Artificial intelligence technologies, skills demand and employment: Evidence from linked job ads data

Peede, Lennert; Stops, Michael Working Paper Artificial intelligence technologies, skills demand and employment: Evidence from linked job ads data IAB-Discussion Paper, No. 15/2024 Provided in Cooperation with: Institute for Employment Research (IAB) Suggested Citation: Peede, Lennert; Stops, Michael (2024) : Artificial intelligence technologies, skills demand and employment: Evidence from linked job ads data, IAB-Discussion Paper, No. 15/2024, Institut für Arbeitsmarktund Berufsforschung (IAB), Nürnberg, https://doi.org/10.48720/IAB.DP.2415 This Version is available at: https://hdl.handle.net/10419/309055 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-sa/4.0/deed.de IAB-DISCUSSION PAPER Articles on labour market issues 15|2024 Artificial intelligence technologies, skills demand and employment: evidence from linked job ads data Lennert Peede, Michael Stops ISSN 2195-2663 Artificial intelligence technologies, skills demand and employment: evidence from linked job ads data Lennert Peede, IAB and University of Erlangen-Nuremberg Michael Stops, IAB and ai:conomics Mit der Reihe „IAB-Discussion Paper“ will das Forschungsinstitut der Bundesagentur für Arbeit den Dialog mit der externen Wissenschaft intensivieren. Durch die rasche Verbreitung von Forschungsergebnissen über das Internet soll noch vor Drucklegung Kritik angeregt und Qualität gesichert werden. The “IAB Discussion Paper” is published by the research institute of the German Federal Employment Agency in order to intensify the dialogue with the scientific community. The prompt publication of the latest research results via the internet intends to stimulate criticism and to ensure research quality at an early stage before printing. Contents Contents ................................................................................................................................... 3 Abstract .................................................................................................................................... 5 Zusammenfassung ................................................................................................................... 5 JEL classification ...................................................................................................................... 5 Keywords ................................................................................................................................. 5 Acknowledgements .................................................................................................................. 6 1 Introduction ....................................................................................................................... 7 2 Effect channels of AI activities .......................................................................................... 10 3 Data and descriptive findings ........................................................................................... 11 3.1 AI activities and job ads ..................................................................................................... 11 3.2 Establishments and AI activities – some stylized facts ..................................................... 16 3.3 AI vacancies and the potential exposure of AI (and other) tools on jobs at the establishment level ............................................................................................................ 18 3.4 Skill change and employment growth .............................................................................. 21 4 Empirical strategy and results .......................................................................................... 25 4.1 Baseline model ................................................................................................................... 25 4.2 Skills demand turnover ...................................................................................................... 26 4.3 Employment ....................................................................................................................... 31 5 Robustness Checks ........................................................................................................... 36 5.1 Employment and employees’ qualification ...................................................................... 36 5.2 Exploiting the panel dimension of the data ...................................................................... 39 6 Conclusion ........................................................................................................................ 43 References .............................................................................................................................. 45 A Appendix on additional tables .......................................................................................... 47 B Appendix on additional figures ......................................................................................... 51 C Appendix on data representativeness ............................................................................... 53 D Appendix on establishment properties ............................................................................. 55 Figures ................................................................................................................................... 61 Tables ..................................................................................................................................... 61 Imprint ................................................................................................................................... 63 IAB-Discussion Paper 15|2024 5 Abstract We study how artificial intelligence (AI) affects labour demand at the establishment level. We use the share of AI related vacancy postings at the establishment level to measure efforts to develop, implement or use AI technologies. Low overall AI vacancy shares show that we study a phase of early AI adoption. At the establishment level, the AI vacancy share relates to a small reduction in those skills which are not related to AI technologies. We further find no effects on overall employment growth but slightly higher employment growth in jobs for highly skilled workers. Zusammenfassung Wir untersuchen, wie künstliche Intelligenz (KI) die Arbeitsnachfrage auf der Betriebsebene beeinflusst. Um die Aktivitäten in der Entwicklung, Implementierung oder Nutzung von KITechnologien zu messen, verwenden wir den Anteil derjenigen Stellenausschreibungen, die einen Bezug zu KI haben. Niedrige KI-Stellenanteile insgesamt zeigen, dass wir eine frühe Phase der KI-Einführung untersuchen. Auf der Betriebsebene hängt der KI-Stellenanteil mit einem relativ geringen Rückgang derjenigen Kompetenzanforderungen zusammen, die nicht mit KITechnologien in Verbindung stehen. Darüber hinaus finden wir keine Auswirkungen auf die Gesamtbeschäftigung in den Betrieben, aber ein leicht höheres Beschäftigungswachstum in Jobs mit hoch komplexen Tätigkeiten. JEL classification J23, J24, J63, O33 Keywords Artificial Intelligence, Vacancies, Skills, Employment IAB-Discussion Paper 15|2024 6 Acknowledgements We would like to thank Silke Anger, Julia Engel, Sabrina Genz, Myrielle Gonschor, Pascal Heß, Simon Janssen, Marc Levels, Alan Manning, Raymond Montizaan, Markus Nagler, Patrick Nüß, Michael Oberfichtner, Roland Rathelot, Nicholas Rounding, Duncan Roth, Jens Ruhose, Jason Sockin, Eduard Storm, and Simon Wiederhold for providing specific comments and suggestions. We are also grateful for the feedback provided by various participants at the “Knowledge, Skills, Behaviours” workshop at University of Brighton 2022, the TASKS VI conference in Nuremberg 2022, the RWI Research Seminar in Essen 2023, the Learning and Work Seminar of the University of Maastricht in 2023, the internal workshop of BIBB/IAB/ROA in Maastricht 2023, the ”Conference on Artificial Intelligence and the Economy” in Berlin 2023, the annual meetings of the European Economic Association (EEA) in Barcelona 2023, the European Association of Labor Economists (EALE) in Prague 2023, the Verein für Sozialpolitik in Regensburg 2023, the “Skills for the Future: Navigating the Digital, Green, and Social Transitions in European Labour Markets” workshop of the ELMI network in Luxemburg 2023, the annual meetings of the Scottish Economic Society in Glasgow 2024, the Society for Labor Economists 2024 in Portland (Oregon), the Canadian Economics Association in Toronto 2024, the Munich Summer Institute PhD workshop in Munich 2024, and the ”Economics of Digitization” PhD Workshop in Warsaw 2024. Michael Stops thanks the ai:conomics project consortium financed by the Federal Ministry of Labour and Social Affairs (BMAS) for financial support. IAB-Discussion Paper 15|2024 7 1 Introduction The increased availability of big data and machine learning algorithms enabled a large number of innovations in the area of artificial intelligence (AI). These new technologies fostered again the public debate about technological automation of human labour. AI chatbots are a popular example which could potentially substitute humans in a broad range of tasks (Eloundou et al. 2023). At the same there is large and growing number of very specific applications; for example, Automated Optical Inspection machines could potentially substitute humans in the specific tasks of inspecting printed circuit boards. Due to its increasing capabilities, AI technologies are expected to have important quantitative and qualitative impacts on the demand for human labour. However, the directions of these implications are ex-ante unclear: AI can replace human workers, AI can complement the work done by humans or AI requires the performance of human tasks that are new to the workforce (Acemoglu and Restrepo 2018). All this can happen at the same time, because AI as a technology is broadly defined and its implications on existing work places may be conditional on its specific forms, purposes, and how the workplaces are organised (Acemoglu and Restrepo 2020).1 In this paper, we contribute to the scarce empirical literature on labour demand effects of AI technologies at the establishment level during a phase of early AI adoption (2015-2019). To do so, we use novel and highly representative job ads text data to extract those skill requirements from the job descriptions that are required to use, implement, or develop AI (AI skills henceforth). The observed demand for AI skills serves as an indicator for different establishment activities in using, implementing or developing AI (AI activity henceforth). These job ads texts are provided by the Federal Employment Agency in Germany. Based on an establishment identifier, we can directly link the job ads data to rich administrative establishment data, which allows us to observe establishment characteristics related to AI activity. We use this data to test whether and how AI activity is related to a general change of non-AI hard skill requirements (non-AI skills henceforth) and to employment growth at the establishment level. The major contribution of this paper is to address the current lack of adequate establishment data (Raj and Seamans 2019) and to analyse the impact of AI at the establishment level. We have full access to the original job ads text data and do not have to rely on web scraping methods to obtain the data. This ensures full control of the text data analyses. Moreover, the job ads data is provided with a row of further meta information like, beside others, the establishment identifier, the very detailed occupational information, and the number of vacancies per job ad. Since the job ads receive full support by the Federal Employment Agency, severe measurement errors of these meta information are unlikely. The establishment identifier allows us to link our data to administrative data and to analyse, at the establishment level, employment growth in total, in jobs with different required skill levels, or in worker groups with different education levels. First, we find a low but increasing share of AI vacancies posted in the German labour market, i.e., the share of vacancies containing at least one AI skill. The share ranges from approximately 0.02 1 We consider AI technologies as available algorithms that process, identify, and act on patterns in unstructured data, like speech data, text, or images in systematic ways for different purposes, together with the machines, devices, and services that are controlled by these algorithms. IAB-Discussion Paper 15|2024 8 per cent in 2015 to 0.22 per cent in 2019. Throughout this paper we will use the AI vacancy share as an empirical measure for AI activity, i.e., different establishment activities in using, implementing or developing AI. Therefore, the low AI vacancy shares indicate that Germany was in the starting phase of companies’ AI activities between 2015 and 2019. Consistent with the US literature (e.g., Alekseeva et al. 2021, Acemoglu et al. 2022a), AI activity differs strongly across economic sectors. We find the highest AI vacancy shares in the information and communication technologies sector and the professional services sector, which we consider as industries where the development of AI and the provision of corresponding services are the main objectives. In contrast, we find lower AI vacancy shares in economic sectors in which establishments are more likely to use AI, like manufacturing and finance. Second, we find that AI activity has diminishing effects on the demand for other non-AI skills at the establishment level. We start by measuring establishments’ changes in the demand for nonAI skills using different skill change indices based on Deming and Noray (2020) and Acemoglu et al. (2022a). Changes in the requirement of non-AI skills are an observable indication for changes in the task content besides the tasks that are directly related to the use, implementation or development of AI. We distinguish between a negative skill change index, that quantifies the rate of change for skills with decreasing demand, and a positive skill change, that quantifies the rate of change for skills with increasing demand. The negative skill change could indicate a displacement of skill requirements (and hence tasks) while the positive skill change could indicate the introduction of new skill requirements (and hence new tasks). Then, the net skill change index considers both directions of skill changes and hence the net effect on non-AI skills or tasks. We find that AI activity is related to a lower positive skill change and a higher negative skill change although both point estimates are not statistically significant. This indicates that AI activity is related to a less pronounced introduction of new non-AI skills and a stronger decline in the demand for other non-AI skills simultaneously. Since both effects contribute to a lower demand for non-AI skills, we find that AI activity is related to a decline of 0.01 non-AI skills per posted vacancy over four years as indicated by the net skill change index. Overall, although AI activity is related to a decline in the demand for non-AI skills, the small magnitude provides evidence against a sizeable displacement of human tasks due to AI technologies. Third, in line with the diminishing effects on the skill change, we find no significant overall employment growth that is related to AI activity. The analysis of employment growth by required skill levels additionally reveals that AI activity is related to a slightly larger employment growth rate in highly complex jobs. Overall, these findings indicate that among considered establishments AI technologies tend to increase the demand for highly specialised workers which can implement or develop those technologies whereas we do not find any effect for other employees, yet. So far, the empirical evidence about the effects of AI on labour demand at the establishment level is limited and the results are mixed. In particular, none of the existing studies uses detailed administrative data to study employment effects of AI with the exception of Genz et al. (2021) who, however, analyse the effects of a broad set of Industry 4.0 technologies (including AI but IAB-Discussion Paper 15|2024 15 suggesting that on average highly skilled employees are more exposed to AI technology than lower skilled employees (Felten et al. 2018, Webb 2020).9 Figure 2: Share of AI vacancies by required skill levels Note: Data on the vacancies are from the BA-JOBBÖRSE. Vacancies from temporary work agencies are excluded. The AI vacancy share is the share of vacancies requiring at least one AI skill. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. The job ads from the BA-JOBBÖRSE have the particular feature that they can be directly linked with the Establishment History Panel (BHP) that is also provided by the Federal Employment Agency. The BHP is a cross-sectional data set for German establishments from 1975 to the most recent year. It entails data on all establishments which have at least one employee subject to social security contributions in Germany (Ganzer et al. 2022). In the empirical analysis we exploit the establishments’ employment levels for the relevant period and make use of a broad set of provided establishment characteristics as control variables. Based on this data, Figure 3 shows the AI vacancy share by industries. Though relatively low in their level, the AI vacancy shares show similarities to the patterns presented in Acemoglu et al. (2022a) and Babina et al. (2022) for the US labour market. According to the results, AI activity mainly takes places in the sectors of ICT and professional services. These sectors typically either develop or implement AI technologies for other sectors. In the following, we will refer to the both sectors as the AI producing sectors. The other sectors rather utilise AI technologies and hence we will refer to these sectors as the AI using sectors. We find that AI activities in these sectors are considerably lower than in the AI producing sectors. The manufacturing and finance sectors reveal highest AI activities whereas other sectors show even lower levels of AI activity. For the AI using sectors we generally don’t find upward dynamics in the observation period.10 9 AI exposure means the overlap of capabilities of AI technologies and required worker skills in each occupation. Therefore, the index primarily aims at the question which of the skills are not longer required from the worker, but the index also points to those occupations (and their required skill levels) that can be potentially complemented by AI skills. 10 As for the overall AI vacancy share the same conclusion applies based on shares of establishments that post at least one AI vacancy across industries (see Figure A3 in the appendix). IAB-Discussion Paper 15|2024 16 Figure 3: Share of AI vacancies per sector Note: Data on the vacancies are from the BA-JOBBÖRSE. Vacancies from temporary work agencies are excluded. The AI vacancy share is the share of vacancies requiring at least one AI skill. Data on the sectors of the posting establishments are from the BHP. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). 3.2 Establishments and AI activities – some stylized facts The exact linkage of the job ads data and the establishment data allows us to further exploit establishment characteristics that are related to AI activity. We characterise establishments with and without AI vacancy postings in 2015 along further vacancy posting behaviour in 2015 and further establishment properties. Further establishment properties include the establishmentspecific wage premia (the so-called AKM effects) provided by Bellmann et al. (2020) for the period 2010 to 2017, based on Abowd et al. (1999) and firstly applied for German establishments by Card et al. (2013), and variables referring to 2015 from the BHP. The variables for 2015 are overall employment, establishment age in years, and employment shares in jobs of the four different required skill levels. This documentation of stylised facts on establishment characteristics guides the selection of control variables in our empirical analysis. IAB-Discussion Paper 15|2024 17 Table 2: Summary statistics for establishment properties with and w/o AI activity in 2015 Leere Zelle Mean Median With AI activity (1) w/o AI activity (2) Difference (2)-(1) (3) With AI activity (4) w/o AI activity (5) Further vacancy posting Number of all vacancies in 2015 9.839 3.413 -6.426*** 5.000 2.000 Establishment properties AKM effect 2010-2017 (log points) referring to 2015 0.312 0.183 -0.130*** 0.350 0.197 Overall employees] 410.333 96.644 -313.690** 113.000 28.000 Establishment age (years) 21.043 20.110 -0.933 19.000 28.000 Employment share unskilled jobs [%] 9.708 20.389 10.681 5.495 13.333 Employment share skilled jobs [%] 47.662 62.877 15.215*** 48.905 66.667 Employment share complex jobs [%] 23.092 10.459 -12.633*** 17.241 4.878 Employment share highly complex jobs [%] 19.538 6.272 -13.266*** 14.286 1.587 Notes: All industries included. Temporary work excluded. Establishments with overall employment growth above the 95th percentile are excluded. All variables (other than the AKM effects) refer to the year 2015. AI activity means that establishments posted at least one AI vacancy in 2015. Column (3) shows significance levels from mean comparisons establishments with AI activity (1) and without AI activity (2) based on t-tests. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Table 2 shows the mean and the median of the considered variables for establishments with and without AI activity in 2015.11 We excluded establishments with an extraordinary growth in their employment.12 According to Table 2 we find, first, that establishments with AI activity post more overall vacancies in 2015 than other establishments on average (approximately 10 vacancies vs. 3 vacancies with medians of 5 vs. 2 vacancies). Second, establishments with AI activity in 2015 reveal larger AKM effects 2010-2017 (0.31 log points vs. 0.18 log points on average with means close to the medians in both groups). Third, establishments with AI activity in 2015 tend to be larger (mean of approximately 410 employees vs. 97 employees). We observe a relatively large difference in mean and median overall employment within both groups. For both groups the mean exceeds the median (approximately 113 employees vs. 28 employees) by a factor of approximately four. This larger average establishment size confirms further evidence for Germany (Rammer et al. 2021) and for the USA (Acemoglu et al. 2022b, Acemoglu et al. 2023) relying on survey data. Fourth, establishments with AI activity have a similar mean and median establishment age (mean of approximately 21 years vs. 20 years). Taken together with the finding that AI vacancies are rather posted by larger establishments points against the notion that start up establishments strongly select 11 We discuss the distribution of each these variables in Appendix. 12 In doing so, we excluded establishments with an employment growth that lies above the 95th percentile of the overall employment growth rates distribution across all establishments. IAB-Discussion Paper 15|2024 18 themselves into AI activities in 2015.13 Fifth, establishments with AI activity tend to have higher employment shares in complex jobs (approximate mean of 23 vs. 10 per cent) and highly complex jobs (20 vs. 6 per cent). The opposite is true for the employment shares in unskilled (approximately 10 vs. 20 per cent) and skilled jobs (approximately 48 per cent vs. 63 per cent). This finding of higher employment shares in (highly) complex jobs is consistent with findings for US data (Babina et al. 2023). 3.3 AI vacancies and the potential exposure of AI (and other) tools on jobs at the establishment level We now want to explore whether the observed AI activities are conditional on the specific occupational structure within the establishment or, as the current debate about AI suggests, whether AI as a ”general purpose” technology is utilized for (almost) all jobs. We utilize the AI exposure measure by Webb (2020), that links information on concrete functions of existing AI technology with tasks that have to be performed within occupations. This measure implies indeed that the usage of AI is conditional on specific typical tasks within occupations. We now descriptively evaluate this assumption by analysing the relationship of the exposure measures by Webb (2020) and the AI vacancy share within the establishment. The Webb exposure indices quantify the overlap of abilities of AI and (traditional) Software technologies and occupational tasks using patent data and occupations’ task descriptions from O*NET provided by the U.S. Department of Labor. The overlap is defined as a 5-digit occupational AI or Software exposure score. Thereby, the Webb AI and Software exposure measures predict each at which intensity typical tasks within occupations could be potentially performed by the respective technology. A low value indicates that few tasks in an occupation can be automated by the respective technology and a high value indicates that a large fraction of tasks may be automated. To compute the AI and software exposure for each establishment we aggregate the occupationspecific AI and Software exposure indices at the establishment level following Acemoglu et al. (2022a) based on individual data from the Integrated Employment Biographies (IEB) provided by the Federal Employment Agency.14 Formally, we construct the measures of Webb AI and Software exposure in 2015 at the establishment level as 𝑊𝑒𝑏𝑏 𝑒𝑥𝑝𝑜𝑠𝑢𝑟𝑒,  =   ∈  𝑒𝑚𝑝,  𝑒𝑚𝑝, ∗ 𝑊𝑒𝑏𝑏 𝑠𝑐𝑜𝑟𝑒  where 𝑊𝑒𝑏𝑏 𝑠𝑐𝑜𝑟𝑒  are Webb exposure scores of the respective technology 𝑐 ∈ {𝐴𝐼,𝑆𝑜𝑓𝑡𝑤𝑎𝑟𝑒} for each 5-digit occupation 𝑜 ∈𝑂 in an establishment 𝑒. These occupation-specific scores are weighted by ,  ,  which are the employment shares in the respective occupations within the 13 The potential selection of startups into AI activity is highly relevant for our empirical analysis, in which we examine the relationship of establishment level outcomes and AI activity, because establishment growth trajectories in start-ups may differ from those of other establishments. 14 The IEB provides longitudinal individual employment spells for the universe of employment subject to social security (Schmucker et al. 2023). IAB-Discussion Paper 15|2024 19 establishment in 2015. In contrast to Acemoglu et al. (2022a) our direct link of the job ads data to administrative data allows us to compute the 𝑊𝑒𝑏𝑏 𝑒𝑥𝑝𝑜𝑠𝑢𝑟𝑒,  measures for each establishment relying on the actual employment structure instead of the employment structure indicated in the job postings. The weighted occupational AI exposure scores are then summed up for all occupations within the establishment 𝑜 ∈𝑂. Finally, we standardise the AI exposure measure in the sample so that it has a mean zero and standard deviation of one. By averaging these exposure scores at the establishment level each of the resulting average values quantifies how strongly the task content of the whole establishment is exposed to automation by either AI or software technology. Table 3 shows the results from estimating the relationship between the AI vacancy share measured in per cent and the standardised AI and Software exposure measure. Columns (1) to (3) refer to the sample including all establishments and columns (4) to (6) refer to the sample excluding the AI producing sectors ICT and professional services. Columns (1) and (4) refer to the relationship of the AI vacancy share and Webb AI exposure only, columns (2) and (5) to the relationship of the AI vacancy share and Webb Software exposure only, and finally columns (3) and columns (6) refer to the relationship of the AI vacancy share and the both Webb exposure measures. All specifications include variables for the number of overall posted vacancies in 2015 and for establishment properties, i.e., AKM effects from 2010-2017, and establishment size, age, economic sector and federal state referring to 2015. IAB-Discussion Paper 15|2024 20 Table 3: Relationship of AI activity and Webb AI/software exposure 2015 Leere Zelle (1) (2) (3) (4) (5) (6) Dependent variable: AI vacancy share in 2015 All establishments AI using sectors Webb AI exposure 2015 0.056∗∗∗ (0.020) Leere Zelle 0.079∗∗∗ (0.027) 0.041∗∗ (0.017) Leere Zelle 0.056∗∗ (0.024) Webb software exposure 2015 Leere Zelle 0.023 (0.020) - 0.040 (0.027) Leere Zelle 0.017 (0.016) - 0.027 (0.023) Observations 33310 33310 33310 31630 31630 31630 Covariates: Further vacancy posting Number of all vacancies in 2015 yes yes yes yes yes yes Establishment properties AKM effects 2010-2017 yes yes yes yes yes yes ...referring to 2015 Establishment size yes yes yes yes yes yes Establishment age yes yes yes yes yes yes Economic sectors yes yes yes yes yes yes Federal states yes yes yes yes yes yes Notes: This tables shows the relationship between our AI activity measures and the Webb (2020) AI and software exposure index at the establishment level in 2015. Our AI activity measure is the AI vacancy share in 2015. The Webb AI (software) exposure index is a weighted average of occupation-specific AI (software) scores at the establishment level where employment shares of the respective occupations are the weights. We estimate the model with OLS. Included covariates are the number of all vacancies in 2015, AKM effects 2010-2017 and establishment size, establishment age, federal state, economic sector in 2015. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). AI exposure index from Webb (2020). We find that the AI vacancy share in 2015 positively correlates with the Webb AI exposure. Column (1) shows that, considering all establishments in the sample, an increase in the Webb AI exposure by one standard deviation is associated with a higher AI vacancy share by 0.056 percentage points (standard error of 0.024). The estimate is highly statistically significant at the one per cent level. The next column (2) shows that Webb software exposure does not correlate with the AI vacancy share as the point estimate is much lower 0.023 percentage points but the standard error is nearly the same (0.023) as for AI exposure. In the specification with both Webb exposure indices, the coefficient on the Webb AI exposure becomes slightly stronger (0.079 pp) and remains highly statistically significant at the one per cent level. The main results change only slightly once we exclude AI producing sectors ICT and professional services and thereby focusing on AI using sectors. The point estimate for the coefficient of the Webb AI exposure is slightly lower compared to the full sample (0.041 vs. 0.056). However, again we find no significant relationship between the Webb Software exposure and the AI vacancy share and after including both Webb exposure measures the coefficient of the Webb AI exposure is slightly lower but statistically significant at the five per cent level whereas the coefficient or the Webb software exposure remains insignificant. Overall, we find a robust relationship between the AI vacancy share in 2015 and the Webb AI exposure while the AI vacancy share is unrelated to the Webb Software exposure. This finding IAB-Discussion Paper 15|2024 21 confirms that the demand for AI skills also reflects that the potential usage of AI is conditional on occupational structures with specific tasks that are performed within the establishment. We will take these findings into account by including control variables for the occupational structure at the establishment level. Particularly, we will make use of the same individual data linked with our establishment data from the Integrated Employment Biographies (IEB). Therefore, we can exploit how the variation of AI skills demand within comparable occupational structures have an impact on skills and employment in the establishment. 3.4 Skill change and employment growth We now explore changes for 2015 to 2019, first, in the establishment’s skills demand and, second, of the establishment’s employment. We construct establishment level skill change indices similar to those in Deming and Noray (2020), Acemoglu et al. (2022a). Assuming that skill requirements in job ads approximate which tasks workers must complete within an establishment, changes in those skill requirements for establishments with AI activity are an observable indication of changes in the labour task content. The emergence of new skills indicates that the establishment introduced new tasks. Analogously, the disappearance of skills indicates a displacement of tasks which previously human workers performed. If more skills disappear from an establishment’s job ad than new ones emerge, this indicates a redundancy of skills required by the establishment. If AI activity is related to a sizeable skill redundancy, we see this as evidence for a sizeable displacement of human tasks. We measure the net skill change from 𝑡 = 2015 to 𝑡 = 2019 as 𝑛𝑒𝑡 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒, = ∑ ,  ,  − ,  ,     , with s denoting a specific skill from the universe of total skills 𝑆. The measure relies on the relative occurrence of skills. Therefore, we sum all occurrences of a certain skill 𝑠 required by establishment e in a given year 𝑠𝑘𝑖𝑙𝑙, 15, and divide this by the number of all vacancies of establishment 𝑒 in year 𝑡𝑣, . Finally, we sum up the differences in the relative occurrences of all posted skills between both years 2015 and 2019 in each establishment and refer to this measure as the 𝑛𝑒𝑡 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒. This measure includes two different types of skill changes: skills which establishments demand more frequently over time (𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒) and those skills which establishments demand less frequently over time (𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒). A negative net skill change for an establishment results from the negative skill change being larger than the positive skill change. We interpret this as a proxy for the indication of a sizeable skill redundancy in this establishment. The net skill change indicator differs to the measure proposed by Deming and Noray (2020) as we do not sum up absolute values of skill changes but allow negative and positive skill changes to balance out. Thereby, our net skill change measure can take both negative and positive values. This enables us to assess which direction of the skill change, i.e., whether the positive or the 15 Note that in case a vacancy lists a skill at more than one place in the text, the skill is nevertheless counted only once. IAB-Discussion Paper 15|2024 22 negative skill change dominates, and to decompose the total skill change into the negative skill change and the positive skill change: 𝑛𝑒𝑡 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒,=𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒,− 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒, where we group skills which appear more often in vacancies in 2019 relative to 2015 and skills which appear more seldom or tend to disappear in 2019 relative to 2015: 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒, =  𝑚𝑎𝑥 𝑠𝑘𝑖𝑙𝑙,  𝑣,  − 𝑠𝑘𝑖𝑙𝑙,  𝑣,  ,0    𝑛𝑎𝑔𝑒𝑎𝑡𝑖𝑣𝑒 𝑠𝑘𝑖𝑙𝑙 𝑐ℎ𝑎𝑛𝑔𝑒, =  𝑚𝑖𝑛 𝑠𝑘𝑖𝑙𝑙,  𝑣,   − 𝑠𝑘𝑖𝑙𝑙,  𝑣,  ,0    Our positive and negative skill change indices differ from Acemoglu et al. (2022a) since we assign each establishment a positive as well as a negative skill change value. Moreover, we strictly separate both skill change directions and consider for the positive skill change only those skills which appear at least with the same frequency while not adjusting for diminishing skills (and vice versa). Thereby, we can exploit the variation in the negative skill change and the positive skill change in each establishment regardless of which effect dominates. A hard skills dictionary provided by Stops et al. (2021) serves as the base to compute the skill change indicators.16 Thereby, besides others we capture the change in the task content that arises due to displacement of tasks, introduction of complementary AI development/maintaining tasks (e.g., further software or programming skills), and the introduction of new tasks that are complementary to the use of AI (e.g., new tasks due to new business models). 16 The hard skills dictionary consists of 7,270 skill terms (excluding AI terms) that are clearly defined in their meaning; this includes 10,116 keywords and 23,158 different keyword combinations. For the interpretation of the number of keywords, it must be noted that word stems are counted. These where generated as part of the pre-processing procedure where a German word stemming procedure was adopted (the base stemming procedure is CISTEM by Weissweiler and Fraser (2018)). This implies that the number of potentially identifiable words not reduced to the word stem is much larger. IAB-Discussion Paper 15|2024 23 Table 4: Summary statistics for outcome variables Outcome Variable Mean Median Min Max Skill change indices (2015-2019): Net skill change 0.483 0.500 -4.086 5.000 Positive skill change 1.633 1.000 0.000 13.000 Negative skill change 1.150 0.667 0.000 12.000 Employment growth (2015-2019) [%]: Overall employment growth 7.817 5.263 -99.687 100.000 Employment growth (unskilled jobs) 6.777 0.000 -100.000 200.000 Employment growth (skilled jobs) 5.826 0.000 -100.000 120.000 Employment growth (complex jobs) -2.812 0.000 -100.000 100.000 Employment growth (highly complex jobs) -0.748 0.000 -100.000 100.000 Notes: All industries included. Skill changes and employment growth refer to the period 2015 to 2019. The skill change indices are measured as the differences of the totals of occurrences of skills relative to all job ads of an establishment in the respective years; see also the definitions in the text. Employment growth rates are in per cent. We exclude potential outlier establishments. For the skill change indices, we exclude establishments with the lowest 2.5 per cent and the highest 2.5 per cent values in the net skill change change index. For the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Unskilled jobs require no formal qualification or only short term training. Skilled jobs require a formal vocational education training of at least 2 years. Complex jobs require a university degree or master craftman’s certificate. Highly complex jobs require a university degree or similar and, beyond that, profound professional experience or further formal highly specialised qualification certificates like a doctorate or a habilitation. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). The upper part of Table 4 presents descriptive statistics for our skill changes indices based on our sample. For our analysis, we drop potential outliers by excluding establishments with the largest and the lowest 2.5 percentile of the net skill change distribution. We find that, on average, the establishments reveal a slightly positive net skill change with a mean of approximately 0.49 (median: 0.5), resulting from a positive skill change (mean: 1.63) that is a larger than the negative skill change (mean: 1.15). IAB-Discussion Paper 15|2024 24 Table 5: Descriptive Statistics of the outcome variables for establishments with AI activity and establishments without AI activity in 2015 Leere Zelle Mean Median With AI activity w/o AI activity With AI activity w/o AI activity Skill change indices (2015-2019): Net skill change 0.091 0.485 0.000 0.500 Positive skill change 3.255 1.629 3.250 1.000 Negative skill change 3.164 1.144 3.251 0.667 Employment growth (2015-2019) [%]: Overall employment growth 13.306 7.802 12.685 5.263 Employment growth (unskilled jobs) 10.218 6.768 0.000 0.000 Employment growth (skilled jobs) 13.060 5.807 9.878 0.000 Employment growth (complex jobs) 10.538 - 2.849 0.000 0.000 Employment growth (highly complex jobs) 14.443 - 0.787 10.819 0.000 Notes: All industries included. Skill changes and employment growth refer to the period 2015 to 2019. The skill change indices are measured as the differences of the totals of occurrences of skills relative to all job ads of an establishment in the respective years; see also the definitions in the text. Employment growth is measured in per cent. For the skill change indices we exclude establishments with the lowest 2.5 per cent and the highest 2.5 per cent values of the net skill change. AI activity refers to posting at least one AI vacancy in 2015. For the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Unskilled jobs require no formal qualification or only short term training. Skilled jobs require a formal vocational education training of at least 2 years. Complex jobs require a university degree or master craftman’s certificate. Highly complex jobs require a university degree or similar and, beyond that, profound professional experience or further formal highly specialised qualification certificates like a doctorate or a habilitation. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). We now distinguish establishments with AI activities and establishments without AI activities. The upper part of Table 5 shows the mean and median of the considered skill change variables. Establishments with AI activity in 2015 have a larger turnover of skills, i.e., they have a larger negative skill change (3.16 vs. 1.14 skills per vacancy) and a higher positive skill change on average (3.26 vs. 1.63 skills per vacancy). The same conclusion holds for the median of both indices. However, for establishments with AI activity both indices are of a more similar magnitude so that AI activity is related to a lower (but still positive) net skill change. We further calculate employment growth rates in per cent from 2015 to 2019 based on the linked establishment history panel (BHP). Here we can distinguish employment growth in total from employment growth in either unskilled, skilled, complex, and highly complex jobs. The lower part of Table 4 shows descriptive statistics for employment growth for all establishments in our sample. Again, we drop potential outliers by excluding establishments with the highest larges five percentile of the respective employment growth rates distribution. Overall, and for unskilled and skilled jobs we find positive employment growth rates between 5.93 per cent (for skilled jobs), 6.78 per cent (for unskilled jobs) and 7.82 per cent in total. Employment growth rates were negative, though in lower magnitudes, for complex jobs (-2.81 per cent) and, quite smaller, for highly complex jobs (-0.75 per cent). IAB-Discussion Paper 15|2024 31 In the fully specified model (see col. 9 of Table 7) the effect on the net skill change amounts to - 0.011 skills per vacancy which is a similar magnitude like the estimate of the same regression based on the full sample of establishments. The estimate is statistically significant at the five per cent level. The estimates for the effects on the positive and negative skill change rates (-0.006 and 0.005 skills per vacancy, resp.) are also rather similar compared to the corresponding point estimates from the sample including all establishments. Both point estimates are not statistically significant in this sample, too. Thus, in establishments that rather utilize AI we also find only a weak relation between non-AI skills redundancies and AI activities. Overall, considering all establishments we find only small effects of AI activity on the skills turnover. Although AI activity is related to a slightly higher rate at which skills become redundant, the effect is small. For AI using sectors we find similar results. Since we assume that changes in the skills demand indicate changes in the labour task content, our results suggest only small changes in the labour task content related to AI activity apart from the introduction of AI-specific tasks. 4.3 Employment Next, we analyse establishment employment growth in per cent across all jobs and across jobs in each of the four required skill levels. Thus, the coefficient 𝛽 denotes a change in employment growth in percentage points according to a one percentage point change of the AI vacancy share. AI activity in 2015 associated with subsequent lower employment growth either in total or in a particular required skill level would provide evidence for displacement effects outweighing reinstatement effects within the establishment. Table 8 provides the estimation results of nine different specifications. Again, we start by estimating the bivariate relationship of employment growth and AI activity in 2015 (col. 1) and subsequently include covariates into the model (col. 2-8) and, finally, include all covariates (col. 9). IAB-Discussion Paper 15|2024 32 Table 8: Employment growth 2015 - 2019 and AI skills demand 2015, all establishments Leere Zelle (1) (2) (3) (4) (5) (6) (7) (8) (9) Overall Employment Growth AI vacancy share in 2015 0.055 (0.049) 0.006 (0.049) 0.006 (0.049) 0.002 (0.048) - 0.003 (0.049) - 0.007 (0.047) 0.006 (0.048) - 0.016 (0.045) - 0.030 (0.045) Observations 33771 33771 33771 33771 33771 33771 33771 33771 33771 Unskilled jobs AI vacancy share in 2015 0.112 (0.122) 0.182 (0.126) 0.181 (0.126) 0.182 (0.126) 0.181 (0.126) 0.179 (0.125) 0.182 (0.126) 0.186 (0.126) 0.175 (0.125) Observations 34169 34169 34169 34169 34169 34169 34169 34169 34169 Skilled jobs AI vacancy share in 2015 0.037 (0.068) 0.016 (0.067) 0.016 (0.067) 0.011 (0.066) 0.011 (0.066) 0.008 (0.067) 0.020 (0.067) - 0.001 (0.066) -0.011 (0.066) Observations 33653 33653 33653 33653 33653 33653 33653 33653 33653 Complex jobs AI vacancy share in 2015 0.154∗∗ (0.077) 0.054 (0.074) 0.053 (0.074) 0.049 (0.073) 0.057 (0.073) 0.033 (0.072) 0.057 (0.074) 0.026 (0.072) 0.024 (0.072) Observations 33699 33699 33699 33699 33699 33699 33699 33699 33699 Highly complex jobs AI vacancy share in 2015 0.218∗∗∗ (0.073) 0.152∗∗ (0.076) 0.152∗∗ (0.076) 0.148∗∗ (0.075) 0.155∗∗ (0.076) 0.135∗ (0.073) 0.142∗ (0.075) 0.151∗∗ (0.076) 0.132∗ (0.072) Observations 34094 34094 34094 34094 34094 34094 34094 34094 34094 Covariates: Further vacancy posting Average AI vacancy share 2016-19 no yes yes yes yes yes yes yes yes Number of all vacancies in 2015 no no yes no no no no no yes Establishment properties AKM effects 2010-2017 no no no yes no no no no yes ... referring to 2015 Establishment size no yes yes yes yes yes yes yes yes Establishment age no no no no yes no no no yes Economic sectors no no no no no yes no no yes Federal states no no no no no no yes no yes Occupational shares no no no no no no no yes yes Notes: This table reports the estimation results for regressing the overall employment growth and employment growth in jobs differentiated by skill requirement level on the AI vacancy share in 2015. Unskilled jobs require no formal qualification or only short term training. Skilled jobs require a formal vocational education training of at least 2 years. Complex jobs require a university degree or master craftman’s certificate. Highly complex jobs require a university degree or similar and, beyond that, profound professional experience or further formal highly specialised qualification certificates like a doctorate or a habilitation. All industries are included. For regressions with the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Integrated Employment Biographies (IEB). IAB-Discussion Paper 15|2024 33 Across all specifications the coefficient for the effect on overall employment growth is statistically insignificant. In most specifications the overall employment growth is positively related to AI activity but turns negative when we include establishment age (col. 5), economic sectors (col. 6) and/or occupational shares (col. 8) as covariates into the model. After including all covariates (col. 9), the establishments have a lower employment growth of 0.03 percentage points per 1 percentage point increase in the AI vacancy share on average. Given an average overall employment growth of 7.8 per cent in the sample considering all establishments21, this effect on the growth rate is rather small and not statistically significant. However, we observe heterogeneity of employment growth differences across the required skill levels. The estimated effect for employment growth in unskilled jobs is rather stable but statistically insignificant across all specifications. Compared to the bivariate relationship (0.112, see col. 1), the estimated effect on employment growth in unskilled jobs changes only slightly (0.175) in the fully specified model (col. 9). However, considering the fully specified model the estimated effect on employment growth in these jobs has the highest magnitude compared to the other required skill levels. In skilled jobs the magnitude of the point estimate in the fully specified model is similar to the point estimate for overall employment growth (-0.011) but also not statistically significant. In complex jobs an increase in AI activity in 2015 by one percentage point is associated with an increase in employment growth by 0.154 percentage points. The estimate is statistically significant at the 5 per cent level. However, in the fully specified model the effect on employment growth is much smaller than in the bivariate relationship but also positively associated with AI activity (0.024). Despite of different magnitudes we find no significant effects on employment growth in unskilled, skilled and complex jobs once we include the covariates. We only find a statistically significant effect in employment growth in highly complex jobs related to AI activity. Without any covariates we estimate positive employment growth by 0.218 percentage points which is significant at the one per cent level. However, even after including all covariates establishments have a 0.132 percentage points higher employment growth rate in highly complex jobs per 1 percentage point increase in the AI vacancy share in 2015. The estimate is significant at the 10 per cent level. Notably, once we control for further AI vacancy postings in 2016-2019 and establishment size (col. 2), the magnitude of the estimated coefficient is rather stable across specifications (col. 3-9). The estimated effect in employment growth is positive and indicates a reversed direction to the average employment growth in highly complex jobs across all establishments of -0.716 per cent in our sample (Table 4). Given that the estimated coefficient is much lower than one and is related to a one percentage point increase in the AI vacancy share, the effect is small. Since additionally, on average the share of employees in highly complex jobs is rather low across establishments (6.3 per cent), the observed higher employment growth in highly complex jobs does not translates into a higher overall employment growth. Next, we again exclude the ICT and the professional services sectors from our analysis (Table 9). The point estimate of -0.065 after controlling for all covariates has a similar magnitude like the estimate based on all establishments for overall employment growth. Again, as in the sample considering all establishments the point estimate is not statistically significant. The major 21 See Table 4 for average outcome growth rates for establishments without AI activity and others. IAB-Discussion Paper 15|2024 34 difference of the sample excluding AI producing sectors relative to the sample of all establishments is the insignificant estimate of 0.126 in employment growth in highly complex jobs due to a slightly smaller magnitude and a larger standard error. IAB-Discussion Paper 15|2024 35 Table 9: Employment growth 2015 - 2019 and AI skills demand 2015, establishments in sectors outside ICT and professional services Leere Zelle (1) (2) (3) (4) (5) (6) (7) (8) (9) Overall Employment Growth AI vacancy share in 2015 -0.002 (0.052) -0.042 (0.055) -0.042 (0.055) -0.045 (0.054) -0.046 (0.055) -0.046 (0.055) -0.040 (0.055) -0.052 (0.051) -0.065 (0.051) Observations 32082 32082 32082 32082 32082 32082 32082 32082 32082 Unskilled jobs AI vacancy share in 2015 0.168 (0.160) 0.232 (0.163) 0.232 (0.163) 0.231 (0.163) 0.231 (0.163) 0.204 (0.163) 0.233 (0.163) 0.219 (0.165) 0.206 (0.164) Observations 32388 32388 32388 32388 32388 32388 32388 32388 32388 Skilled jobs AI vacancy share in 2015 -0.001 (0.063) -0.008 (0.061) -0.008 (0.061) -0.011 (0.060) -0.010 (0.061) -0.015 (0.062) -0.005 (0.062) -0.022 (0.059) -0.028 (0.059) Observations 31893 31893 31893 31893 31893 31893 31893 31893 31893 Complex jobs AI vacancy share in 2015 0.138 (0.086) 0.035 (0.083) 0.035 (0.083) 0.031 (0.083) 0.037 (0.083) 0.021 (0.082) 0.038 (0.083) 0.009 (0.083) 0.004 (0.082) Observations 32034 32034 32034 32034 32034 32034 32034 32034 32034 Highly complex jobs AI vacancy share in 2015 0.204∗∗ (0.084) 0.139 (0.086) 0.139 (0.086) 0.135 (0.084) 0.141∗ (0.085) 0.131 (0.081) 0.134 (0.084) 0.139 (0.086) 0.126 (0.081) Observations 32483 32483 32483 32483 32483 32483 32483 32483 32483 Covariates: Further vacancy posting Average AI vacancy share 2016-19 no yes yes yes yes yes yes yes yes Number of all vacancies in 2015 no no yes no no no no no yes Establishment properties AKM effects 2010-2017 no no no yes no no no no yes ... referring to 2015 Establishment size no yes yes yes yes yes yes yes yes Establishment age no no no no yes no no no yes Economic sectors no no no no no yes no no yes Federal states no no no no no no yes no yes Occupational shares no no no no no no no yes yes Notes: This table reports the estimation results for regressing the overall employment growth and employment growth in jobs differentiated by skill requirement level on the AI vacancy share in 2015. Unskilled jobs require no formal qualification or only short term training. Skilled jobs require a formal vocational education training of at least 2 years. Complex jobs require a university degree or master craftman’s certificate. Highly complex jobs require a university degree or similar and, beyond that, profound professional experience or further formal highly specialised qualification certificates like a doctorate or a habilitation. Establishments in ICT and professional services are excluded. For regressions with the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Integrated Employment Biographies (IEB). IAB-Discussion Paper 15|2024 36 Overall, in line with our analysis for the relationship of the skill change of establishments and AI activity, we do not find supportive evidence for sizeable displacement effects. In the sample considering all establishments we find a higher employment growth in highly complex jobs. This may reflect the requirement of additional employees for jobs where the development or implementation of AI technologies are the main tasks. Our finding is also in line with recent literature that show that highly educated workers tend to benefit from AI activity (Albanesi et al. 2023, Babina et al. 2023) in terms of employment growth. 5 Robustness Checks 5.1 Employment and employees’ qualification As a first robustness check we change from the job requirement perspective to the employee perspective. Instead of considering the different required skill levels we now define the worker groups according to the employees’ qualification levels. A change of the number of jobs in a certain required skill level must not necessarily coincide with a change in the number of employees with the corresponding qualification level. The reason is that establishments occasionally employ individuals on jobs that require normally a qualification that is lower or a higher than the individual’s qualification (see, for instance, Rohrbach-Schmidt and Tiemann 2016, Erdsiek 2021). In what follows, we present estimates of employment growth for worker groups defined by the workers’ reached qualification level. Hereby we can distinguish unskilled employees without a formal qualification, qualified employees with a certified vocational educational training of at least 2 years, and highly qualified employees with a university degree or similar. Table 10 presents the results for all establishments and Table 11 presents the results for establishments without AI producing sectors ICT and professional services. Again, the first panel reports results for overall employment growth and, therefore, repeats the previous results in the first panels of Table 8 or Table 9, respectively. IAB-Discussion Paper 15|2024 37 Table 10: Employment growth 2015 - 2019 by employee’s qualification levels, and AI skills demand, all establishments Leere Zelle (1) (2) (3) (4) (5) (6) (7) (8) (9) Overall Employment Growth AI vacancy share in 2015 0.055 (0.049) 0.006 (0.049) 0.006 (0.049) 0.002 (0.048) -0.003 (0.049) -0.007 (0.047) 0.006 (0.048) -0.016 (0.045) -0.030 (0.045) Observations 33771 33771 33771 33771 33771 33771 33771 33771 33771 Unskilled Employment Growth AI vacancy share in 2015 -0.043 (0.101) -0.075 (0.101) -0.075 (0.101) -0.076 (0.101) -0.082 (0.101) -0.089 (0.100) -0.074 (0.101) -0.093 (0.100) -0.101 (0.101) Observations 34210 34210 34210 34210 34210 34210 34210 34210 34210 Qualified Employment Growth AI vacancy share in 2015 0.036 (0.048) -0.014 (0.049) -0.014 (0.049) -0.019 (0.048) -0.019 (0.049) -0.028 (0.047) -0.013 (0.048) -0.039 (0.045) -0.049 (0.045) Observations 33918 33918 33918 33918 33918 33918 33918 33918 33918 Highly Qualified Employment Growth AI vacancy share in 2015 0.203∗∗∗ (0.055) 0.074 (0.063) 0.074 (0.063) 0.068 (0.061) 0.075 (0.063) 0.047 (0.061) 0.055 (0.062) 0.069 (0.062) 0.037 (0.059) Observations 33611 33611 33611 33611 33611 33611 33611 33611 33611 Covariates: Further vacancy posting Average AI vacancy share 2016-19 no yes yes yes yes yes yes yes yes Number of all vacancies in 2015 no no yes no no no no no yes Establishment properties AKM effects 2010-2017 no no no yes no no no no yes ... referring to 2015 Establishment size no yes yes yes yes yes yes yes yes Establishment age no no no no yes no no no yes Economic sectors no no no no no yes no no yes Federal states no no no no no no yes no yes Occupational shares no no no no no no no yes yes Notes: This table reports the estimation results for regressing overall employment growth and employment growth in jobs differentiated by qualification levels on the AI vacancy share in 2015. All sectors are included but we exclude temporary work agencies. For regressions with the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Integrated Employment Biographies (IEB). IAB-Discussion Paper 15|2024 38 Table 11: Employment growth 2015 - 2019 by employee’s qualification levels, and AI skills demand, Establishments w/o AI producing sectors Leere Zelle (1) (2) (3) (4) (5) (6) (7) (8) (9) Overall Employment Growth AI vacancy share in 2015 -0.002 (0.052) -0.042 (0.055) -0.042 (0.055) -0.045 (0.054) -0.046 (0.055) -0.046 (0.055) -0.040 (0.055) -0.052 (0.051) -0.065 (0.051) Observations 32082 32082 32082 32082 32082 32082 32082 32082 32082 Unskilled Employment Growth AI vacancy share in 2015 -0.138 (0.109) -0.149 (0.106) -0.149 (0.106) -0.150 (0.107) -0.153 (0.107) -0.154 (0.107) -0.145 (0.107) -0.159 (0.107) -0.162 (0.109) Observations 32454 32454 32454 32454 32454 32454 32454 32454 32454 Qualified Employment Growth AI vacancy share in 2015 0.016 (0.055) -0.015 (0.056) -0.015 (0.056) -0.019 (0.054) -0.018 (0.057) -0.023 (0.054) -0.014 (0.056) -0.029 (0.050) -0.040 (0.050) Observations 32229 32229 32229 32229 32229 32229 32229 32229 32229 Highly Qualified Employment Growth AI vacancy share in 2015 0.198∗∗∗ (0.070) 0.074 (0.081) 0.074 (0.081) 0.070 (0.078) 0.075 (0.081) 0.062 (0.077) 0.068 (0.079) 0.071 (0.078) 0.051 (0.074) Observations 31837 31837 31837 31837 31837 31837 31837 31837 31837 Covariates: Further vacancy posting Average AI vacancy share 2016-19 no yes yes yes yes yes yes yes yes Number of all vacancies in 2015 no no yes no no no no no yes Establishment properties AKM effects 2010-2017 no no no yes no no no no yes ... referring to 2015 Establishment size no yes yes yes yes yes yes yes yes Establishment age no no no no yes no no no yes Economic sectors no no no no no yes no no yes Federal states no no no no no no yes no yes Occupational shares no no no no no no no yes yes Notes: This table reports the estimation results for regressing overall employment growth and the employment growth in jobs differentiated by qualification levels on the AI vacancy share in 2015. Establishments from ICT and professional services are excluded. Additionally, we exclude temporary work agencies. For regressions with the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Integrated Employment Biographies (IEB). IAB-Discussion Paper 15|2024 39 The further panels show results for employment growth by the different qualification levels. Overall, we see only small differences of the employment growth rates related to AI activity: first, the point estimates for unskilled employment growth now turn negative in the two tables implying a lower employment growth for unskilled employees in establishment with AI activity. However, the effects are insignificant. For qualified employment growth we find similar results as for the required skill levels. The point estimates of the growth rate differences are also negative and of similar magnitude as for the skilled jobs. The second difference is that the positive employment growth for the highly qualified employees turns insignificant based on the sample with all establishments in the fully specified model (col. 9 in each of the Table 8 or Table 9). Moreover, the magnitude of the growth rate difference is much smaller for all establishments (0.037, col. 9 in Table 8). When we exclude establishments from ICT and professional services, the point estimate has a magnitude of 0.051 (col. 9 in Table 9). However, both estimates are insignificant. The results again point to the absence of sizeable displacement effects. However, in contrast to our main analysis we find no higher growth in highly qualified employment. This may be driven by the fact that the group of highly qualified employees entails a larger group than employees in highly complex jobs; because formally the group of highly qualified employees comprises a mix of qualifications that allow individuals access to either complex or highly complex jobs. 5.2 Exploiting the panel dimension of the data Next, we exploit the panel structure of our data to estimate a short run relationship between AI activity and employment growth. Thereby, we test whether the results in the main analysis change if we exploit further variation in the AI vacancy share from other years. Since our job ads data are repeated cross sections, we do not observe all job postings of an establishment in each year. Hence, we have a unbalanced panel of establishments and their job posting activity in each year. Moreover, together with the small time span we can consider, which is five years, we cannot include a large number of lags in the model (e.g., Babina et al. 2022 find that AI investments translate into employment growth after about three years). However, in line with our previous analyses the findings from this exercise further provide no evidence for sizeable displacement effects. We estimate the following model as a robustness check: ∆𝑦, = 𝛼+ 𝛽 𝑣,  𝑣,  +𝑥′,𝛾+𝜖, where we test for a relationship of the AI vacancy share in 𝑡 − 1 and employment growth from the same period 𝑡−1 to the next period 𝑡. The vector 𝑥′, the same covariates as in the main analysis and we additionally control for year fixed effects. We start by estimating equation (3) with pooled OLS and contrast the results with the results from using a fixed effects estimator. Moreover, to address the unbalanced nature of our panel we estimate the same model but IAB-Discussion Paper 15|2024 40 restrict the data to establishments for which we have at least three observations. As for our main analysis we measure employment growth in per cent and we estimate standard errors that are robust to heteroscedasticity. Table 12: Employment growth and AI skills demand in the panel data set, all establishments Leere Zelle OLS (1) OLS (2) OLS (3) FE (4) FE (5) Overall Employment AI vacancy share (t-1) 0.016∗∗ (0.007) 0.014∗ (0.007) 0.013 (0.008) -0.000 (0.011) -0.000 (0.011) Observations 448460 448460 321464 448460 321464 Unskilled jobs AI vacancy share (t-1) 0.016 (0.012) 0.015 (0.012) 0.004 (0.016) 0.018 (0.029) 0.012 (0.030) Observations 446619 446619 320507 446619 320507 Skilled jobs AI vacancy share (t-1) 0.013 (0.010) 0.011 (0.010) 0.010 (0.011) -0.001 (0.018) -0.001 (0.018) Observations 454456 454456 324268 454456 324268 Complex jobs AI vacancy share (t-1) 0.017 (0.012) 0.016 (0.012) 0.011 (0.014) 0.002 (0.022) -0.004 (0.022) Observations 442545 442545 318002 442545 318002 Highly complex jobs AI vacancy share (t-1) 0.012 (0.010) 0.012 (0.010) 0.022∗ (0.012) -0.008 (0.021) -0.017 (0.019) Observations 439970 439970 317487 439970 317487 Covariates: Fixed Effects Year fixed effects yes yes yes yes yes Further vacancy posting (t-1) Number of all vacancies yes yes yes yes yes Establishment properties AKM effects 2010-2017 no yes yes no no ... referring to t-1 Establishment size yes yes yes yes yes Establishment age yes yes yes yes yes Economic sectors yes yes yes yes yes Federal states yes yes yes yes yes Occupational shares yes yes yes yes yes Observations requirement Restricted to ≥ 3 observations no no yes no yes Notes: This table reports the estimation results for regressing the overall employment growth and employment growth in jobs differentiated by skill requirement level on the AI vacancy share using panel data. Unskilled jobs require no formal qualification or only short term training. Skilled jobs require a formal vocational education training of at least 2 years. Complex jobs require a university degree or master craftman’s certificate. Highly complex jobs require a university degree or similar and, beyond that, profound professional experience or further formal highly specialised qualification certificates like a doctorate or a habilitation. All industries are included. For regressions with the employment growth rates we exclude establishments with the 5 per cent highest growth rates. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.10 Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Integrated Employment Biographies (IEB). IAB-Discussion Paper 15|2024 47 A Appendix on additional tables Table A1: Overview of the AI skills from the AI dictionary Number AI skill 1 AI ChatBot 2 AI KIBIT 3 ANTLR 4 AWS Panorama 5 AdaBoost algorithm 6 Affective Computing 7 Amazon CodeGuru 8 Amazon Comprehend 9 Amazon Comprehend Medical 10 Amazon DevOps Guru 11 Amazon Forecast 12 Amazon Fraud Detector 13 Amazon HealthLake 14 Amazon Kendra 15 Amazon Lex 16 Amazon Lookout für Equipment 17 Amazon Lookout für Metrics 18 Amazon Lookout für Vision 19 Amazon Monitron 20 Amazon Omics 21 Amazon Personalize 22 Amazon Polly 23 Amazon Rekognition 24 Amazon SageMaker 25 Amazon Textract 26 Amazon Transcribe 27 Amazon Translate 28 Apertium 29 Applicant Tracking System 30 Artificial Intelligence 31 Augmented Analytics 32 Automated Driving 33 Automated optical inspection (AOI) 34 Autonomous Driving 35 Autonomous Systems 36 Azure AI Content Safety 37 Azure Anomaly Detector 38 Azure Bot Service 39 Azure Cognitive Search 40 Azure Cognitive Services 41 Azure Content Moderator 42 Azure Custom Vision 43 Azure Data Science Virtual Machines 44 Azure Databricks 45 Azure Form Recogniser 46 Azure Health Bot 47 Azure Immersive Reader 48 Azure Kinect DK 49 Azure Language Understanding (LUIS) 50 Azure Machine Learning 51 Azure Metrics Advisor 52 Azure Open Datasets 53 Azure OpenAI Service 54 Azure Personaliser 55 Azure Project Bonsai 56 Azure QnA Maker 57 Azure Speaker Recognition 58 Azure Speech translation 59 Azure Speech-to-Text 60 Azure Translator IAB-Discussion Paper 15|2024 48 Number AI skill 61 Azure Video Indexer 62 Bayesian optimization 63 BindsNET 64 Blue Prism 65 Boosting 66 Business intelligence 67 Caffe 68 Character generation 69 Character recognition 70 ChatGPT 71 Chi Square Automatic Interaction Detection (CHAID) 72 Classification Algorithms 73 Clustering Algorithms 74 Cognitive Computing 75 Colab 76 Collaborative Filtering 77 Computational Linguistics 78 Computer Vision 79 Curated Shopping 80 DALL-E 81 Dauerstrichradar 82 Decision Trees 83 Deep Learning 84 Deeplearning4j 85 Dimensionality Reduction 86 Direction generation 87 Direction recognition 88 Distinguo 89 Electromechanical Systems 90 Embedded Vision 91 Environment Perception 92 Expert System 93 Face generation 94 Face recognition 95 Feature Extraction 96 GPT-1 97 GPT-2 98 GPT-3 99 GPT-4 100 Generative Adversarial Networks 101 Google AI Infrastructure 102 Google AutoML 103 Google Cloud Machine Learning Platform 104 Google Contact Center AI 105 Google Dialogflow 106 Google Document AI 107 Google Media Translation 108 Google Natural Language API 109 Google Recommendations AI 110 Google Text-to-Speech 111 Google Translation AI 112 Google Vertex AI 113 Google Video AI 114 Google Vision AI 115 Gradient boosting 116 H2O 117 IBM Cloud Paks 118 IBM Watson 119 IPSoft Amelia 120 Image Processing 121 Image Recognition 122 Image Tagging 123 Image generation 124 Information Extraction 125 Ithink 126 KNIME 127 Keras 128 Kernel Methods 129 Knowledge Engineering IAB-Discussion Paper 15|2024 49 Number AI skill 130 Knowledge Extraction 131 Knowledge Representation and Reasoning 132 Kubeflow 133 Latent Dirichlet Allocation 134 Latent Semantic Analysis 135 Least absolute shrinkage and selection operator 136 Legal Technology 137 Lexalytics 138 Lexical Acquisition 139 Lexical Semantics 140 Libsvm 141 Lidar 142 Long Short-Term Memory (LSTM) 143 MLPACK (C++ library) 144 MLlib 145 MXNet 146 Machine Learning 147 Machine Learning Operations (MLOps) 148 Machine Translation 149 Machine Vision 150 Madlib 151 Mahout 152 Mask R-CNN 153 Matplotlib 154 Microsoft Cognitive Toolkit 155 Mlflow 156 Mlpy 157 MoSes 158 Modular Audio Recognition Framework 159 Motion Planning 160 Motoman Robot Programming 161 ND4J (software) 162 Natural Language Inference 163 Natural Language Processing 164 Natural Language Toolkit 165 Natural Language Understanding 166 Nearest Neighbor Algorithm 167 Neural Networks 168 Neuromorphic Computing 169 Numpy 170 Object Recognition 171 Object Tracking 172 OpenCV 173 OpenNLP 174 Path Planning 175 Pattern Recognition 176 Perceptron 177 Predictive Maintenance 178 Predictive Models 179 Pybrain 180 Random Forests 181 RapidMiner 182 Recommender Systems 183 Reinforcement Learning 184 Remote Sensing 185 Robot Framework 186 Robot Operating System (ROS) 187 Robot Programming 188 Robot learning 189 Robotic Process Automation 190 Robotic Systems 191 Semantic Driven Subtractive Clustering Method (SDSCM) 192 Sentiment Analysis / Opinion Mining 193 Sentiment Classification 194 Servo Drives/Motors 195 Shogun 196 Simultaneous Localization and Mapping (SLAM) 197 Speech Recognition 198 Speech generation IAB-Discussion Paper 15|2024 50 Number AI skill 199 Stochastic Gradient Descent 200 Superml 201 Supervised Learning 202 Support Vector Machines 203 TensorFlow 204 TensorQuant 205 Text Mining 206 Text generation 207 Text recognition 208 Text to Speech 209 Theano 210 Tokenization 211 Torch 212 Unsupervised Learning 213 Video generation 214 Video processing 215 Video recognition 216 Video-based Driver Assistance Systems 217 Virtual Agents 218 Visual inventory management 219 Voicebot 220 Vowpal Wabbit 221 Weka 222 Word2Vec 223 Xgboost 224 Zero-shot learning 225 alteryx 226 kernLab 227 mlr3 228 pytorch 229 scikit-learn 230 spaCy 231 uipath IAB-Discussion Paper 15|2024 51 B Appendix on additional figures Figure A1: AI establishment shares and AI vacancy shares Note: Vacancies from temporary work agencies are excluded. The AI vacancy share is defined as the share of vacancies containing at least one AI skill in the job description on all vacancies. The AI establishment share is calculated by dividing the number of establishments that post at least one AI vacancy in a given year by the number of all establishments in our sample. Both shares are measured in per cent. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. IAB-Discussion Paper 15|2024 52 Figure A2: AI establishment shares weighted by employment Note: Vacancies from temporary work agencies are excluded. The AI establishment share is calculated by dividing the number of establishments that post at least one AI vacancy in a given year by the number of all establishments in our sample. The shares are measured in per cent. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Figure A3: Shares of establishments posting at least one AI vacancy across industries Note: Vacancies from temporary work agencies are excluded. Establishments with overall employment growth above the 95th percentile are excluded. The AI establishment shares are calculated by dividing the number of establishments that post at least one AI vacancy in a given year and by the number of all establishments in a given industry. The shares are measured in per cent. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). IAB-Discussion Paper 15|2024 53 C Appendix on data representativeness We compare the structure of the job ads data with the structure of the IAB JVS to get information on how representative our job ads data is. Both data sets can be characterised as cross sectional data sets. Our job ads data refers to mid of October of each year. The IAB JVS refers to an unspecific point in time in the 4th quarter of each year. Therefore both data refer roughly to the same time period. We compare the vacancy shares across industries and required skill levels. Figure A4 shows vacancy shares of the job ads data and the IAB JVS for 2015 across economic sectors. In 2015 there are only slight differences in the sectoral shares of the job ads data and the IAB JVS. The Information and Communication Technology (ICT) sector seems to be slightly underrepresented in the job ads data. In contrast, the trade and car maintanance, the construction and the manufacturing sector are slightly overrepresented. The next Figure Figure A5 shows the corresponding sectoral shares for 2019. In 2019 the overrepresentation in the manufacturing and the trade and car maintenance sector is larger compared 2015. Simultaneously, other services (besides professional services) are now more underrepresented. Regarding the required skill levels, in the IAB JVS the information on ”complex jobs” and ”highly complex jobs” are aggregated. For the comparison, we, therefore, also aggregated the information on these skill levels in our job ads data. We find only small differences (Figure A6). As for the sectoral shares, in 2015 the distributions of the of the IAB JVS and the job ads data across required skill levels match very closely. In 2019 ”skilled jobs” are slightly underrepresented in the while the group of complex and ”highly complex jobs” are slightly overrepresented in the job ads data compared to the IAB JVS. Figure A4: Industry shares in the vacancy data and the IAB Job Vacancy Survey (2015) Note: Data on the vacancies are from the BA-JOBBÖRSE. Vacancies from temporary work agencies are excluded. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. IAB Job Vacancy Survey (see Bossler et al. 2022 for details). IAB-Discussion Paper 15|2024 54 Figure A5: Industry shares in the vacancy data and the IAB Job Vacancy Survey (2019) Note: Data on the vacancies are from the BA-JOBBÖRSE. Vacancies from temporary work agencies are excluded. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. IAB Job Vacancy Survey (see Bossler et al. 2022 for details). IAB-Discussion Paper 15|2024 55 Figure A6: Shares in the vacancy data and the IAB Job Vacancy Survey across required skill levels (2015 and 2019) (a) 2015 (b) 2019 Note: Vacancies from temporary work agencies are excluded. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. IAB Job Vacancy Survey (see Bossler et al. 2022 for details). D Appendix on establishment properties In this section we describe the distribution of establishment property variables across establishments either with or without AI activity in 2015 more in detail (see also section 3.2 of the main text). Figure A7 shows the establishment shares posting between 1-9, 10-19, 20-49, 50-199, 200-499 or 500 and more vacancies for both groups. Establishments with AI activity are less likely to post 1-9 vacancies but more likely to post between 10-19, 20-49 or 50-199 vacancies in 2015. However, there are no establishments with AI activity with 200-499 or 500 and more posted vacancies while there are very few other establishments with overall vacancies within these ranges. IAB-Discussion Paper 15|2024 56 Figure A7: Distribution of overall vacancy postings 2015 across establishments Note: Vacancies from temporary work agencies are excluded. Establishments with overall employment growth above the 95th percentile are excluded. Source: JOBBÖRSE of the German Federal Employment Agency (FEA). Job ads with full support by the FEA. Cross sections for the years 2015 to 2019 with a reference date of October 15th of each year. Establishment History Panel (BHP). Figure A8 shows the distribution of establishments across ten AKM effects 2010-2017 value bins. Those bins, with exception of the two bins containing the minimum and the maximum value, include establishments within an AKM effect value range of 0.1 log points. The distribution of establishments with AI vacancy posting in 2015 along AKM effects is clearly right skewed compared to the distribution of other establishments. This indicates that establishments with AI activity at this early stage of development of AI tend to pay higher establishment-specific wage premia to their employees than other establishments in our sample. Imprint IAB-Discussion Paper 15|2024 Date of publication November 22, 2024 Publisher Institute for Employment Research of the Federal Employment Agency Regensburger Str. 104 90478 Nürnberg Germany Rights of use This publication is published under the following Creative Commons Licence: Attribution – ShareAlike 4.0 International (CC BY-SA 4.0) https://creativecommons.org/licenses/by-sa/4.0/deed.de Download of this IAB-Discussion Paper https://doku.iab.de/discussionpapers/2024/dp1524.pdf All publications in the series “IAB-Discussion Paper” can be downloaded from https://iab.de/en/publications/iab-publications/iab-discussion-paper-en/ Website https://iab.de/en/ ISSN 2195-2663 DOI 10.48720/IAB.DP.2415 Corresponding author Dr. Michael Stops Phone: +49 911 179-4591 Email: [email protected]