AI skills on the labour market
Authors:
Lennert Peede, Michael Stops
Abstract:
Our study contributes to the question what job advertisements reveal about Artificial Intelligence (AI) activity among German firms. To do so, we analyze the demand for those skills on the German labor market, which are required to adopt, apply or develop AI methodologies (AI skills henceforth). Furthermore, we analyze how firm specifics and the labor demand for particular groups of workers are related to this. Particularly, we pursue three leading questions. Firstly, we analyze how pronounced AI activity is in Germany and how firms with AI activity are characterized. Secondly, we assess whether current AI activity is driven by automation. Thirdly, we test whether automation by AI leads to sizeable displacement effects. Our study is work in progress; in our eventual presentation, we also discuss our further analysis plan.
In the recent years machine learning algorithms became more and more technically feasible and this resulted, together with an increased availability of “big data”, in a large number of innovations in the area of AI. Brynjolfsson, Rock, and Syverson (2019) even claim that AI has the potential to become a ‘general purpose technology’ in the future. Cockburn, Henderson, and Stern (2019) document that AI activities foster innovation and development of new businesses and services.
At the same time the implications of AI for the labor markets are widely discussed. On the one hand, some studies argue that AI is mainly utilized to simulate human intelligence and, therefore, to perform tasks that were previously performed by human beings (Taddy 2019). On the other hand, other studies refer to AI as a technology that can potentially exceed humans’ cognitive capabilities. Thus, such AI technology would be complementary to workers’ tasks implying new innovations (Brynjolfsson, Rock, and Syverson 2019). Both perspectives may be true, because AI as technology is broadly defined and its implications on existing work places may be conditional on its specific forms, purposes, and how the workplaces are organized (Acemoglu and Restrepo 2020).
Only recently the technical opportunities to collect recent, large, and highly disaggregated data sets were considerably expanded and allowed first efforts to determine the prevalence of AI use. The studies by Alekseeva et al. (2021); Babina et al. (2020), and Acemoglu et al. (2022) belong to a strand of literature that aims at a systematic exploration whether there has indeed been a major increase in AI adoption as opposed to just extensive media coverage of AI. These studies are based on job advertisement data for the USA and argue that this kind of data renders it possible to find examples of AI technologies that either complement workers’ tasks or replace them. In contrast, job ads data for the German labor
market as a source for empirical labor market research has been currently very seldom exploited and, to the best of our knowledge, never for identifying AI activities in German firms. Our study aims to close these gaps and, beyond that, makes use of the opportunity to directly link this data with rich administrative firm data.
In doing so, we observe signals for the AI adoption of firms in Germany and its implications. AI adoption means that firms begin to implement AI technology, either in order to complement or to displace tasks previously performed by workers. The information in job ads referring to the requirement of AI-related activities can be interpreted as such signals of AI adopting firms (see also Acemoglu et al. (2022)).
In the first step, we set up a unique dictionary on AI terms for German job ads to measure the extent of the demand for AI skills in job advertisements. Hereby, we refer to well documented AI technologies listed by a row of information sources and combine these to a dictionary that is more comprehensive than each of these sources alone (see, for reference, Acemoglu et al. (2022) Alekseeva et al. (2021), Taddy (2019), and OECD.AI (2022)). Our dictionary closely captures the standard keywords which relate to state-of-the-art machine- learning driven AI technologies. We then assign the AI-related skills as proposed by Squicciarini and Nachtigall (2021) to the following categories: Generic AI keywords, AI applications, AI approaches, and AI-related software and libraries. By assessing the relative occurrence of these distinctive categories over time we draw conclusions with respect to the stage of development of AI technology in Germany. For instance, a higher occurrence of specific software and libraries in contrast to generic keywords in job descriptions indicates a higher level of specialization within a firm.
We then use job ads text data published and provided by the JOBBÖRSE of the Federal Employment Agency (BA-JOBBÖRSE). The BA-JOBBÖRSE is one of the largest online job portals in Germany. It is free of charge for job seekers as well as for companies. Job openings can be searched for directly (without registration) on the BA-JOBBÖRSE site1.
We use cross sectional samples of all job advertisements referring to job offers in the years 2013 and later for 2015 to 2019 and actively maintained by the Federal Employment Agency’s public placement service. According to this selection, more than 2.9 million advertisements for 4.6 million vacancies are considered in total. They include a minimum of 238 thousand advertisements for 380 thousand vacancies for the first observation in 2013 and a maximum of 474 thousand advertisements for 716 thousand vacancies in 2019.
In general, establishments can create one job ad for multiple jobs with the same characteristics for different work locations. If only the number of job ads is considered, labor demand would be quantitatively underestimated and regional assignment would often not be possible. However, the magnitude of labor demand must be taken into account when determining the importance of the signal words in job ads. Therefore, in our analysis, we weight each job ad by the number of jobs behind it for each work location. The availability of such information is a big advantage of our data compared with job ads data that is crawled or scraped from the web, because this data comes without such information.
Dictionary and job ads data texts are further prepared for the analysis according to Stops et al. (2021): The texts in the job descriptions undergo various pre-processing steps that include, e.g., standardization, stemming, and tokenization. Furthermore, a segmentation procedure was conducted in order to extract the relevant part of the job ads text for the analysis of the AI skill terms: In addition to the self-introduction of the company or business, job advertisements contain the job description of a position; the requirements that the sought-after skilled workers must bring with them; and functional parts with legal references, for example, regarding the consideration of certain groups of people or information on the further application procedure. We are particularly interested in the job description of the position and the requirements that the sought-after workers must bring with them as it most precisely should reveal true preferences for AI-related skills and hence AI activity. With the help of a machine learning algorithm-based classification procedure, we therefore mark the part of the job advertisement texts relevant for us accordingly as “job description” and distinguish it from the remaining part being not relevant for the analyses in this study.
With this data at hand, we identify the occurrence of AI skill terms and compute the ratio of the absolute occurrences of AI skill terms in the job ads by occupations and regions. We finally link these data with rich administrative data on firm specifics provided by the Federal Employment Agency. Using administrative data at the establishment level allows us track establishments’ performance more precisely than with pure vacancy data as done so far in the literature.
So far, we find increasing AI activity as both the number and the ratio of AI vacancies relative to all vacancies increase over the sample period. We further find that the relative occurrence of AI vacancies is rather low compared with the US labor market (compare with Acemoglu et al. (2022)). More specifically, firms increasingly post skill requirements for generic AI-related keywords but seldom require specific skills related to respective software, libraries or approaches. This let us suppose, that firms may be interested on employees with general AI capabilities expecting that these experts may foster the AI adoption in these firms. All in all, we conclude that AI adoption in Germany is in an early stage of development and German firms do not seem to be rather specialized yet.
Moreover, evaluating the sectoral decomposition of AI vacancies and the vacancies’ required skill levels indicate that a relatively large part of AI activity is complementing workers. We hypothesize that tasks with relatively low skill requirements are more susceptible for automation. Since AI vacancies are mainly posted for jobs requiring relatively high skill levels, we conclude that AI adoption mainly introduces new tasks for specialized workers. Additionally, AI activity mainly takes places in AI-producing sectors (i.e., Information and Communications Technology and Professional Services) instead of AI-using sectors. In these sectors AI is expected to introduce new tasks and is not used for automation. To support this evidence, we test whether a task content, which is susceptible for automation by AI, is related to AI activity at the establishment level using regression techniques following Acemoglu et al. (2022). We find this relationship to be weak and conclude that AI activity is currently not driven by establishments adopting AI to automate their task content.
Albeit we argued so far that AI activity is not mainly driven by automation, we still observe a considerable amount of AI vacancies with low skill requirements. According to the plausible assumption that tasks with lower complexity tend to be easier automatable (cf. Acemoglu and Restrepo (2018)), we interpret this as potential signal that at least some worker groups are affected by automation with AI. Hence, we further plan to analyze whether automation with AI even at an early stage of development leads to sizeable displacement effects.
Due to the linked administrative data, we are able to expand this analysis with precise employment outcomes. Hereby, in the further course of the study we will analyze the establishments’ overall employment as well as the establishments’ employment in each skill group.
Generally, our study is work in progress. Besides the plan to add the employment level analyses, we will also conclude our eventual presentation by discussing the further research plan that includes the exploration of the mechanisms behind the observed relationships from both a theoretical and empirical perspective and, consequently, utilizes adequate micro data to validate our findings.
References
Acemoglu, Daron, David Autor, Jonathon Hazell, and Pascual Restrepo. 2022. “Artificial Intelligence and Jobs: Evidence from Online Vacancies.” Journal of Labor Economics, 40(S1): S293-S340.
Acemoglu, Daron, and Pascual Restrepo. 2018. “The Race between Man and Machine.” American Economic Review, 108(6): 1488–542.
Acemoglu, Daron, and Pascual Restrepo. 2020. “The wrong kind of AI? Artificial intelligence and the future of labour demand.” Cambridge Journal of Regions, Economy and Society, 13(1): 25–35.
Alekseeva, Liudmila, José Azar, Mireia Giné, Sampsa Samila, and Bledi Taska. 2021. “The demand for AI skills in the labor market.” Labour Economics, 71: 1–27.
Babina, Tania, Anastassia Fedyk, Alex X. He, and James Hodson. 2020. “Artificial Intelligence, Firm Growth, and Industry
Concentration.” SSRN Electronic Journal.
Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. 2019. “Artificial Intelligence and the Modern Productivity Paradox. A Clash of Expectations and Statistics.” In National Bureau of Economic Research Conference Report, The Economics of Artificial Intelligence. An Agenda, ed. Ajay Agrawal, Joshua Gans, and Avi Goldfarb, 23–57.
Cockburn, Ian, Rebecca Henderson, and Scott Stern. 2019. “The Impact of Artificial Intelligence on Innovation: An Exploratory Analysis.” In National Bureau of Economic Research Conference Report, The Economics of Artificial Intelligence. An Agenda, ed. Ajay Agrawal, Joshua Gans, and Avi Goldfarb, 115–46.
OECD.AI. 2022. “AI jobs and skills. Live Data.” OECD AI Policy Observatory. https://oecd.ai/en/data?selectedArea=ai-jobs- and-skills.
Squicciarini, Mariagrazia, and Heike Nachtigall. 2021. “Demand for AI skills in jobs: Evidence from online job postings.”
OECD Science, Technology and Industry Working Papers No. 2021/03.
Stops, Michael, Ann-Christin Bächmann, Ralf Glassner, Markus Janser, Britta Matthes, Lina-Jeanette Metzger, Christoph Müller, and Joachim Seitz. 2021. “Extracting Skill Requirements from Job Ads – the “Machbarkeitsstudie Kompetenz- Kompass”” IAB Forschungsbericht 7/2021. https://doku.iab.de/forschungsbericht/2021/fb0721.pdf.
Taddy, Matt. 2019. “The Technological Elements of Artificial Intelligence.” In National Bureau of Economic Research Conference Report, The Economics of Artificial Intelligence. An Agenda, ed. Ajay Agrawal, Joshua Gans, and Avi Goldfarb, 61–88.