carefully. Finally, this paper paves the way for further research. As more data on AI becomes available progressively, it would be relevant to study the implications of AI adoption on inequality. Given that this technology has the potential to both substitute and complement high-skilled workers, it would be interesting to investigate further whether its adoption influences overall employment composition or instead exacerbates wage inequality, particularly among workers more exposed to it. 28 References Acemoglu, D. and Autor, D. (2011). Skills, Tasks and Technologies: Implications for Employment and Earnings. In Card, D. and Ashenfelter, O., editors, Handbook of Labor Economics, volume 4, chapter 12, pages 1043–1171. Elsevier. Acemoglu, D., Autor, D. H., Hazell, J., and Restrepo, P. (2022). Artificial Intelligence and Jobs: Evidence from Online Vacancies. Journal of Labor Economics, 40(S1):S293–S340. Acemoglu, D., Autor, D. H., and Lyle, D. (2004). Women, war, and wages: The effect of female labor supply on the wage structure at midcentury. Journal of political Economy, 112(3):497–551. Acemoglu, D. and Restrepo, P. (2020). Robots and jobs: Evidence from US Labor Markets. Journal of Political Economy, 128(6):2188–2244. Acemoglu, D. and Restrepo, P. (2022). Tasks, automation, and the rise in US wage inequality. Econometrica, 90(5):1973–2016. Aghion, P., Antonin, C., and Bunel, S. (2019). Artificial Intelligence, Growth and Employment: The Role of Policy. Economie et Statistique / Economics and Statistics, 510-51151:149–164. Akerman, A., Helpman, E., Itskhoki, O., Muendler, M.-A., and Redding, S. (2013). Sources of Wage Inequality. American Economic Review, 103(3):214–219. Autor, D. H. (2013). The “task approach” to labor markets: An overview. Journal for Labour Market Research, 46(3):185–199. Autor, D. H. (2022). The Labor Market Impacts of Technological Change: From Unbridled Enthusiasm to Qualified Optimism to Vast Uncertainty. NBER Working Paper Series 30074, National Bureau of Economic Research (NBER), Cambridge, MA. Autor, D. H., Levy, F., and Murnane, R. J. (2003). The Skill Content of Recent Technological Change: An Empirical Exploration. The Quarterly Journal of Economics, 118(4):1279– 1333. Barth, E., Røed, M., Schøne, P., and Umblijs, J. (2020). How Robots Change Within-Firm Wage Inequality. IZA DP No. 13605, Intitute of Labor Economics (IZA). Bessen, J., Denk, E., and Meng, C. (2022). The remainder effect: How automation complements labor quality. Boston University School of Law Research Paper Series, (22-3). Bessen, J. E. (2016). How computer automation affects occupations: Technology, jobs, and skills. Law & Economics Working Paper No. 15-49, Boston University School of Law. Biewen, M. and Seckler, M. (2019). Unions, internationalization, tasks, firms, and worker characteristics: A detailed decomposition analysis of rising wage inequality in Germany. The Journal of Economic Inequality, 17:461–498. 29 Blum, B. S. (2008). Trade, technology, and the rise of the service sector: The effects on US wage inequality. Journal of International Economics, 74(2):441–458. Boddin, D. and Kroeger, T. (2022). Servitization, Inequality, and Wages. Labour Economics, 77:102011. Bonney, K., Breaux, C., Buffington, C., Dinlersoz, E., Foster, L. S., Goldschlag, N., Haltiwanger, J. C., Kroff, Z., and Savage, K. (2024). Tracking firm use of AI in real time: A snapshot from the Business Trends and Outlook Survey. NBER Working Paper 32319, National Bureau of Economic Research (NBER). Brall, F. and Schmid, R. (2023). Automation, robots and wage inequality in germany: A decomposition analysis. LABOUR, 37(1):33–95. Brynjolfsson, E., Jin, W., and McElheran, K. (2021). The power of prediction: predictive analytics, workplace complements, and business performance. Business Economics, 56:217–239. Brynjolfsson, E., Rock, D., and Syverson, C. (2017). Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics. NBER Working Paper 24001, National Bureau of Economic Research (NBER). Calvino, F. and Fontanelli, L. (2023). A portrait of AI adopters across countries: Firm characteristics, assets’ complementarities and productivity. OECD Science, Technology and Industry Working Papers 2023/02, Organization for Economic Cooperation and Development (OECD). Card, D. (2001). The Effect of Unions on Wage Inequality in the U.S. Labor Market. Industrial and Labor Relations Review, 54(2):296–315. Card, D. (2022). Who set your wage? American Economic Review, 112(4):1075–1090. Card, D., Cardoso, A. R., Heining, J., and Kline, P. (2018). Firms and labor market inequality: Evidence and some theory. Journal of Labor Economics, 36(S1):S13–S70. Card, D., Lemieux, T., and Riddell, W. C. (2004). Unions and wage inequality. Journal of Labor research, 25:519–559. Chiacchio, F., Petropoulos, G., and Pichler, D. (2018). The impact of industrial robots on EU employment and wages: A local labour market approach. Bruegel Working Paper 2018/02, Bruegel. Cirillo, V., Sostero, M., and Tamagni, F. (2017). Innovation and within-firm wage inequalities: Empirical evidence from major European countries. Industry and Innovation, 24(5):468–491. Corrado, C., Haskel, J., and Jona-Lasinio, C. (2021). Artificial intelligence and productivity: An intangible assets approach. Oxford Review of Economic Policy, 37(3):435–458. 30 Cortes, G. M. (2016). Where have the middle-wage workers gone? A study of polarization using panel data. Journal of Labor Economics, 34(1):63–105. Czarnitzki, D., Fernández, G. P., and Rammer, C. (2023). Artificial intelligence and firmlevel productivity. Journal of Economic Behavior & Organization, 211:188–205. Damioli, G., Van Roy, V., Vértesy, D., and Vivarelli, M. (2024). Drivers of employment dynamics of ai innovators. Technological Forecasting and Social Change, 201:123249. Dauth, W., Findeisen, S., Suedekum, J., and Woessner, N. (2021). The adjustment of labor markets to robots. Journal ofthe European Economic Association, 19(6):3104–3153. Engberg, E., Görg, H., Lodefalk, M., Javed, F., Längkvist, M., Monteiro, N. P., Nordås, H. K., Schroeder, S., and Tang, A. (2024). AI Unboxed and Jobs: A Novel Measure and Firm-Level Evidence from Three Countries. IZA Discussion Papers 16717, Institute of Labor Economics (IZA). Ernst, E., Merola, R., and Samaan, D. (2019). Economics of Artificial Intelligence: Implications for the Future of Work. IZA Journal of Labor Policy, 9(1). Esping-Andersen, G. (2007). Sociological Explanations of Changing Income Distributions. American Behavioral Scientist, 50(5):639–658. Fana, M., Cirillo, V., Guarascio, D., and Tubiana, M. (2020). A Comparative National Tasks Database. JRC Technical Report 2020/13, Joint Research Centre (JRC), European Commission. Felten, E., Raj, M., and Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12):2195–2217. Felten, E. W., Raj, M., and Seamans, R. (2019). The occupational impact of artificial intelligence: Labor, skills, and polarization. NYU Stern School of Business. Available at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3368605. Fernández-Macías, E. and Arranz-Muñoz, J. M. (2020). Occupations and the recent trends in wage inequality in Europe. European Journal of Industrial Relations, 26(3):331–346. Fernandez Macias, E., Urzi Brancati, M. C., Wright, S., and Pesole, A. (2023). The Platformisation of Work. Technical Report JRC133016, Publications Office of the European Union, Luxembourg. Filippi, E., Bannò, M., and Trento, S. (2023). Automation technologies and their impact on employment: A review, synthesis and future research agenda. Technological Forecasting and Social Change, 191:122448. Filippucci, F., Gal, P., Jona-Lasinio, C., Leandro, A., and Nicoletti, G. (2024). The impact of artificial intelligence on productivity, distribution and growth: Key mechanisms, initial evidence and policy challenges. OECD Artificial Intelligence Papers 15, Organization for Economic Cooperation and Development (OECD). 31 Fortin, N., Lemieux, T., and Firpo, S. (2011). Occupational tasks and changes in the wage structure. IZA Discussion Papers 5542, Intitute of Labor Economics (IZA). Foster-McGregor, N., Nomaler, Ö., and Verspagen, B. (2021). Job Automation Risk, Economic Structure and Trade: a European Perspective. Research Policy, 50(7):104269. Frank, M. R., Autor, D. H., Bessen, J. E., Brynjolfsson, E., Cebrian, M., Deming, D. J., Feldman, M., Groh, M., Lobo, J., Moro, E., Wang, D., Youn, H., and Rahwan, I. (2019). Toward understanding the impact of artificial intelligence on labor. Proceedings of the National Academy of Sciences of the United States of America, 116(14):6531–6539. Frey, C. B. and Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114:254–280. Georgieff, A. (2024). Artificial intelligence and wage inequality. OECD artificial intelligence papers 13, Organization for Economic Cooperation and Development (OECD). Georgieff, A. and Hyee, R. (2021). Artificial intelligence and employment: New cross-country evidence. OECD Social, Employment and Migration Working Papers 265, Organization for Economic Cooperation and Development (OECD). Goos, M. and Manning, A. (2007). Lousy and Lovely Jobs: The Rising Polarization of Work in Britain. The Review of Economics and Statistics, 89(1):118–133. Goos, M., Manning, A., and Salomons, A. (2009). Job polarization in Europe. American Economic Review, 99(2):58–63. Goos, M., Manning, A., and Salomons, A. (2014). Explaining job polarization: Routinebiased technological change and offshoring. American Economic Review, 104(8):2509–26. Graetz, G. and Michaels, G. (2018). Robots at work. The Review of Economics and Statistics, 100(5):753–768. Gregory, T., Salomons, A., and Zierahn, U. (2022). Racing with or Against the Machine? Evidence on the Role of Trade in Europe. Journal of the European Economic Association, 20(2):869–906. Guarascio, D., Reljic, J., and Stöllinger, R. (2023). Artificial intelligence and employment: A look into the crystal ball. GLO Discussion Paper No. 1333, Global Labor Organization (GLO). Hall, P. A. and Soskice, D. (2001). An Introduction to Varietes of Capitalism. In Varieties of capitalism: The institutional foundation of comparative advantage, pages 1–70. Oxford University Press, New York. Hardy, W., Keister, R., and Lewandowski, P. (2018). Educational upgrading, structural change and the task composition of jobs in europe. Economics of Transition, 26(2):201– 231. 32 Harkness, S. (2010). The contribution of women’s employment and earnings to household income inequality: A cross-country analysis. LIS Working Paper Series 531, Luxembourg Income Study (LIS). Heß, P., Janssen, S., and Leber, U. (2023). The effect of automation technology on workers’ training participation. Economics of Education Review, 96:102438. Jaccoud, F., Petit, F., Ciarli, T., and Savona, M. (2024). Automation and Employment Over the Technology Life Cycle: Evidence from European Regions. CESifo Working Paper 10987, CESifo. Jung, J. and Mercenier, J. (2014). Routinization-biased Technical Change and Globalization: understanding Labor Market Polarization. Economic Inquiry, 52(4):1446–1465. Jurkat, A., Klump, R., and Schneider, F. (2022). Tracking the rise of robots: The IFR database. Jahrbücher für Nationalökonomie und Statistik, 242(5-6):669–689. Kaltenberg, M. and Foster-McGregor, N. (2020). The impact of automation on inequality across Europe. Working Paper Series 31, UNU-MERIT. Kim, C. and Sakamoto, A. (2008). The Rise of Intra-Occupational Wage Inequality in the United States, 1983 to 2002. American Sociological Review, 73(1):129–157. Kollmeyer, C. (2013). Family Structure, Female Employment, and National Income Inequality: A Cross-National Study of 16 Western Countries. European sociological review, 29(4):816–827. Marcolin, L., Miroudot, S., and Squicciarini, M. (2019). To be (routine) or not to be (routine), that is the question: A cross-country task-based answer. Industrial and Corporate Change, 28(3):477–501. Mishel, L., Shierholz, H., and Schmitt, J. (2013). Don’t Blame the Robots: Assessing the Job Polarization Explanation of Growing Wage Inequality. EPI–CEPR Working Paper 19, Economic Policy Institute (EPI). Montobbio, F., Staccioli, J., Virgillito, M. E., and Vivarelli, M. (2022). Robots and the origin of their labour-saving impact. Technological Forecasting and Social Change, 174:121122. Montobbio, F., Staccioli, J., Virgillito, M. E., and Vivarelli, M. (2024). Labour-saving automation: A direct measure of occupational exposure. The World Economy, 47(1):332– 361. Mueller, H. M., Ouimet, P. P., and Simintzi, E. (2017). Wage inequality and firm growth. American Economic Review, 107(5):379–383. Nedelkoska, L. and Quintini, G. (2018). Automation, skills use and training. OECD Social, Employment and Migration Working Papers 202, Organization for Economic Cooperation and Development (OECD). 33 OECD (2015). The Innovation Imperative: Contributing to Productivity, Growth and WellBeing. OECD Publishing. OECD (2023). Recommendation of the council on artificial intelligence. Accessed: 2024-1018. Ozgul, P., Fregin, M.-C., Stops, M., Janssen, S., and Levels, M. (2024). High-skilled Human Workers in Non-Routine Jobs are Susceptible to AI Automation but Wage Benefits Differ between Occupations. Technical report. Piva, M. and Vivarelli, M. (2018). Innovation, jobs, skills and tasks: A multifaceted relationship. GIORNALE DI DIRITTO DEL LAVORO E DI RELAZIONI INDUSTRIALI, 2018:599–619. Prytkova, E., Petit, F., Li, D., Chaturvedi, S., and Ciarli, T. (2024). The Employment Impact of Emerging Digital Technologies. Working Paper. Song, J., Price, D. J., Guvenen, F., Bloom, N., and von Wachter, T. (2018). Firming Up Inequality. The Quarterly Journal of Economics, 134(1):1–50. Squicciarini, M. and Staccioli, J. (2022). Labour-saving technologies and employment levels: Are Robots Really Making Workers Redundant? OECD Science, Technology and Industry Policy Papers 124, Organization for Economic Cooperation and Development (OECD). Trajtenberg, M. (2018). AI as the Next GPT: A Political-Economy Perspective. NBER Working Paper 24245, National Bureau of Economic Research (NBER). Urzì Brancati, M.-C., Gonzalez, V. I., and Fernández-Macías, E. (2023). The rise of digital monitoring and algorithmic management of work: Empirical insights. Technical report, Publications Office of the European Union. JRC Working Paper, JRC133407. van der Velde, L. (2020). Within Occupation Wage Dispersion and the Task Content of Jobs. Oxford Bulletin of Economics and Statistics, 82(5):1161–1197. Vannuccini, S. and Prytkova, E. (2024). Artificial Intelligence’s new clothes? A system technology perspective. Journal of Information Technology, 39(2):317–338. Webb, M. (2020). The impact of artificial intelligence on the labor market. Working Paper. Weksler, G. and Lastra, F. (2022). Guidowe/occupationcross: initial release. Software, Available at https://doi.org/10.5281/zenodo.7025097. Zolas, N. J., Kroff, Z., Brynjolfsson, E., McElheran, K. S., Beede, D., Buffington, C., Goldschlag, N., Foster, L., and Dinlersoz, E. (2021). Advanced Technologies Adoption and Use by U.S. Firms: Evidence from the Annual Business Survey. NBER Working Paper Series 28290, National Bureau of Economic Research (NBER). 34 Appendices A Appendix A A.1 Compatibilization between ISCO-08 and ISCO-88 Since there is a break in the ISCO classification from 2010 onwards, we apply a crosswalk between ISCO-88 and ISCO-08. This is done by implementing the function isco88_to_isco08_ndigit in the R package ‘occupationcross’ (Weksler and Lastra,2022). Drawing upon the ISCO correspondence tables provided by ILO, this function takes the individual observation at ISCO-88 2-digit (or 3-digit) level and draws an assignment to an ISCO-08 code among all the possible crosses38. We have used this function to convert ISCO-88 into ISCO-08 for the 2002 and 2006 waves of the SES. A.2 Descriptives of AI and robot penetration Figure A.1: Sum of venture capital investments in AI by country. Notes: Figure A.1 shows the evolution of the sum of investments in AI in the top fifteen European countries (in U.S. dollars). Source: OECD.AI (2022), visualisations powered by JSI using data from Preqin, accessed on 28/11/2022, www.oecd.ai 38For more details on the package refer to R package ‘occupationcross’. 35 Table A.1: Annual growth rate of robot density. 1995-2017. Austria 5.98 (3.26) Belgium 3.73 (5.64) Czech Republic 18.53 (9.77) Denmark 10.27 (4.92) Finland 4.43 (6.13) France 3.85 (4.32) Germany 5.71 (3.94) Hungary 18.09 (21.67) Italy 4.23 (4.22) Netherlands 8.76 (5.03) Slovakia 13.50 (21.10) Spain 7.48 (7.07) Sweden 4.20 (2.97) Total 8.37 (10.86) Notes: Own elaboration based on International Federation of Robotics (IFR) and EU-KLEMS. Robot density is the ratio between robot stock and employment in thousand. 36 A.3 Theil decomposition at two and three-digit levels The Theil index is as follows: T=1 N N X i=1 yi ¯yln yi ¯y Where yiis the wage of individual i,Nis the total number of employees and ¯yis the average wage of the total number of employees N. It can be further decomposed into between and within-occupation inequality in the following way: T= O X o=1 No N ¯yo ¯y ∗To | {z } Within-Occupation + O X o=1 No N ¯yo ¯yln ¯yo ¯y | {z } Between-Occupation where Nois the number of employees in occupation o,Ois the total number of occupations, ¯yois the mean wage of occupation o, and Tois the Theil index for occupation o. In this way, the within-group component is the weighted average of the Theil index across occupations. On the other hand, the between-group component represents the inequality stemming from differences between occupations. This is calculated based on the disparities in average wage over occupations in relation to the average of the entire working population. In the context of this paper, yrepresents the real hourly wage. 37 Figure A.6: Correlation between standard deviation of the real hourly wage by occupation and exposure to AI and robots (a) AI (b) Robot Notes: Figure A.6 shows the correlation between the standard deviation of the real hourly wage gap and the AI score (panel A.6a) and the robot score (panel A.6b) by occupation. The robot and AI scores are at ISCO-08 2-digit level. They are built using the indexes provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. We mapped ISCO-88 to ISCO-08 to harmonize occupations across the different waves. The details of the harmonization can be consulted in appendix A. For comparability reasons, we restrict the sample to euro countries. These are Belgium, Cyprus, Estonia, Finland, France, Italy, Lithuania, Luxemburg, Netherlands, Poland, Portugal, Slovakia, and Spain. Source: Own elaboration based on EU-SES. 44 Figure A.7: Distribution of the p90/p10 wage gap before and after log transformation Notes: Figure A.7 shows the distribution of the real hourly wage gap between the p90/p10 at the occupation-country-year cell before and after it is log transformed. Own elaboration based on EU-SES. B Appendix B B.1 Construction of control variables for within occupation inequality Share of female by occupation We compute the share of female employment on total employment for each occupation-country-year cell at ISCO-08 2-digits. Blau index We estimate the Blau index for each occupation-country-year cell by applying the following equation Blau_indexoct = 1 −X fLofct Loct 2 (2) Where Lofct is the level of employment in occupation ofor category f(i.e. male and female) in country cat time tand Loct is the total employment in occupation o, country cat year t. The index varies between 0 and 1, where values close to 0 (1) indicate lower (higher) diversity. Share of unionization To account for collective agreement coverage, we created a dummy variable taking value equal to 1 if the worker has any sort of agreement and 0 if she has 45 no agreement at all, as Table B.1 depicts. While this classification is somewhat nuanced, given that the type of collective agreement can significantly impact workers’ bargaining power, it is important to note that not all countries provide data for every category, posing challenges for comparison. For consistency, we opted for this approach. Subsequently, we calculated the proportion of workers covered by any form of collective agreement within each occupation-country-year cell. Table B.1: Classification to account for collective agreement Dissemination in SES Union dummy A National level or interconfederal agreement 1 B Industry agreement 1 C Agreement for individual industries in individual regions 1 D Enterprise or single employer agreement 1 E Agreement applying only to workers in the local unit 1 F Any other type of agreement 1 N No collective agreement exists 0 Notes: Own elaboration based on EU-SES classification. Share of employment in manufacturing We calculate the share of persons employed in manufacturing on total employment for each occupation-country-year cell at ISCO-08 2-digits. Herfindahl-Hirschman Index for Occupations (at industry level) is calculated by aggregating the square root of the share of employment in the different industries in each occupation. The indicator is as follows: HHIoct = I X i=1 Loict Loct 2 (3) Where Loict is the level of employment in occupation oin industry iin country cat time tand Loct is the overall employment in occupation o, country cat year t. In this sense, a higher value of the index indicates that employment is concentrated in fewer industries and lower levels imply that employment in that occupation is more evenly spread across industries. 46 Herfindahl-Hirschman Index for Occupations (for high-skill employment) is calculated by aggregating the squares of the shares of high-skilled workers—defined as those with tertiary or higher education—in each occupation. The indicator is expressed as follows: HHIoct =XLohct Loct 2 (4) Where Lohct is the level of high-skilled employment in occupation oin in country cat time tand Loct is the overall employment in occupation o, country cat year t. A higher value of the index for a specific occupation indicates a greater concentration of high-skilled employees within that occupation. Number of occupations within major subgroups We aggregate the number of 4-digits occupations within each ISCO08 2-digit occupations. Standard deviation of technology exposure We calculate the standard deviation by major 2-digit ISCO08 groups of the AI and robot exposure scores, along with the routine intensity indexes—both manual and cognitive—defined for each occupation at the 4-digit level. Task intensity measures We retrieved task data from the O*NET database, which is at the SOC level. In particular, we resort to the Abilities, Skills, Work Activities, and Work Context data files from O*NET. Following Hardy et al. (2018), we select task-related items that capture routine and non-routine activities, distinguishing between cognitive and manual dimensions. Task intensity scores for each SOC occupation is then calculated by averaging the importance scores within each of the four categories. Finally, occupations are mapped to ISCO-08 classifications. B.2 Descriptive statistics C Appendix C. Robustness checks C.1 Regressions with alternative control variables 47 Table B.2: Composition of the sample (unbalanced) Country 2002 2006 2010 2014 2018 Total BE 24 24 20 21 17 106 BG 37 37 37 37 37 185 CY 31 33 30 33 33 160 CZ 37 37 37 37 37 185 EE 37 37 34 35 34 177 ES 37 37 37 35 37 183 FI 37 37 35 35 34 178 FR 37 37 36 36 36 182 HU 25 25 24 25 23 122 IT 36 0 0 36 36 108 LT 36 36 33 34 32 171 LU 34 34 0 32 0 100 NL 37 37 37 37 37 185 NO 0 37 37 37 0 111 PL 0 37 37 37 36 147 PT 36 37 37 37 37 184 RO 37 37 36 35 37 182 SE 0 0 0 0 37 37 SK 36 36 36 36 36 180 Total 554 595 543 615 576 2,883 Notes: This Table shows the composition of the sample used in the analysis. Our unit of analysis are country-occupation-year, therefore each cell of the table refers to the number of occupations included. Table B.3: Composition of the sample (balanced) Country 2002 2006 2010 2014 2018 Total BG 37 37 37 37 37 185 CY 29 29 29 29 29 145 EE 33 33 33 33 33 165 ES 35 35 35 35 35 175 FI 32 32 32 32 32 160 FR 36 36 36 36 36 180 HU 22 22 22 22 22 110 NL 37 37 37 37 37 185 PT 36 36 36 36 36 180 RO 35 35 35 35 35 175 SK 36 36 36 36 36 180 Total 368 368 368 368 368 1,840 Notes: This Table shows the composition of the sample used in the analysis. Our unit of analysis are country-occupation-year, therefore each cell of the table refers to the number of occupations included. 48 Table C.1: Dependent variable: Log of real hourly wage gap. Balanced panel. (1) (2) (3) Wage gap p90/p10 Wage gap p90/p50 Wage gap p50/p10 Robot exposure Webb -0.341∗∗∗ -0.170∗∗∗ -0.171∗∗∗ (0.031) (0.016) (0.020) AI exposure Webb 0.305∗∗∗ 0.178∗∗∗ 0.128∗∗∗ (0.076) (0.053) (0.037) Constant 1.018∗∗∗ 0.590∗∗∗ 0.428∗∗∗ (0.082) (0.050) (0.047) Observations 1840 1840 1840 R20.516 0.500 0.410 Country FE Yes Yes Yes Year FE Yes Yes Yes Manufacturing share Yes Yes Yes Sex Yes Yes Yes Union Yes Yes Yes Nro Ocupp Yes Yes Yes HHI High skill Yes Yes Yes Notes:∗∗∗ p < 0.01;∗∗ p < 0.05;∗p < 0.1. Robust standard errors between parentheses. Table 5.1 displays the results of regressing the log of the real hourly wage gap on the AI and robot exposure score at the occupation-country-year level. The AI and robot occupational scores are at ISCO-08 2-digit level. They are built using the index provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. For years prior to 2010, we mapped ISCO-88 to ISCO-08. The details of the harmonization can be consulted in appendix A. Control variables for each occupation-country-year cell include: the share of employment in manufacturing, the share of female employment, the share of unionized workers, and the HHI index of high-skill employment. Additional control variables at the occupational level include: the number of 4-digit occupations within each ISCO-08 2-digit occupation, the standard deviation of robot and AI exposure at the 4-digit level within ISCO-08 2-digit occupations, the standard deviation of routine task content at the 4-digit level within ISCO-08 2-digit occupations. Country and year fixed effects are also included. For further details on the construction of the control variables refer to appendix B.1. Own elaboration based on the five waves of the EU-SES. 49 Table C.2: Dependent variable: Log of real hourly wage gap. Alternative control for gender (1) (2) (3) Wage gap p90/p10 Wage gap p90/p50 Wage gap p50/p10 Robot exposure Webb -0.293∗∗∗ -0.137∗∗∗ -0.155∗∗∗ (0.024) (0.013) (0.016) AI exposure Webb 0.563∗∗∗ 0.284∗∗∗ 0.279∗∗∗ (0.045) (0.030) (0.025) Constant 0.907∗∗∗ 0.612∗∗∗ 0.295∗∗∗ (0.063) (0.038) (0.035) Observations 2883 2883 2883 R20.533 0.490 0.424 Country FE Yes Yes Yes Year FE Yes Yes Yes Manufacturing share Yes Yes Yes Blau Index (gender) Yes Yes Yes Union Yes Yes Yes Nro Ocupp Yes Yes Yes HHI High skill Yes Yes Yes Notes:∗∗∗ p < 0.01;∗∗ p < 0.05;∗p < 0.1. Robust standard errors between parentheses. Table 5.1 displays the results of regressing the log of the real hourly wage gap on the AI and robot exposure score at the occupation-country-year level. The AI and robot occupational scores are at ISCO-08 2-digit level. They are built using the index provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. For years prior to 2010, we mapped ISCO-88 to ISCO-08. The details of the harmonization can be consulted in appendix A. Control variables for each occupation-country-year cell include: the share of employment in manufacturing, the Blau index, the share of unionized workers, and the HHI index of high-skill employment. Additional control variables at the occupational level include: the number of 4-digit occupations within each ISCO-08 2-digit occupation, the standard deviation of robot and AI exposure at the 4-digit level within ISCO-08 2-digit occupations, the standard deviation of routine task content at the 4-digit level within ISCO-08 2-digit occupations. Country and year fixed effects are also included. For further details on the construction of the control variables refer to appendix B.1. Own elaboration based on the five waves of the EU-SES. 50 Table C.3: Dependent variable: Log of real hourly wage gap. (1) (2) (3) Wage gap p90/p10 Wage gap p90/p50 Wage gap p50/p10 Robot exposure Webb -0.332∗∗∗ -0.163∗∗∗ -0.168∗∗∗ (0.023) (0.013) (0.015) AI exposure Webb 0.273∗∗∗ 0.166∗∗∗ 0.108∗∗∗ (0.053) (0.037) (0.028) Constant 1.013∗∗∗ 0.606∗∗∗ 0.407∗∗∗ (0.066) (0.042) (0.037) Observations 2883 2883 2883 R20.569 0.500 0.472 Country FE Yes Yes Yes Year FE Yes Yes Yes HHI index Yes Yes Yes Sex Yes Yes Yes Union Yes Yes Yes Nro Ocupp Yes Yes Yes HHI High skill Yes Yes Yes Notes:∗∗∗ p < 0.01;∗∗ p < 0.05;∗p < 0.1. Robust standard errors between parentheses. Table 5.1 displays the results of regressing the log of the real hourly wage gap on the AI and robot exposure score at the occupation-country-year level. The AI and robot occupational scores are at ISCO-08 2-digit level. They are built using the index provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. For years prior to 2010, we mapped ISCO-88 to ISCO-08. The details of the harmonization can be consulted in appendix A. Control variables for each occupation-country-year cell include: the HHI index of industry concentration, the share of female employment, the share of unionized workers, and the HHI index of high-skill employment. Additional control variables at the occupational level include: the number of 4-digit occupations within each ISCO-08 2-digit occupation, the standard deviation of robot and AI exposure at the 4-digit level within ISCO-08 2-digit occupations, the standard deviation of routine task content at the 4-digit level within ISCO-08 2-digit occupations. Country and year fixed effects are also included. For further details on the construction of the control variables refer to appendix B.1. Own elaboration based on the five waves of the EU-SES. 51 Table C.4: Dependent variable: Log of real hourly wage gap. Controlling for labor institutions (1) (2) (3) Wage gap p90/p10 Wage gap p90/p50 Wage gap p50/p10 Robot exposure Webb -0.343∗∗∗ -0.179∗∗∗ -0.164∗∗∗ (0.025) (0.014) (0.016) AI exposure Webb 0.190∗∗∗ 0.129∗∗∗ 0.061∗∗ (0.057) (0.036) (0.031) Constant 0.956∗∗∗ 0.595∗∗∗ 0.361∗∗∗ (0.114) (0.069) (0.068) Observations 2109 2109 2109 R20.494 0.431 0.410 Country FE Yes Yes Yes Year FE Yes Yes Yes Manufacturing share Yes Yes Yes Sex Yes Yes Yes Union Yes Yes Yes Nro Ocupp Yes Yes Yes HHI High skill Yes Yes Yes Notes:∗∗∗ p < 0.01;∗∗ p < 0.05;∗p < 0.1. Robust standard errors between parentheses. Table 5.1 displays the results of regressing the log of the real hourly wage gap on the AI and robot exposure score at the occupation-country-year level. The AI and robot occupational scores are at ISCO-08 2-digit level. They are built using the index provided by Webb (2020) at the SOC-2010 and converted into ISCO-08. For years prior to 2010, we mapped ISCO-88 to ISCO-08. The details of the harmonization can be consulted in appendix A. Control variables for each occupation-country-year cell include: the share of employment in manufacturing, the share of female employment, the share of unionized workers, and the HHI index of high-skill employment. Additional control variables at the occupational level include: the number of 4-digit occupations within each ISCO-08 2-digit occupation, the standard deviation of robot and AI exposure at the 4-digit level within ISCO-08 2-digit occupations, the standard deviation of routine task content at the 4-digit level within ISCO-08 2-digit occupations. Country and year fixed effects are also included. It also includes the log of employment protection legislation index from the OECD (individual and collective dismissals). For further details on the construction of the control variables refer to appendix B.1. Own elaboration based on the five waves of the EU-SES. C.2 Robustness checks with alternative AI and robot exposure measures Table C.5: Spearman correlations between AI and robot measures at ISCO-08 4-digit level AI Webb Rob Webb AI Felten et al AI Prytkova et al Engberg et al Rob Prytkova et al Rob Montobbio et al AI Webb 1.00 Rob Webb 0.27∗∗∗ 1.00 AI Felten et al 0.09 -0.68∗∗∗ 1.00 AI Prytkova et al 0.11∗0.07 0.10 1.00 Engberg et al 0.02 -0.72∗∗∗ 0.93∗∗∗ 0.10∗1.00 Rob Prytkova et al 0.28∗∗∗ 0.34∗∗∗ -0.08 0.41∗∗∗ -0.18∗∗∗ 1.00 Rob Montobbio et al 0.30∗∗∗ 0.51∗∗∗ -0.35∗∗∗ 0.19∗∗∗ -0.34∗∗∗ 0.48∗∗∗ 1.00 ∗p < 0.05,∗∗ p < 0.01,∗∗∗ p < 0.001 Notes: This table shows the Spearman correlation between the different alternatives of the AI and robot scores at ISCO-08 4-digit level. Correlations 52 Table C.6: Spearman correlations between AI and robot measures at ISCO-08 3-digit level AI Webb Rob Webb AI Felten et al AI Prytkova et al Engberg et al Rob Prytkova et al Rob Montobbio et al AI Webb 1.00 Rob Webb 0.29∗∗ 1.00 AI Felten et al -0.04 -0.80∗∗∗ 1.00 AI Prytkova et al 0.21∗0.06 0.11 1.00 Engberg et al -0.12 -0.84∗∗∗ 0.94∗∗∗ 0.14 1.00 Rob Prytkova et al 0.43∗∗∗ 0.40∗∗∗ -0.17 0.49∗∗∗ -0.26∗∗ 1.00 Rob Montobbio et al 0.40∗∗∗ 0.56∗∗∗ -0.39∗∗∗ 0.33∗∗∗ -0.41∗∗∗ 0.65∗∗∗ 1.00 ∗p < 0.05,∗∗ p < 0.01,∗∗∗ p < 0.001 Notes: This table shows the Spearman correlation between the different alternatives of the AI and robot scores at ISCO-08 3-digit level. Table C.7: Spearman correlations between AI and robot measures at ISCO-08 2-digit level AI Webb Rob Webb AI Felten et al AI OECD AI Prytkova et al Engberg et al Rob Prytkova et al Rob Montobbio et al AI Webb 1.00 Rob Webb 0.28 1.00 AI Felten et al -0.00 -0.85∗∗∗ 1.00 AI OECD 0.07 -0.87∗∗∗ 0.94∗∗∗ 1.00 AI Prytkova et al 0.22 -0.14 0.27 0.31 1.00 Engberg et al -0.08 -0.87∗∗∗ 0.96∗∗∗ 0.93∗∗∗ 0.35∗1.00 Rob Prytkova et al 0.60∗∗∗ 0.37∗-0.11 -0.11 0.35∗-0.17 1.00 Rob Montobbio et al 0.53∗∗∗ 0.62∗∗∗ -0.46∗∗ -0.50∗∗ 0.19 -0.44∗∗ 0.68∗∗∗ 1.00 ∗p < 0.05,∗∗ p < 0.01,∗∗∗ p < 0.001 Notes: This table shows the Spearman correlation between the different alternatives of the AI and robot scores at the ISCO-08 2-digit level. 53