Artificial intelligence and the wellbeing of workers

Abstract

This study explores the relationship between artificial intelligence (AI) and workers’ well-being and health using longitudinal survey data from Germany (2000–2020). Using a measure of occupational exposure to AI, we explore an event study design and a difference-in-differences approach to compare AI-exposed and non-exposed workers. Before AI became widely available, there is no evidence of differential pre­trends in workers’ well-being and health. We find no evidence of a sizeable negative impact of AI on workers’ well-being and mental health. If anything, there is evidence of an improvement in health status and health satisfaction, which may be explained by the decline in job physical intensity. Overall, our results are consistent with the lack of negative effects of AI on the labor markets.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-98241-3.

Introduction

Over the past few years, there has been a striking increase in the adoption of artificial intelligence (Al) by firms worldwide. The emergence of generative AI, like ChatGPT, has brought about a substantial surge in public interest in AI. Media and scholars have highlighted how this new technology has the potential to reshape our everyday existence as well as our cognitive and professional processes¹. The level of investment in AI is increasing at high rates². As of 2022, approximately 50% of companies reported using AI technologies in at least one business area³. AI may have transformative impacts on economic growth, health care, safety, and transportation and may reduce the costs and barriers to information access, education, and training^4–6. Similar to other technological changes, AI can help reduce work-related risks⁷. Recent studies have investigated the impact of AI on labor market outcomes and productivity^2,8–11. However, to the best of our knowledge, no existing study has examined the impact of AI on workers’ well-being and health using longitudinal survey data. This study therefore aims to address this gap in the literature.

In this study, we examine how AI adoption in the workplace affects workers’ well-being, economic concerns, and health using longitudinal data from the German Socio-Economic Panel (SOEP). Germany provides a compelling case for studying the labor market effects of technological change due to its strong labor institutions, including unions and extensive employment protection legislation^12–14. These institutions play a critical role in shaping AI adoption by negotiating terms that mitigate worker displacement and facilitate equitable transitions. The role of works councils and worker participation in firm governance has also been emphasized in recent research^15,16. In contrast, countries like the US, where labor protections are weaker, experience more abrupt disruptions and greater worker vulnerability^13,17. Germany’s industrial structure, with its strong foundation in high-skill manufacturing (e.g., automotive, machinery) and specialized services (e.g., finance, IT), shapes the way AI is adopted across sectors. If AI-driven automation were to significantly impact labor markets in Germany, its effects might vary depending on sector-specific characteristics and workforce training systems. In manufacturing, AI could plausibly serve as a tool for precision and efficiency, complementing rather than replacing skilled labor—potentially due to the country’s strong vocational training system and emphasis on high-value production. However, in contrast, certain service sectors, such as customer support, might be more susceptible to automation-driven job restructuring, depending on how AI adoption unfolds across different industries. More broadly, the extent to which AI displaces or augments jobs may hinge on factors such as task complexity, the need for human oversight, and the adaptability of training programs to evolving technological demands. By analyzing these sectoral variations within Germany’s distinct institutional framework, our study provides nuanced insights into how AI adoption interacts with labor protections, skill development, and economic structures to influence employment outcomes. Finally, Germany’s commitment to AI is underscored by its 2018 Artificial Intelligence Strategy, backed by a €5 billion investment. AI adoption among German firms has grown significantly; before 2016, only about 2% of firms reported using AI, a figure that rose to 10% by 2021¹⁸.

To conduct our analysis, we use the measure of occupational exposure to AI developed by Webb (2019)¹⁹. We define this measure of occupational exposure based on workers’ initial occupations observed in the sample. To mitigate concerns that the rising importance of AI may have affected the self-selection of workers in their initial occupations, we restrict the sample to individuals who entered the labor market before 2010, i.e., before the advent of AI technology in Germany. Having classified occupations according to their degree of exposure to AI, we employ an event study design and a difference­in-differences (DiD) approach by comparing workers in high- and low-exposure occupations before and after the significant increase in the adoption of AI across German firms in 2010. Our identification strategy hinges on the assumption of parallel trends in the outcomes of interest between AI-exposed and non-exposed workers prior to 2010, and thus before the significant roll-out of AI in Germany. Our analysis supports this assumption by showing no evidence of any significant differences in our outcome variables in the pre­ trends, that is, during the period preceding the major wave of AI adoption (before 2010). We also conduct the same analysis using an alternative metric of AI exposure leveraging a new set of questions on AI-related technologies in the workplace, introduced for the first time in the 2020 SOEP wave.

The results of our baseline analysis suggest no evidence of a negative effect of AI on workers’ well-being. If anything, we find that AI is associated with significant improvements in self-rated health and health satisfaction, which may be consistent with the evidence of a decline in job physical intensity. We also find no evidence of significant effects of AI on the likelihood of reporting economic concerns nor on a number of metrics related to workers’ mental health. Overall, these results are consistent with prior studies documenting that AI exposure did not cause job losses^9,20,21. Finally, we examine how the impact of AI on workers’ well-being and health differs based on various factors such as gender, age, industry, geographic location (East vs. West Germany), skill level, and union membership.

The paper is organized as follows. In Section “Background”, we present the previous literature, the conceptual framework, and discuss our contribution and the background. In Section “Data and empirical specification”, we present the data and the empirical specification. In Section “Results”, we report our main results as well as a robustness test and heterogeneity analyses. Section “Conclusion” concludes the paper.

Background

Previous work

Our paper builds on and extends recent research examining the impact of AI on labor market outcomes and worker well-being. Prior studies have largely found positive effects of AI exposure on employment and wages at the industry or occupational level. For example, Felten et al. (2019)²² document small wage gains in AI-exposed occupations in the US, while Gathmann and Grimm (2022)²³ find a positive relationship between AI exposure and employment in Germany, particularly in the service sector. Acemoglu et al. (2022)²⁰ analyze U.S. labor markets using establishment-level vacancy data from 2010 onward, finding rapid AI adoption, particularly in firms with tasks suited for automation. Their findings suggest that AI-exposed firms increase AI-related hiring while simultaneously reducing non-AI hiring and altering skill demands. Despite these firm-level shifts, they find no significant impact on overall employment or wage growth in AI-exposed industries and occupations, indicating that AI’s displacement effects may currently outweigh productivity-driven job creation. When examining the heterogeneity of results across occupations, Bonfiglioli et al. (2025)²⁴ find evidence that AI exposure led to job losses across US commuting zones, particularly for low-skill and production workers, while benefiting high-wage and STEM occupations. Unlike other technologies, AI’s impact is driven by services rather than manufacturing, contributing to automation of jobs and rising inequality.

At the same time, researchers have raised concerns that AI could accelerate the erosion of middle-class job security by automating tasks without creating sufficient new roles for human workers²⁵.Whether AI complements or displaces human labor depends not only on the extent of automation but also on which tasks are automated and which workers are affected²⁶. In this regard, Brekelmans and Petropoulos (2020)²⁷ highlight that mid-skilled occupations are among the most vulnerable to AI-driven disruption. However, some scholars suggest that AI can help reduce job performance inequalities by improving efficiency and decision-making support^10,28.

While existing studies have extensively examined labor market effects of AI, relatively little attention has been paid to its broader impact on worker well-being. Our work addresses this gap, contributing to a growing body of research exploring the effects of automation technologies on workers’ health and psychological outcomes.

Previous research on the effects of automation on well-being has primarily examined industrial robots and mechanized systems^13,29. However, AI represents a distinct form of automation, as it relies on computer-based learning and cognitive processing rather than physical manipulation⁹. Unlike traditional robotics, which primarily displaces routine manual tasks, AI has the potential to automate non-routine cognitive tasks that were once considered resistant to mechanization^22,25,30. This shift suggests that AI may not only disrupt routine jobs but also reshape knowledge-based professions, placing even highly educated workers at risk of automation-driven changes. Moreover, as AI systems become more capable of complex reasoning, decision-making, and problem-solving, their impact on the labor market is likely to extend beyond simple task automation, influencing job design, skill requirements, and career trajectories across various industries.

The influence of AI on employee well-being may operate through multiple channels. On one hand, AI-driven automation can reduce physical strain in labor-intensive jobs, potentially improving physical health. On the other hand, AI adoption can increase cognitive and emotional demands in knowledge-intensive occupations, altering job content in ways that either enhance or undermine job satisfaction. Additionally, shifts in workplace dynamics—such as changes in perceived job security, workplace autonomy, and the sense of purpose derived from work—may further affect workers’ experiences with AI.

Worker attitudes toward AI remain mixed. Some global surveys indicate rising concerns about the consequences of AI on job opportunities³¹ yet a recent Pew study finds that US workers in AI-exposed industries do not perceive AI as an immediate threat³². AI has the potential to enhance productivity and complement human skills, but it can also displace workers in certain roles. As with past technological revolutions, the ultimate labor market impact of AI will depend on the evolving balance between complementarity and substitution between AI and human labor^20,33,34. Moreover, AI alters the nature of work itself, influencing job satisfaction, professional identity, and the perceived dignity of labor³⁵.

Whether the beneficial effects of AI on labor market outcomes offset or even outweigh its displacement effects remains an empirical question—particularly in the short term, as workers and labor markets undergo a period of transition and adaptation to this new general-purpose technology.

Conceptual framework

This study builds on task-based theories of technological change^33,36, which conceptualize AI as a transformative force that reallocates tasks between humans and machines. Similarly to robots, AI automates routine and physically demanding tasks, enabling workers to transition into higher-skill, cognitively demanding roles. However, unlike robots, AI can also automatize non-routine tasks³⁰. The dual role of AI—as both a complement and a substitute to human labor—is central to understanding its heterogeneous effects. In our context, complementarity arises when AI reduces physical strain and augments workers’ capabilities, enhancing productivity and job satisfaction. Conversely, substitutability occurs when AI displaces workers, increasing job insecurity and workplace anxiety.

AI adoption, as documented by Acemoglu et al. (2022)²⁰, follows a pattern where firms with task structures conducive to AI integration experience substantial labor reconfigurations. While automation’s displacement effects have been widely studied, the role of AI in hazardous and physically strenuous occupations offers a distinct perspective on its labor market consequences. We hypothesize that AI’s integration into hazardous tasks—such as those involving exposure to toxic environments, heavy lifting, or repetitive strain—can reduce workplace injuries and long-term health risks. This hypothesis aligns with prior research on automation’s potential to enhance worker safety by shifting high-risk activities to machines, thereby improving physical well-being¹³.

Acemoglu et al. (2022)²⁰ highlight that AI-exposed establishments tend to reduce overall hiring, indicating that productivity gains from AI do not necessarily lead to net job creation. This trend, coupled with AI’s ability to automate both routine and certain non-routine cognitive and abstract tasks, suggests that a broader range of workers—including those in knowledge-based professions—may experience job displacement pressures. Such uncertainty can have significant psychological consequences, as fears of automation-related job loss contribute to chronic stress, financial insecurity, and diminished workplace morale. The effects are likely to be unevenly distributed, disproportionately impacting low-wage workers while also altering career trajectories for higher-skilled professionals, ultimately reinforcing patterns of economic inequality across different skill levels.

Thus, while AI adoption holds promise for improving workplace safety and productivity, its net effect on worker well-being remains uncertain. Whether the benefits—such as reduced workplace hazards and improved job quality—outweigh the disruptions caused by job displacement and economic instability depends on how AI is integrated into labor markets. This duality underscores the need for policies that not only facilitate AI integration in ways that protect workers’ health but also address the economic and social ramifications of AI-driven labor shifts.

Furthermore, our study builds upon traditional job quality frameworks by examining AI-induced transformations in physical job intensity, cognitive demands, and health outcomes. While much of the public debate on digitalization has focused on job quantity (i.e., job creation vs. job loss), its effects on job quality are equally significant and warrant further attention^37,38. Martin and Hauret (2022)³⁷ identify six key dimensions of job quality commonly examined in the literature: labor income, workplace safety, working time and work-life balance, job security, skill development and training, and employment-related relationships and work motivation. Traditional models conceptualize job quality through physical, cognitive, and emotional dimensions. However, emerging technologies—especially AI—are reshaping these dimensions in profound ways. AI-driven workplace changes can lead to new forms of cognitive strain, shifts in workplace autonomy, and evolving skill demands, all of which have direct consequences for workers’ well-being. Further research is needed to fully understand how digitalization—particularly AI—modifies job demands, psychological stress, and long-term employment conditions.

AI technologies frequently automate repetitive, physically demanding, and hazardous tasks, thereby alleviating physical labor for workers. In manufacturing, AI-powered machinery has increasingly replaced tasks such as assembly-line work and heavy lifting, while in service industries, AI tools like virtual assistants reduce administrative workloads, complementing physical task reductions. As AI shifts job responsibilities from physical execution to supervisory or decision-making roles, the overall physical strain on workers declines. Empirical evidence supports this trend; for instance, Gihleb et al. (2022)¹³ document significant reductions in physically demanding work as automation becomes more prevalent. Using a physical burden metric from the SOEP, we examine the link between AI exposure and reduced physical job intensity.

Finally, our study engages with theories of institutional mediation¹³ to examine how labor market institutions shape AI’s effects on workers. Specifically, we hypothesize that Germany’s strong labor protections, high unionization rates, and employment legislation may moderate the adverse consequences of AI on worker well-being. Institutions that provide employment security, reskilling opportunities, and worker protections may buffer against the negative impacts of automation, reducing stress and job displacement fears. This contrasts with more flexible labor markets, where AI-induced disruptions may lead to greater economic precarity.

Our contribution

Our research highlights the importance of Germany’s unique institutional context, characterized by strong labor protections, extensive union representation, and comprehensive employment legislation. These factors, combined with Germany’s gradual adoption of AI technologies, create an environment where AI is more likely to complement rather than displace worker skills, mitigating some of the negative labor market effects observed in countries like the US. Germany’s institutional framework, marked by strong labor protections, widespread union representation, and comprehensive employment regulations, plays a crucial role in shaping the effects of AI adoption. These structural safeguards, along with the country’s more gradual integration of AI technologies, foster an environment where AI is more likely to enhance rather than replace worker skills, mitigating some of the negative labor market effects observed elsewhere. However, as Bonfiglioli et al. (2025)²⁴ highlight, AI exposure has led to job losses in the US, particularly among low-skilled and production workers, while benefiting high-wage and STEM occupations. Given AI’s distinct impact on service industries rather than traditional manufacturing, workers in routine-intensive service sectors may still face heightened risks of displacement, even within Germany’s more protective labor market. These theoretically ambiguous effects make Germany a particularly interesting case for examining the interaction between AI adoption and labor market institutions, raising an important empirical question about the extent to which these protections can shield workers from displacement while enabling technological progress.

In addition, we explore heterogeneity in outcomes based on worker characteristics (e.g., gender, education, union membership) and regional differences (e.g., East vs. West Germany), offering a more granular perspective on the potential effects of AI on labor markets and worker well-being.

Unlike studies such as Nazareno and Schiff (2021) and Liu (2023)^29,39, which rely on cross-sectional data and broad measures of automation exposure⁴⁰, our analysis is uniquely AI-focused. We leverage the Webb measure of exposure to AI alongside a self-reported metric from the SOEP, ensuring a more precise assessment of the effects of AI on worker well-being and labor market dynamic. Using longitudinal data from the SOEP over the period 2000–2020, we employ an event study analysis and a DiD design to address selection bias and control for individual fixed effects. This methodological approach enables us to capture long-term trends and examine nuanced outcomes such as life satisfaction, mental health, and physical health—dimensions that have often been overlooked in prior research.

AI in Germany

The roll-out of AI in Germany accelerated only recently. As noted by Gathmann and Grimm (2022)²³, patent applications for AI technologies started to grow strongly only after 2015, and more significantly in 2017 and 2018. The innovation survey conducted by ZEW- Leibniz Centre for European Economic Research provides a consistent longitudinal perspective on AI adoption in Germany¹⁸. Specifically, the most recent wave of the innovation survey contains information on AI adoption percentages for 2021. AI use was not widespread before 2010, and the rate of AI adoption was extremely low before 2016. To be conservative, we chose 2010 as the beginning of our treatment period. AI adoption rates have increased substantially over the last few years. While only 2% of firms adopted AI before 2016, this number rose to 6% in 2019 and 10% in 2021¹⁸. Regarding the diffusion of AI across industries, the leading adopters of AI technology in 2019 were finance (24%) and IT (21%), followed by skilled services (18%) such as legal, architecture, consulting, and research. Conversely, the laggards in AI adoption include mining (1.6%), miscellaneous business services (2.3%), and transportation (5.3%). These cross-sectoral differences in AI adoption are qualitatively reflected in our individual-level data on AI exposure from the SOEP, with IT and finance being the most exposed (see Figure A.1 in the Appendix). As deatiled in Rammer et al. (2022), 75% of the firms in the finance sector that used AI technologies in 2019 began using AI in 2016. The share of the chemical and pharmaceutical sectors is 74%, whereas that of electronics and machines is 68%.

Increasing rates of AI adoption across German firms were accompanied by the German government’s investment in AI. In 2018, the German Federal Government launched its Artificial Intelligence Strategy and pledged to invest approximately 5 billion euros by 2025 in AI development. For these reasons, Germany is an interesting country for analyzing the effects of rising exposure to AI on the well-being and health of workers.

Data and empirical specification

Data

Data and codes are available at https://osf.io/kagdj. The data source for our analysis comes from the German Socio-Economic Panel (SOEP) version 37. The SOEP is a representative longitudinal survey of households and individuals in Germany, covering a wide range of topics, including job characteristics, health, and well-being, since 1984^41,42. Its rich data structure provides several unique features that make it particularly well-suited for this study.

First, the SOEP includes detailed occupational information classified according to the International Standard Classification of Occupations (ISCO), which allows us to merge these data with the occupational measure of AI exposure developed by Webb (2019)¹⁹. Using natural language processing (NLP) techniques, Webb (2019) identifies verb-noun pairs in job descriptions from the O*NET database and compares them with similar pairs in AI-related patent descriptions. The degree of overlap between these pairs quantifies the susceptibility of specific tasks within occupations to being performed by AI technologies. By applying this measure to historical data on software and industrial robots, Webb (2019) demonstrates its validity, showing that occupations highly exposed to previous automation technologies experienced significant declines in employment and wages. To incorporate the Webb measure into our data, we used a cross-walk provided by the Bureau of Labor Statistics, which links O*NET Standard Occupational Classification codes to ISCO codes. This cross-walk matches to the four-digit ISCO-08 codes used in the SOEP dataset. Based on this measure, we construct our key explanatory variable: a dummy variable that indicates whether a worker is employed in an occupation that has a larger (i.e., above the median) fraction of such overlapping tasks. These workers with above median AI exposure are classified as more exposed to AI. This task-based approach provides an objective, granular, and internationally comparable measure of AI exposure, which is less prone to biases present in self-reported data.

Second, the 2020 SOEP wave provides information on the use of automatic digital systems and their frequency of use⁴³. In 2020, the survey included a new module aimed at measuring individual-level exposure to AI in the workplace. Employed respondents are asked a battery of questions about their current exposure to various digital systems and are required to indicate the frequency of interaction with these systems on the job. Figure A.2 in the Appendix shows the question module which covers five broad areas in digital systems: (1) natural language processing; (2) image and video processing; (3) text processing; (4) information processing and evaluation; and (5) knowledge gathering. Thus, as an alternative to the Webb measure of AI exposure, we construct an occupational metric based on the proportion of SOEP respondents within each occupation who report using AI tools in the workplace. One of the advantages of the SOEP questionnaire is that workers are interviewed indirectly about their use of AI technologies, thereby avoiding potential measurement errors due to their familiarity with the notion of AI. The use-frequency categories are “several times a day”; “on a daily basis”; “on a weekly basis”; “less often”; “never.” The answer distribution of these items is skewed; that is, most respondents reported that they never used these systems. The answer distributions are shown in Figures A.3 to A.7 in the Appendix. For these five items, positive exposure to AI (i.e., at least infrequent use) ranges from 20 to 30%. Using this information, we construct a broad measure of AI exposure at the individual level. In practice, we identify whether a person interacted with any type of digital system in their job, at least on a weekly basis. We have also created a more targeted measure that is specifically related to generative AI, focusing on tasks typically associated with its use. By concentrating on the use of AI for information processing and evaluation (4), as well as knowledge gathering (5), we construct a metric based on these specific use categories. This allows us to better capture the novel applications of AI, which have gained significant attention with the rise of large language models. When we conduct our analyses using this refined AI­measure, we find that our key findings remain robust. We then compute the average degree to which a three-digit ISCO occupation is directly exposed to AI technology. As in the case of the Webb measure, the SOEP-based measure of AI exposure is a dummy variable that indicates whether a worker is employed in an occupation with high (i.e., above the median) exposure to AI. Figure A.8 in the Appendix illustrates the predictive occupation categories for individual exposure to AI. Unsurprisingly, most AI exposed occupations include programmers and IT workers. White collar workers and skilled professionals are among those most exposed to new technologies.

Both measures have distinct strengths and limitations. The Webb (2019) measure offers an objective, task-based assessment of AI exposure that avoids the potential biases of self-reported data. It provides a systematic way to quantify the impact of AI across occupations, but it relies on US-based O*NET data and patent classifications, which may not fully reflect Germany’s specific labor market characteristics. On the other hand, the SOEP-based measure benefits from its grounding in workers’ self-reported interactions with AI technologies and its direct relevance to the German context. However, it is vulnerable to self-reporting biases, such as awareness and recall biases, which can affect the reliability of subjective metrics. Workers who perceive AI as beneficial may overstate positive effects, while those who feel threatened by AI may exaggerate negative impacts. These biases can influence the interpretation of results, particularly for our outcomes of subjective well-being. To mitigate these concerns, we use the Webb (2019) measure as our primary metric of AI exposure, given its objectivity and methodological rigor. At the same time, we incorporate the SOEP-based measure in robustness checks to capture complementary insights and discuss differences in results across the two measures. We also employ individual fixed effects and control for a wide range of time-varying individual characteristics, such as education, occupation, and industry, to address potential confounding factors. This dual approach allows us to provide a more comprehensive analysis while highlighting the importance of understanding measurement biases in studies of AI’s impact on worker well-being.

Finally, the dataset contains a set of self-reported indicators of satisfaction with respect to different life domains. In our study, we focus on life and job satisfaction as the primary outcome variables for workers’ well-being. These measures are recorded on an 11-point Likert scale, ranging from 0 (“very dissatisfied”) to 10 (“very satisfied”). Respondents are also asked about several domain-specific concerns, including worries about job security and personal economic situation, which serve as additional outcomes. They respond using a scale ranging from “not concerned at all” to “somewhat concerned” and “very concerned.” While single-item measures (e.g., job satisfaction, life satisfaction, self-rated health) lack the nuance of multi-item scales and may have greater measurement error, they are widely used in large-scale datasets like the SOEP due to their efficiency, reduced respondent burden, and demonstrated validity. Self-rated health, for example, has been shown to correlate strongly with objective health indicators and mortality riskand single-item life satisfaction measures, commonly used in happiness economics, have been validated against more detailed well-being assessments^44,45.

The SOEP provides several measures of individual overall health. For our empirical analysis, we use the standard self-assessed health metric, defined on a 5-point scale, where one indicates the poorest health and five represents the best. We also use the previously mentioned 11-category scale for health satisfaction. Furthermore, the SOEP has several metrics of mental well-being, including the Mental Component Score (MCS) derived from the SF-12 questionnaire, a validated composite measure of mental and emotional well-being derived from multiple items^44,46. In contrast, the anxiety measure is a single-item construct that relies on self-reported frequency of anxiety, assessed using a 5-point scale, with one representing “very rarely” and five indicating “very often”. To be specific, we begin with approximately 736,000 observations over the period 1984–2020. Restricting the sample to individuals who entered the labor market before 2010 and maintained their first-year occupation over time reduces it to approximately 446,000 observations. Limiting the period to 2000–2020 further narrows the sample to about 280,000 observations. Focusing on workers aged 18–59 at the time of the interview reduces the sample to approximately 213,000 observations. Incorporating the Webb measure using the Bureau of Labor Statistics cross-walk (linking O*NET SOC codes to ISCO-08) results in about 172,000 observations. After excluding cases with missing data on observable characteristics, the final sample consists of approximately 162,000 observations. Additionally, the SOEP includes information on individuals’ labor market outcomes. We use three main labor market indicators: employment, monthly gross labor income, and weekly working hours.

To minimize the risk of AI driving the selection into occupations, we restrict the analysis to individuals who entered the SOEP data with non-missing information on their occupation before 2010 and maintained their occupation fixed over time. The working sample is constructed as follows. We consider the survey years 2000–2020 and restrict our attention to workers aged 18–59 with nonmissing information on all covariates. As mentioned above, we restrict the sample to individuals who entered the labor market before 2010, namely, before the diffusion of AI across German firms. This restriction has consequences for the age distribution in our sample, particularly when focusing on the later years of our analysis and the impact of AI technology, which in Germany increased significantly only after 2015. Therefore, our results should be interpreted as the effects of AI exposure on middle-aged and older workers. After these restrictions, the final longitudinal sample consists of approximately 162,000 person-year observations with non-missing occupational information, derived from about 18,500 individuals. The sample size varies depending on the outcome variable used in the regression model. To be specific, we begin with approximately 736,000 observations over the period 1984–2020. Restricting the sample to individuals who entered the labor market before 2010 and maintained their first-year occupation over time reduces it to approximately 446,000 observations. Limiting the period to 2000–2020 further narrows the sample to about 280,000 observations. Focusing on workers aged 18–59 at the time of the interview reduces the sample to approximately 213,000 observations. Incorporating the Webb measure using the Bureau of Labor Statistics cross-walk (linking O*NET SOC codes to ISCO-08) results in about 172,000 observations. After excluding cases with missing data on observable characteristics, the final sample consists of approximately 162,000 observations. As a robustness check, we repeated the main analysis using the Webb measure with data from SOEP version 39. However, since SOEP versions 38 and 39 lack information on workplace exposure to digital systems, we decided to retain version 37 as our baseline for consistency with the analysis based on the self-reported measure of AI exposure from the SOEP.

Table A.1 in the Appendix displays the descriptive statistics of the main variables used in our analyses. As regards AI exposure, approximately 51% of workers in the sample are classified as having a high exposure (i.e., above the median) to AI according to the Webb measure. The corresponding share for the SOEP measure is 66%, indicating that about two-thirds of workers are employed in occupations where workers reported a high use of AI technology in the workplace. Given the differences in how the Webb and SOEP measures define AI exposure-especially since the Webb measure assesses AI exposure at the task level-it is unsurprising to see significant variations in the proportions of workers classified as highly exposed to AI. The sample can be characterized as mid-career workers, as their average age is approximately 41. The sample is balanced between males and females. Approximately 24% of individuals have a college degree, which identifies the high-skilled workers in our heterogeneity analysis by education. Approximately 60% are married, and the average number of children in the household is close to 0.7. Although, as mentioned above, the sample is composed of slightly older workers, they are fairly representative of the German workforce. Average satisfaction with life and with job are both equal to approximately 7. The mean of the worries scales (job security and personal economic situation) is close to two, and thus centered on the response item indicating that they are “somewhat concerned”. The MCS is close to its normed mean of 50, while the mean of the anxiety frequency is close to 2, which corresponds to the item “rarely feeling anxious.” Labor market statistics reflect a working age sample, with about 84% of respondents currently employed, either part-time or full­ time. On average, workers report a monthly labor income of approximately 3,000 euros for a given year and work nearly 39 h per week.

Model specification

To examine the relationship between AI exposure and workers’ outcomes, we employ two complementary empirical approaches. First, we adopt an event study approach. Therefore, we estimate the following equation:

1

where the index ijst denotes an individual i, who had their first job in an ISCO occupation j, resided in federal state s, and was interviewed in year t. denotes the outcome variable of interest: well-being (life satisfaction and job satisfaction); worries (concerns about job insecurity and personal economic situation); health metrics (health status, health satisfaction, mental component score (MCS), and anxiety frequency); and labor market outcomes (employment, labor income, and hours worked). is a set of calendar year dummies from 2000 to 2020 with the reference period being 2010. Before 2010, there was almost no exposure to AI technology in the workplace. As mentioned earlier, fewer than 1% of businesses were exposed to AI at the time. is a dummy variable equal to one if individual i is highly exposed to AI in their (first) occupation, i.e., an indicator variable that equals one if the exposure to AI is above the median in this occupation. The coefficients of interest are , which capture the average difference in the outcomes of interest between AI-exposed and non-exposed workers over time. is a vector that includes worker-level covariates such as interactions between a gender dummy and a full set of age dummies, the number of children, as well as indicators for marital status and education. denotes the individual fixed effects, which absorb the influence of any time-constant individual heterogeneity. Our specification also includes federal state × year fixed effects (the initial state of residence in the panel is used to eliminate the impact of migration between states, which may be influenced by exposure to AI), . The inclusion of these effects controls for all possible state-level time-varying factors, thereby accounting for the possibility that regions with different occupational structures may experience different time-varying shocks. A potential concern is that technological progress from 2000 to 2020 could be correlated with the measure of exposure to AI. To mitigate potential biases stemming from technological progress, we add to Eq. (1) two sets of fixed effects. First, we include one-digit occupation fixed effects, , thereby exploiting only variation in AI-exposure among individuals working in the same one-digit occupation. These fixed effects absorb the influence of any changes occurring in technologies across one-digit occupations during the period. This set of fixed effects controls for the impact of technology on workers employed in occupations with different task content (e.g., how routine the tasks are and hence how susceptible they are to automation/computerization). Second, all our estimates control for one-digit industry × year fixed effects, , which account for industry-specific shocks over time. For both sets of fixed effects, we use the initial occupation (or industry) observed in the individual’s first year in the sample to control for potential bias from selection into the occupations (or industries) that may be influenced by AI exposure. Finally, represents an idiosyncratic error term.

We then integrate the event study analysis with the results from the DiD design. Formally, we estimate the following model:

2

where the variables , , , parameters , and the error term are defined in the same way as in Eq. (1). is a dummy variable that equals one after 2010. The key coefficient in the DiD specification is , which captures the difference in outcomes for AI-exposed workers after 2010, relative to non-exposed workers. The identifying variation for our coefficients of interest in both equation — and , respectively—stems from changes in AI exposure within occupations and over time. We cluster standard errors at the two-digit occupation level for all estimates.

At this stage, it is worth remarking that our analysis faces two main empirical challenges. First, it does not leverage any quasi-experimental variation in the allocation of workers across AI-exposed and non-exposed occupations. Second, we do not observe the counterfactual evolution of our outcome variables in the absence of AI. We attempt to address these issues in three ways. First, in both equations, we exploit the longitudinal design of the SOEP by including worker fixed effects (). These fixed effects cancel out the important time-invariant confounding factors that could bias our estimates. For example, individuals might sort themselves into occupations with different levels of AI exposure based on pre-determined characteristics, which could simultaneously affect their well-being, concerns about their economic future, and health. Individual fixed effects account for this selection bias. Second, our choice to assign exposure to AI of the initial occupation and to keep only individuals entering the sample before 2010,—before the advent of AI technology in the German industry,—further alleviates selection concerns regarding the movement of workers across occupations in response to AI penetration. Finally, we show that there are flat pre-trends between AI-exposed and non­exposed workers, thereby suggesting that in our setting, the identification assumption of parallel trends in the absence of AI is plausible (see Section “Results”).

Results

Main results

Event studies

Using the Webb measure of AI, Figs. 1 and 2 present the event study estimates of the effect of AI on well-being and health, as described in Eq. (1), namely, the series of estimated coefficients. These figures highlight the dynamic impact of AI exposure during the two periods. The first period, 2000–2010, compares AI­exposed and non-exposed workers when AI was virtually absent. The second period, 2011–2020, analyzes the differences between exposed and non-exposed workers during the initial stages of AI adoption in Germany, and in a period in which AI usage increased substantially among German firms. For all outcomes in Fig. 1, we observe a flat pre-trend from 2000 to 2010, with the coefficients statistically insignificant and close to zero. There is no evidence of a significant change in our measures of workers’ well­being and economic concerns after 2010. Figure 2 displays the event study estimates for the health outcomes. For each outcome, we observe non-significant differences between AI-exposed and non-exposed workers between 2000 and 2010. For anxiety, we have no data before 2007. After 2010, the event study coefficients for mental health and anxiety are small and close to zero. However, we do find some evidence of an improvement in both health status and satisfaction with health.

Fig. 1.

Exposure to AI and well-being, 2000–2020—event study analysis. Data are drawn from the SOEP version 37. The figure shows the point estimates and 95% confidence intervals of the interaction terms between the Webb measure of AI exposure and year dummies taking 2010 as the reference year when estimating the model in Eq. (1).

Fig. 2.

Exposure to AI and health outcomes, 2000–2020—event study analysis. Data are drawn from the SOEP version 37. The figure shows the point estimates and 95% confidence intervals of the interaction terms between the Webb measure of AI exposure and year dummies taking 2010 as the reference year when estimating the model in Eq. (1).

In summary, our event study analysis of workers’ well-being and health reveals no significant change in the trend between AI-exposed and non-exposed workers after 2010. Nonetheless, there is some evidence of an improvement in health status and satisfaction with health.

Difference-in-differences

To gauge the overall effect on our metrics of interest of the Webb measure of exposure to AI, we compare AI-exposed and non-exposed workers before and after the marked increase in AI adoption across German firms. There was virtually no AI roll-out in Germany before 2010¹⁸. Therefore, we conduct a DiD analysis comparing AI-exposed and non-exposed workers before and after 2010. Table 1 reports the main estimates of the effect of AI on workers’ well-being and concerns about their economic situations. The results show that from 2011 onward, AI-exposed workers do not report any significant difference in life satisfaction, job satisfaction and personal economic situation than non-exposed workers (see columns 1 to 4, respectively). Consistent with Fig. 1, we do not observe significant differences in the pre-trends between AI-exposed and non-exposed workers (between 2000 and 2010). Indeed, when testing the hypothesis that the sum of all the pre-trend coefficients for the years preceding the major wave of AI adoption is not significantly different from zero, we fail to reject the null hypothesis for all outcomes (see columns 1 to 4 of Table 1).

Table `1.

Effects of Exposure to AI on Workers’ Well-being – DiD Estimates

Dep. var.	(1)	(2)	(3)	(4)
	Life satisfaction	Job satisfaction	Worries: job security	Worries: own economic situation
Exposed to AI×2011–2020	0.005 (0.019)	− 0.002 (0.039)	− 0.004 (0.016)	− 0.012 (0.008)
Mean of dep. var.	7.045	6.966	1.663	1.961
Std. dev. of dep. var.	1.706	2.067	0.701	0.690
F-test p-value	0.969	0.195	0.327	0.290
(2000+…+2009 = 0)
Observations	159,319	136,121	131,948	159,046

Data are drawn from the SOEP version 37. Life and job satisfaction are measured on an 11-point Likert scale, ranging from 0 (“very dissatisfied”) to 10 (“very satisfied”). Concerns for job security and own economic situation are recorded on a five-point scale, ranging from “not concerned at all” to “very concerned”. Standard errors are reported in parentheses and are clustered at the occupation level. All specifications include individual, age × gender, year × federal state, year × industry, and one-digit occupation fixed effects. Further controls include indicators for education, marital status and number of children.

*Significant at 10%; ** Significant at 5%; ***Significant at 1%.

Table 2 shows the regression coefficients of exposure to AI on health outcomes. Taken together, the results in the table confirm the visual evidence from the event study (see Fig. 2). We find that if anything, exposure to AI leads to an improvement in self­reported health status and satisfaction with health (see columns 1 and 2), whereas there is no evidence of a negative effect on mental health and anxiety (see columns 3 and 4). Reassuringly, for these outcomes as well, we cannot reject the null hypothesis that the sum of all pre-trend coefficients for the years preceding the substantial increase in AI adoption is equal to zero. We acknowledge the potential benefits of using a more nuanced, continuous measure of AI exposure. However, given the primary empirical approach of this study, which relies on an event study and DiD framework, we opted for a dichotomous operationalization of AI exposure. This approach allows us to focus on workers more likely to be exposed to AI. Notably, when using a continuous measure, we continue to find no significant effects on workers’ well-being and mental health. Furthermore, while the baseline results indicate some positive effects on self-reported health and health satisfaction, these effects diminish to zero with the continuous measure. These findings suggest that the dichotomous metric captures the primary effects among workers with higher exposure to AI. We also examined physical health (PCS) as an alternative outcome variable. The results indicate a positive correlation between AI exposure and physical health, with an estimated increase of 0.22 (p-value = 0.126). While the coefficient is not statistically significant, its direction is consistent with our finding that AI adoption led to a mild improvement in self-reported health status.

Table 2.

Effects of exposure to AI on workers’ health outcomes – DiD estimates.

Dep. var.:	(1)	(2)	(3)	(4)
	Health status	Health satisfaction	Mental health	Anxiety
Exposed to AI×2011–2020	0.022** (0.009)	0.042* (0.022)	0.192 (0.193)	− 0.016 (0.017)
Mean of dep. var.	3.556	6.941	49.62	1.948
Std. dev. of dep. var.	0.875	2.042	9.644	0.963
F-test p-value	0.549	0.925	0.577	0.509
(2000+…+2009 = 0)
Observations	159,465	159,408	68,366	81,895

Data are drawn from the SOEP version 37. Self-assessed health status is measured on a 5-point scale, ranging from 1 (“very poor”) to 5 (“very good”). Health satisfaction is recorded on an 11-point Likert scale, ranging from 0 (“very dissatisfied”) to 10 (“very satisfied”). Mental health is assessed using the Mental Component Score (MCS) derived from the SF-12 questionnaire, a validated composite measure of mental and emotional well-being derived from multiple items. The anxiety measure is a single-item construct that relies based on self-reported frequency of anxiety, assessed using a 5-point scale, with one representing “very rarely” and five indicating “very often”. Standard errors are reported in parentheses and are clustered at the occupation level. All specifications include individual, age × gender, year × federal state, year × industry, and one-digit occupation fixed effects. Further controls include indicators for education, marital status and number of children.

*Significant at 10%; ** Significant at 5%; ***Significant at 1%.

Overall, our results suggest that our occupational measure of AI exposure had no significant effects on life and job satisfaction, as well as concerns about personal economic futures. Moreover, occupational AI exposure had no negative effect on mental health, and if anything, there is a mild positive effect on self-reported health status and satisfaction with health. The lack of negative effects on mental health aligns with previous research that found no negative effects of AI on labor market outcomes^20–22. Using our data, Table 3 confirms the absence of any adverse effects on wages and employment. However, there is evidence of a modest reduction in working hours, corresponding to approximately 30 min per week on average. At the same time, as shown in Table A.2 in the Appendix, the evidence of mild positive effects on self-rated health and satisfaction with health may be consistent with the significant reduction in the share of workers employed in jobs with a high physical burden (defined as a score above eight for both physical and psychological burdens, i.e., above the 75th percentile). In this analysis, we used occupational data on physical and psychological burdens. For the years up to 2011, we employed the Kroll (2011)⁴⁷ Job-Exposure Matrix, while for the years following 2011, we used the updated 2018 Job-Exposure Matrix⁴⁸.

Table 3.

Effects of exposure to AI on workers’ labor market Outcomes – DiD estimates.

Dep. var.:	(1)	(2)	(3)
	Employment	Labor income	Working hours
Exposed to AI×2011–2020	− 0.006 (0.009)	47.241 (36.719)	− 0.591** (0.257)
Mean of dep. var.	0.839	2967	38.60
Std. dev. of dep. var.	0.368	2543	12.80
F-test p-value	0.918	0.294	0.628
(2000+…+2009 = 0)
Observations	159,707	133,612	130,410

Data are drawn from the SOEP version 37. Standard errors are reported in parentheses and are clustered at the occupation level. The labor income outcome in column (2) corresponds to monthly gross labor income, whereas the working hours in column (3) refers to weekly working hours. All specifications include individual, age × gender, year × federal state, year × industry, and one-digit occupation fixed effects. Further controls include indicators for education, marital status and number of children.

*Significant at 10%; ** Significant at 5%; ***Significant at 1%.

Robustness checks

To address potential concerns about the use of occupation-level AI exposure measures, we note that our primary metric, based on Webb’s framework, offers a robust and established approach to assessing the susceptibility of tasks to AI within an occupation. By relying on pre-determined occupational characteristics, this measure mitigates endogeneity concerns and provides a consistent basis for comparison with existing literature. However, it assumes uniform exposure within occupations, which may not fully capture individual-level heterogeneity or the dynamic nature of AI adoption. A further limitation of this approach is that the same occupation can involve different tasks and require varying skills across countries⁴⁹. To complement this approach, we repeat the main analysis by incorporating an individual-level measure of AI exposure based on self-reported data from the SOEP about AI-related technologies in the workplace. Thus, the SOEP metric represents the average level of self-reported AI exposure among workers in a given occupation. While this metric is more prone to self-reported bias, it captures direct interactions of workers with AI and allows for a more granular exploration of heterogeneity. Using the SOEP-based measure of AI, we find evidence of small negative effects on life satisfaction and job satisfaction (approximately 0.05 standard deviations). At the same time and consistent with what found with the Webb measure, we confirm the positive effects on self-reported health status and health satisfaction (see Panel A of Tables A.3 and A.4 in the Appendix). Similar results are obtained when constructing a more focused measure specifically related to generative AI, concentrating on tasks commonly associated with its use (see Panel B of Tables A.3 and A.4). As detailed in Sect. 2, in Panel B of Tables A.3 and A.4 we create a more refined measure that is specifically related to generative Al, focusing on the following two use-categories: information processing and evaluation (4), as well as knowledge gathering (5).

Second, we consider alternative metrics for the outcomes of interest. Specifically, for the satisfaction variables, we replace the continuous scales with dichotomous variables equal to one if the respondent indicates a level of satisfaction at or above the median. For concerns and anxiety, we use dummy variables equal to one if the respondent indicated being very concerned. For mental health, we construct a binary variable that equals one if the respondent reports a score at or above the median. As regards self-rated health, we dichotomize it as poor health. The results presented in Tables A.5 and A.6 in the Appendix tend in the same direction and confirm the lack of any significant effects on workers’ well-being and mental health outcomes, while indicating a beneficial effect on physical health. Third, we check the sensitivity of our results to the exclusion of 2020 from the sample. In this case, the major concern is that COVID-19 may have affected well-being and health differently for workers in different occupations. Reassuringly, the DiD estimates reported in Table A.7 and A.8 in the Appendix show that the effects of AI exposure are very similar to the benchmark specification (see Tables 1 and 2). Fourth, expanding the analysis to include the years 2021 and 2022 yields similar results (see Tables A.9 and A.10 in the Appendix). The SOEP did not directly collect individual-level data on exposure to digital systems in the workplace for the years 2021 and 2022. Finally, the inclusion of number of children, marital status, and education dummies as control variables might be problematic as they could potentially be affected by workers’ well­being and health. Tables A.11 and A.12 in the Appendix show that the main results hold true regardless of whether covariates are controlled for or not.

Heterogeneity analyses

In the Appendix, we then present the heterogeneity analyses for workers’ well-being and health along many dimensions: gender, education, unionization, industry, age, and East vs. West Germany (see Tables A.13–A.24). Notably, as shown in Table A.14, the positive impacts of AI exposure on health status and health satisfaction appear to be driven primarily by medium-low educated respondents, which aligns with our findings suggesting that the reduction in physical job intensity may be one of the underlying mechanisms behind the effect of AI on self-reported health status and health satisfaction.

When examining regional heterogeneity, we find that West German respondents report positive effects on health status and satisfaction with health, and exhibit a reduction in anxiety, whereas Eastern Germans experience heightened anxiety associated with AI exposure (see Table A.16). This contrast likely reflects the distinct economic and social contexts of the two regions. Eastern Germany’s higher unemployment, slower economic growth, and greater prevalence of routine, lower-skilled jobs may exacerbate fears of technological displacement. In contrast, Western Germany benefits from stronger labor market stability, higher wages, and better access to retraining programs, which could help mitigate such concerns^50,51. This divergence in AI-related anxiety may also have deeper historical roots. Centralized economic planning in the former German Democratic Republic (GDR) emphasized state-driven technological development at the expense of market adaptability, leaving many industries unprepared for the rapid economic transformations following reunification⁵². Decades of structural economic challenges, combined with historical perceptions of technological lag, may have contributed to a persistent sense of vulnerability toward automation. These findings underscore the importance of addressing regional inequalities through targeted interventions. Expanding retraining programs, enhancing digital literacy initiatives, and increasing investment in high-skill, technology-intensive industries could help reduce dependence on jobs that have higher risk of being replaced by AI and improve economic resilience in economically disadvantaged areas. Such efforts are crucial to ensuring that AI adoption supports, rather than exacerbates, existing disparities in workers’ well-being. We do not find any additional evidence of significant heterogeneity across other groups (see Tables A.17–A.24).

Conclusion

Are workers concerned about the consequences of AI on their labor market opportunities? Recent advances in AI have led to fundamental shifts in daily life. A handful of studies have examined the impact of AI on labor markets and workplace productivity^2,8,9,20. However, little is known about how the AI revolution has affected workers’ well-being and health. Using longitudinal survey data from Germany, this study estimated the effects of exposure to AI technology in the workplace on workers’ well-being and health.

Comparing workers highly exposed to AI with workers employed in less exposed jobs before and after the notable rise in the adoption of AI across German firms in 2010, we found no evidence of significant effects of AI technology on well-being and concerns about the future. While our findings provide no evidence of significant negative effects on mental health, they suggest an improvement in self-reported health status and health satisfaction. This result appears consistent with recent studies that have found no evidence of adverse effects of AI on labor market outcomes, as well as with our own evidence of a decline in job physical intensity.

Our study has a few limitations. Although AI adoption has markedly increased in recent years, we are still in the early phases of the AI revolution. Therefore, it may be premature to draw definitive conclusions about the impact of AI on workers. Furthermore, our main analysis focuses on occupational-level measure of AI exposure, which assumes uniform exposure within occupations and may overlook individual-level variation in AI interactions. While our robustness checks using a self-reported measure of AI exposure suggest small but significant negative effects on well-being, these impacts may be masked at the aggregate level. Additionally, our study cannot generalize to young workers, as we restrict the sample to those who entered the labor market before 2010. Future research could provide deeper insights by analyzing larger datasets and employing more granular AI exposure measures, particularly those distinguishing between varying degrees of AI interaction. Another limitation stems from our use of single-item measures for well-being and health, which, while validated in prior research, may lack the depth of multi-item scales. Furthermore, the self-reported SOEP-based measure of AI exposure introduces potential recall and perception biases, as workers’ subjective views on AI may influence their responses. While we primarily rely on the Webb (2019) measure to mitigate these biases, interpretation should be cautious. Moreover, our results may be influenced by COVID-19-related confounders, particularly in the 2020 wave of SOEP data. Finally, our findings are specific to Germany, where strong labor protections, unionization, and a gradual approach to AI adoption may buffer some of adverse effects of AI on workers. These institutional safeguards differ from those in other countries, particularly those with more flexible labor markets, implying that our results may not fully generalize to different regulatory environments.

While our study is exploratory, it offers preliminary insights into the short-term effects of the AI revolution on workers’ perceptions during this period of technological transition. Future research should further explore the heterogeneous impacts of AI across demographic and regional groups, which is crucial for shaping labor market policies that balance innovation with worker protection.

AI-driven technologies can automate a broader range of tasks than robotics-focused ones⁹, exposing more middle- and high-skilled jobs to automation. The balance between augmentation and substitution remains uncertain, thus requiring further study. Our research explores workers’ perceptions during this transition, highlighting AI’s varied impact on well-being. Understanding these effects can guide labor policies that foster innovation while protecting workers. Policies that safeguard vulnerable employees, develop effective retraining initiatives, and assist workers through technological shifts could be instrumental in reducing the negative impacts of AI on worker well-being while maximizing its benefits.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

J.K. and L.S. conducted the data analysis. O.G., J.K. and L.S. wrote the main manuscript. All authors reviewed the manuscript.

Data availability

Data and codes are available at https://osf.io/kagdj.

Declarations

Competing interests

The authors declare no competing interests.