Technological change and returns to training Open Access

The labor market and employees are confronted with technological change that has far-reaching effects on tasks, job requirements and occupations (e.g. Spitz-Oener, 2006; Dustmann et al., 2009; Autor, 2015). In recent years, the rapid diffusion of artificial intelligence (AI) technologies, most notably large language models (LLMs), has introduced a new wave of disruption with potentially profound effects on the labor market (e.g. Noy and Zhang, 2023; Autor, 2024; Brynjolfsson et al., 2025). Training is often seen as a measure to adapt to new technologies and as a safeguard against possible negative consequences of technological change (e.g. Schmidpeter and Winter-Ebmer, 2021). Against this background, we investigate returns to training and whether returns differ when training and technological change interact. We examine this interaction by analyzing the association between training and wages as well as job mobility for individuals who adopt a new technology at the workplace within the same year and for those without technology adoption.

Analyzing whether returns to training are different for training that is accompanied by new technology from returns to other types of training is important, given that individuals' willingness to bear the costs of training (either the monetary costs or the cost of time and effort) crucially depends on the size of the return. An increase in training participation to adapt to technological change can only be realized if training returns to individuals are positive and not lower than those to other types of training. Otherwise, firms or governments would have to step in and take over even larger parts of the monetary costs, assuming that training has a benefit for firms or societies overall.

There is a large literature documenting the labor market returns to training in terms of wages (see, e.g. Lynch, 1992; Frazis and Loewenstein, 2005; Leuven and Oosterbeek, 2008) or aspects of job mobility (e.g. Parent, 1999; Görlitz and Tamm, 2016), as well as literature documenting the heterogeneity of returns to training based on socio-demographic characteristics (e.g. Ruhose et al., 2019) or type of training ^[1]. We contribute to this strand of literature by analyzing heterogeneity with respect to the parallel adoption of new technology.

In addition, we contribute to the literature on the interplay of technology and training. Most of the literature that has examined the link between technology and training has concentrated on the provision of training or on employee's participation. Lukowski et al. (2021) found that firms with a higher proportion of digital technology users provide more training and once individuals are familiar with digital technologies, they participate less frequently in training. Furthermore, Kleinert and Wölfel (2018) as well as Heβ et al. (2023) show that individuals who are most threatened by technological progress (by having a large share of routine tasks at work or by being highly exposed to robot technology) are less likely to participate in training. Studies that look at returns to training that is accompanied by technology have mostly looked at productivity outcomes instead of individual wage or mobility so far. Boothby et al. (2010) estimate the relationship between technology-training combinations and productivity performance. They find that firms that adopt new technologies and at the same time invest in skills (via training) report higher productivity. Similar results are presented in Bresnahan et al. (2002). They find a positive association between information technology (IT), complementary workplace reorganization (training is part of this) and product and service innovation. Finally, Bartel et al. (2007) find that investments in new computer-based IT improve the efficiency of the production process at all stages and promote the adoption of new human resource practices that include training. Instead of productivity outcomes, we are the first who provide insights on how individual wages and job mobility are related to training, which is conducted alongside the adoption of new technologies at the workplace.

Another contribution of the paper to the literature is to analyze how technological change is associated to worker mobility. Bauer and Bender (2004) find that firms that introduce new organizational and technological change experience lower employment growth rates. Technological change is assessed via the firms' investment in information and communication technology. Dauth et al. (2021) look at robot adoption at the workplace and find more stable employment within firms for incumbents. Robot adoption is measured at the occupation level. More recently, attention has turned to the labor market implications of AI adoption. Much of the emerging literature focuses on the potential wage (e.g. Acemoglu et al., 2022) or productivity effects (e.g. Peng et al., 2023) of AI. Gathmann et al. (2024) look at the impact on worker mobility by investigating displacement effects from AI exposure and find only modest overall impacts. Their evidence suggests that affected workers tend to respond by transitioning into occupations or industries that are less exposed to AI. Contrary to the previous papers our measure of technology has the advantage of being at the individual level and thereby providing more accurate estimates for the individual worker.

Our analysis is based on data from the German Socio-economic Panel (SOEP), which includes many questions on labor market aspects as well as personal characteristics. In comparison to previous studies using the SOEP (e.g. Pischke, 2001; Ruhose et al., 2019), which had to rely on retrospective information on training participation for 3-year periods, which is plagued by considerable recall bias, we exploit recent waves with new questions regarding training participation within the last calendar year. We also use a survey module regarding new technological innovations at the workplace. As outcomes, we analyze wages and two dimensions of job mobility in the form of job changes and promotions.

The analysis is structured in two main analytical sections: First, we descriptively investigate the determinants of training participation and its association with the adoption of new technologies in Section 2.2. Secondly, in Section 4, we examine returns to training, to new technology adoption and the combination of training and technology. We look at returns in terms of wages and two indicators for job mobility. For the main analysis, we estimate fixed effects models that take the longitudinal nature of our data into account. Given that individuals might also differ in time-variant unobserved characteristics, we additionally estimate fixed effects models that control for individual-specific time trends.

Our results indicate that adopting a new technology at the workplace is on average associated with a higher likelihood of participating in training. This association holds even when controlling for an extensive set of observables as well as time-invariant unobservable characteristics. Furthermore, we find that training participation is associated with wages that are on average 0.9% higher. In contrast, new technologies show no statistically discernible association with wages, once individual fixed effects are taken into account. The interaction of training and new technology has a negative sign and is not significant, but findings indicate that new technology adopters do not benefit in terms of wage increases when participating in training. Robustness tests that additionally include individual time trends find similar results, though with larger standard errors. With regard to the mobility outcomes, we can see that training participation is related to a reduction in the probability of job changes. We interpret this decrease in mobility as an increase of the job stability in the current position. This association is only present for training that is not accompanied by technology adoption. In contrast, if training is accompanied by technology adoption, no statistically meaningful relation with job changes is found. We also examine heterogeneity by education level. The results indicate that individuals with lower levels of education do not show wage differences associated with training or new technology adoption, whereas highly educated individuals exhibit a statistically significant positive association between technology adoption and wages.

The remainder of the paper is organized as follows: Section 2 describes the data, presents descriptive statistics and analyzes the association between training participation and new technology adoption. Section 3 discusses the estimation strategy for estimating returns. Results on returns are presented in Section 4. The final section summarizes the findings and offers a conclusion.

The following section describes the data that is used in the analysis and provides the first descriptive evidence on the relation between training participation and new technologies at the workplace.

The analysis is based on data from SOEP v37, a representative longitudinal survey of private households in Germany. It covers a wide range of topics, including household characteristics, income, employment, and education. From 2014 onward, questions regarding training activities performed during the last calendar year are included. Within the period 2015–2018, a question regarding technological innovations at the workplace was included in the annual interviews as well. With respect to outcome variables, we look at log hourly wages and two indicators for job mobility. Specifically, we define indicators for a job change (taking place either by a change of employer or within the firm) and for a promotion within the firm (we count any job change within the firm as a promotion ^[2]).

The analysis is restricted to the four panel waves (2015–2018) with information on both explanatory variables of interest. The estimation sample consists of working-age individuals from age 21–60 who have some kind of employment (full time, part time, marginal, self-employed or apprenticeship). Armed forces and sheltered workshops are excluded. Any observations with item non-response in our training or technology variables are also dropped from the analysis. The resulting sample covers 44,791 person-year observations from 17,856 individuals of which 7,735 are observed in all four waves.

In SOEP, the question on technology adoption is as follows: “Sometimes there are changes in the tools and technologies of the workplace — for example, when new technologies, devices, or working or production processes are introduced. What about you? Have there been any changes of this kind in your job in (the last calendar year)?” In the analysis sample, 22% of employees report that they did adopt a new technology ^[3].

Our measure for training participation refers to any training that was completed in the last calendar year. It encompasses “all types of vocational measures that are designed to build on previous professional training or to pave the way for a change of profession.” The initiator of the training could be the individual, the employer or a government agency. In case a measure was completed, follow-up questions inquire the total number of completed training programs, the number of days spent in training as well as who paid for the training. The overall training participation rate is 30% and relatively stable across all four panel waves. Other studies looking at training within the SOEP find similar participation rates (e.g. Caliendo et al., 2022 or Caliendo et al., 2023). Among all participants, the average duration of training is 9.8 days with a median of 4 days. For training participants with a parallel technology adoption, this duration is slightly higher at 10.3 days with a median of 5 days. In most cases, the training is paid by the employer (85%), which is typical for the German labor market and many other economies (e.g. Pischke, 2001; Bassanini et al., 2007). This aligns with Acemoglu and Pischke (1998, 1999), who demonstrate that employer financing can arise under monopsony power, where firms recoup costs via wage compression and reduced turnover. Among the new technology adopters, the share of (at least partly) employer-financed training is slightly higher at 87% than among nonadopters (84%) and the share of training participants who (at least partly) self-finance training is somewhat lower among new technology adopters than among nonadopters (14 vs 15%).

Descriptive statistics on the characteristics of training participants and nonparticipants are presented in Table 1. There are hardly any differences in the proportion of women between the groups. However, individuals with a migration background are less likely to participate in training, which was also found in previous studies (e.g. Beicht and Walden, 2017). In line with the literature (e.g. Kramer and Tamm, 2018), we find large educational differences, with the low educated being less likely to participate in training, while the highly educated are over-represented in the population of training participants. In terms of job characteristics, it can be seen that training participants are more likely to be in full-time employment. They are also slightly more likely to be self-employed and less likely to be in an apprenticeship. Additionally, training participants are also more frequently employed in large firms (more than 200 workers) than nonparticipants. With regard to technological change, we can see that, on average, 31% of training participants indicate the adoption of a new technology at the workplace, while for nonparticipants this proportion is much lower at around 18%. Our outcome variables also show differences. Training participants earn on average around 4.3 euros more per hour than nonparticipants. They are less likely to change jobs but more likely to be promoted within the same company.

In order to learn more about the correlation between training participation and the adoption of new technologies, Table 2 presents results that account for individual- and job-related characteristics that may affect selection into training. The specification in column (1) only includes the new technology dummy as well as a set of year dummies as controls. Column (2) adds individual controls and job-related characteristics. Column (3) additionally controls for individual fixed effects. The effect of the new technology dummy is positive and highly significant across all specifications. The preferred fixed-effects specification in column (3) indicates that having a new technology at the workplace is associated with a 3.5% points higher probability of participating in a training measure, even when controlling for the whole set of individual and job-related characteristics as well as unobservable time-invariant factors. Overall, individuals who have experienced some technological innovation at their workplace seem to be more likely to participate in training. One reason for this could be that the training is necessary to acquire new skills and knowledge that are required to operate the new technology.

To learn more about the longitudinal pattern of training participation and of new technology adoption,  Appendix Table A1 shows the proportion of individuals with 0, 1, 2, 3 or 4 years of training (of new technology adoption) in the 4-year period of observation. These shares refer to individuals who are part of the analysis sample in all four waves. Almost half of the respondents never experience any training or new technology within the four-year period. Among training participants, a participation in only one year is more common than frequent participations. Similarly, among new technology adopters, a new technological change at the workplace during one year is the most common incidence. Only about a quarter of respondents report more than one technological innovation. In comparison, participating in more than one training measure from 2015 to 2018 is more common at around 37%. A joint occurrence of training and new technology in one year is reported by 15% and in more than one year by hardly 10% of the individuals.

The literature on returns to training has been using several strategies to estimate causal effects, among others fixed-effects methods (e.g. Pischke, 2001), matching (e.g. Ruhose et al., 2019), instrumental variables strategies (e.g. Brunello et al., 2012) and settings where training participation is determined by plausibly exogenous factors (e.g. Leuven and Oosterbeek, 2008; Görlitz, 2011). Because plausibly exogenous factors and credible instruments are hard to come by and because our data provides us with longitudinal information on outcomes and training participation as well as new technology adoption, we use a fixed effects approach to examine the relation between wage, mobility and training.

This is implemented by regressing log hourly wages and the respective mobility indicators on the individual- and job-related characteristics listed in Table 1 as well as dummies for industries (2-digit level NACE code) and for occupations (1-digit level ISCO code) and an individual-specific fixed effect. In these specifications, the control variables of main interest are indicators for training participation, for new technology adoption and the combined occurrence of the two. Specifically, the training and new technology indicators measure the stock of (years of) training and (years of) new technology adoption, rather than simple indicators for participation in the preceding year. Using dummies indicating participation in the preceding year would not make sense in a longitudinal framework, because this would imply that any returns to training (or new technology) are short-lived and vanish completely after one year. Rather, similar to what is typically assumed for schooling, we prefer a specification where each (year with) training adds to the stock of human capital and (potentially) has a long-term impact on outcomes ^[4]. The combined occurrence of training and new technology refers to the stock of years in which both training and new technology adoption took place ^[5].

This fixed-effects approach allows us to account for time-invariant individual characteristics. Estimates can be interpreted as reflecting returns under the assumption that, after controlling for observed time-varying factors, differences in outcome levels between participants and nonparticipants would have followed parallel trends over time in the absence of training (or new technology adoption).However, this might not necessarily hold, given findings in Frazis and Loewenstein (2005) that training participants and nonparticipants might not only differ in wage levels but also in wage trends. To address this concern, we also estimate fixed effects models that control for individual-specific time trends (in the robustness section).

The section on returns starts out with descriptive evidence on wage differences between training participants and nonparticipants. Then the preferred specification based on fixed effects is shown, both for wages and two indicators for job mobility. Afterward, as a robustness tests, we present a specification controlling for individual time trends that takes into account that training participants might not only have different (e.g. wage) levels than nonparticipants due to unobserved characteristics but might also experience different (wage) trends, even in the absence of training. Next, we draw some conclusions on the selection into training in terms of unobservable characteristics that complements findings from Section 2.2. Finally, we probe whether the estimated wage and mobility results are heterogeneous by level of education.

Specifications (1) and (2) of Table 3 show the correlation between wages and training participation as well as new technology adoption. They are estimated using OLS and do not take the longitudinal nature of the data into account. Following most analyses focusing on training returns in a cross-sectional setting, training is measured by a dummy indicating training participation in the preceding year. Specification (1) shows that training participants have wages that are more than a quarter higher than the wages of nonparticipants. According to the interaction term, the wage premium of training participants compared to nonparticipants is smaller when a new technology is adopted in the same year than without a new technology. Yet, individuals who have experienced technological change during the preceding year have wages that are 16% higher on average. Specification (2) indicates that most of the wage premia are due to selection into training and new technology because compared to specification (1), the point estimates drop considerably when controlling for observable individual- and job-related characteristics.

The training premium in specification (2) is at 5.5% and the new technology premium is at 2.2%. This training premium of 5.5% is in the range of training returns estimated in other studies However, it should not be interpreted as a causal effect, as participants and nonparticipants may differ in unobserved characteristics. To account for time-invariant individual heterogeneity, our preferred specification includes individual fixed effects. Results are shown in specification (3) and in this longitudinal setting training and new technology refer to the cumulated stock of years with training and years of new technology adoption.

In the preferred specification, the relation of training to wages remains statistically significant. According to the point estimate, every training is associated with a permanent wage increase of around 0.9%. This estimate is close to the (insignificant) return estimated in Görlitz (2011) that relies on exogenous variation in training participation. Given that training participants have an average of 9.8 days of training and taking into account that most wage returns to a year of schooling are in the range of 7–10% (e.g. Card, 1999), the training return of 0.9% in the preferred specification is much more plausible than that of the OLS specifications. With respect to new technology, the point estimate is small and insignificant in the main specification, i.e. individuals adopting a new technology do not experience wage increases. The training-new technology interaction is insignificant as well, indicating that the wage returns of training do not differ between individuals with and without new technology adoption. However, note that the sum of the training effect and of the interaction term is close to zero and statistically not significant (0.009–0.008 = 0.001 with an F-statistic of 0.10 and a p-value of 0.758) and the sum of the training effect, of the technology effect and of the interaction term is small as well and not significant (0.009 + 0.003–0.008 = 0.004 with an F-statistic of 1.07 and a p-value of 0.301) which would imply that individuals participating in training and adopting a new technology do not fare better than individuals with neither training nor new technology adoption ^[6].

Table 4 shows the results of the preferred fixed effects specification on two mobility indicators. It shows that training has no significant relation to promotions (column 2), but training is associated with a reduction of the likelihood of a job change by 1% point (column 1). Taken together, the results suggest lower turnover and, consequently, higher job stability within the firm. In contrast, new technology adoption shows no clear association with promotions or job changes. Note that while for job change the interaction term for training and new technology is not statistically significant, it is of the same size but of opposite sign to the training coefficient. This implies that training has no significant association with job change if it is accompanied by new technology adoption in the same year (the F-test of significance of the sum of the training effect and the training × new technology interaction is 0.00 with a p-value of 0.987 and the F-test of significance of the sum of the training effect and the new technology effect and the training × new technology interaction is 0.01 with a p-value of 0.921).

Do these fixed effects estimates represent causal effects of training (and new technology) on outcomes? Not necessarily, given that individuals with and without training participation (new technology adoption) might not only differ in (e.g. wage) levels but also in (wage) growth patterns, as has been suggested by Pischke (2001) and Frazis and Loewenstein (2005). To address this concern Table 5 presents estimates that, in addition to individual fixed effects, also control for individual time trends. In this specification, identification of the estimates relies on individuals with at least three observations; therefore, individuals with fewer observations are excluded from the analysis ^[7].

Overall, the point estimates in Table 5 are relatively similar to those in Tables 3 and 4. Yet, the inclusion of more than 8,000 individual time trends leads to an increase in standard errors, which approximately triple for most outcomes and coefficients of interest. Accordingly, some of the previously significant results turn statistically insignificant. Specifically, for wages, the point estimate of training is 0.6% (slightly smaller than the 0.9% of the preferred specification) and insignificant (due to three-times larger standard errors than in the preferred specification). The point estimate of training for job change is at −1.4% points, i.e. it is even larger than the − 1.0% point in the preferred specification, but now becomes insignificant because of less precise standard errors. Since point estimates change only slightly, we interpret the insignificance of findings for wages and job change as a lack of power rather than a truly zero effect and conclude that our main findings are generally not challenged.

As a second test of robustness, we excluded two of our control variables, which might be considered as bad controls, namely information on part-time vs full-time employment and marginal vs regular employment. Both indicators might potentially be endogenous to training and technology adoption. Excluding these controls does not change our findings (results available from the authors upon request).

As a third test of robustness, we examined the endogeneity of training and technology adoption by controlling for leads of the explanatory variables. Any significance of those leads would indicate either reverse causality or anticipation effects. Note that controlling for leads results in a smaller simple size because outcomes from the first wave of data are dropped, given that no information on leads is available for this first wave. Table 6 shows that none of the leads of the technology indicator, of the training indicator or of the interaction of both is statistically significant at the 5% level. Thus, the results in Table 6 do not challenge our main findings.

Subsection 2.2 already presented information on the selection into training for observable characteristics and the drop of the wage premia of training and of new technology between Specifications (1) and (2) of Table 3 indicated that selection into training (and into new technology) in terms of observable time-variant characteristics is strong. This confirms previous findings on selectivity into training (e.g. Bassanini et al., 2007) and the adoption of new technology (e.g. Bartel and Sicherman, 1999). This subsection looks at selection in terms of unobservable characteristics.

The decrease of the wage premium of training between the specifications with and without fixed effects, i.e. Specifications (2) and (3) of Table 3, indicates that sorting into training is not only based on observable but also on unobservable characteristics. To learn more about the sorting in terms of unobservables, column (1) of Table A2 in  Appendix presents results from an analysis similar in spirit of Bartel and Sicherman (1999) where the prediction of the fixed effect from Specification (3) of Table 3 is used as outcome and regressed on information on the frequency of training participation and new technology adoption within the four-year period and observable characteristics that are constant over time (i.e. gender and migration status). The regression uses one observation per individual and focuses on those individuals who are observed in all four waves.

Results indicate that training participants are positively selected and that frequent training participants (i.e. those with 2 or more years with training in the 4-year period) are somewhat more positively selected than occasional training participants (i.e. those with 1 year with training in the 4-year period). Similarly, new technology adopters are positively selected as well and the size of selectivity is similar to training participants, at least for frequent new technology adopters. Occasional new technology adopters appear to be less selected than occasional training participants and less than frequent new technology adopters. The additional relation between parallel training participation and new technology adoption is not significant, i.e. for those individuals, the amount of selectivity equals the sum of the selectivity observed for training and the selectivity observed for new technology adoption. Overall, the positive selection might indicate that training participants and new technology adopters have traits that lead to higher wages, such as motivation and other relevant skills or it might represent the wage effects of training and of new technology adoption that took place before 2015.

Column (2) of Table A2 shows results when using the individual time trend estimated from Specification (1) of Table 5 as the dependent variable. According to these results, there are hardly any differences in individual-specific wage trends by the frequency of training or the frequency of new technology adoption. Only one of the coefficients is statistically significant. This corroborates that the estimates for wages are similar in Specification (3) of Table 3 and Specification (1) of Table 5 as it indicates that wage levels differ between training participants and nonparticipants and between new technology adopters and nonadopters, but not wage trends. This contrasts with findings in Frazis and Loewenstein (2005).

In this subsection, we investigate whether the estimated returns to training and to new technology differ by level of education, as previous research suggests ^[8]. We differentiate between those with low education (i.e. no vocational degree), those with medium education (i.e. vocational degree or degree from the highest school track Abitur, which is our reference category) and those with high education (i.e. college degree). The estimated specifications are similar to our preferred specification, i.e. they include fixed effects but no individual time trends, and the explanatory variables of main interest are interacted with the group indicators. Results are shown in Table A3 in  Appendix.

We find that the return to training on wages is absent for low educated individuals (while the interaction term is not significant, the point estimate of the interaction with low education is larger and of opposite sign to the training effect in the baseline group and an F-test of significance of the sum of the baseline training effect and low education interaction is 0.03 with a p-value of 0.868). Furthermore, while new technology has no relation to wages of employees on average, it does have a significant positive association with wages of highly educated individuals (the F-test on the sum of the baseline new technology effect and the high education interaction is 11.71 with a p-value of 0.001), but apparently only if this is not accompanied by training. This is in line with evidence showing that technological change is skill-biased and in favor of highly educated individuals (e.g. Bartel and Sicherman, 1999). In contrast, for medium-educated individuals in the reference group, the new technology variable now indicates even a negative association with wage, which is significant at the 10% level. The heterogeneity in the wage associations of technological change by level of education, favoring highly educated individuals and harming medium-educated individuals, conforms with results from Cortes (2016). With regard to mobility, associations between training and job change are only present for medium-skilled individuals in the reference category but do not exist for low or highly educated employees (F-statistic for the sum of the baseline training effects and the interaction terms being equal to zero is 0.83 for low and 0.10 for high education with p-values of 0.361 and 0.746, respectively). Note that for medium-skilled individuals in the reference category, the interaction term of training and new technology is of opposite sign to the training effect, similar to the preferred result in Table 4 and now also statistically significant. This implies that training is correlated with increased job stability (of medium-skilled employees) only if it is not accompanied by new technology.

Labor markets around the world are undergoing tremendous changes due to technological progress. New machines or processes at the workplace can fundamentally transform the nature of work, and the rise of AI, particularly generative AI and LLMs, has the potential to accelerate these shifts. Training is often regarded as the key instrument for individuals to keep up with the changing demands for skills and qualifications. For workers, additional human capital is often acquired through training. Using panel data from Germany, we investigate the interaction of training and new technologies in regard to wage returns and two indicators for job mobility. Our results show that new technologies go along with increased training participation rates. We also find that wages of training participants are on average about 0.9% higher, which is comparable to estimates reported in studies using exogenous variation in training participation. In contrast, no statistically significant association is observed between new technology adoption and wages. With regard to the joint occurrence of training and new technology adaption, on the one hand, we do not find that the interaction term is statistically significant when looking at the association with wages – hinting at similar wage returns for training with and without parallel technology adaption. Yet, on the other hand, when looking at the sum of the training and interaction coefficients and F-statistics for joint significance, the relation between wages and training is closer to zero and not statistically significant for training participants with technology adoption, while it is significant for training participants without technology adoption. With respect to the mobility indicators, we observe that training is associated with a lower probability of working in a new job. This result hints towards a binding mechanism of human capital as individuals who acquire knowledge are less likely to leave their employer, which the theory predicts for firm-specific human capital. Like for wages, the interaction term for the parallel occurrence of training and new technologies adoption is not significant, but there is evidence that the binding mechanism of training is not significant for training participants with parallel technology adoption.

Our results in terms of returns to training mostly confirm previous findings. What we do not find are any larger returns to training when combined with new technology adoption. In contrast, there is evidence that the combined occurrence of new technology adoption and of training participation does not make individuals better off in terms of wages or job stability compared to individuals experiencing neither training nor new technology adoption. This suggests that training that is triggered by new technologies is not associated with higher returns than other training ^[9]. This casts doubt on whether increased training participation in the future, which is required by technological change, will come from a larger willingness of employees to pay for training.

By now, it is mostly firms paying for training, which is plausible when firms have monopsony power (Acemoglu and Pischke, 1998, 1999). An increased willingness of employees to invest financially and personally in training may be facilitated by higher wage returns, for instance, if monopsony power diminishes and workers receive a greater share of productivity gains. Expanding the use of training certificates and competence-based credentials could contribute to this by mitigating information asymmetries about employee productivity (Acemoglu and Pischke, 2000).

Furthermore, we find that individuals on average do not profit from new technologies at their workplace in terms of wage or job mobility aspects. At the same time, however, new technologies on average do not pose a threat to wages (conditional on staying employed). Yet, similar to Cortes (2016), we find heterogeneous relations between highly educated individuals who benefit from technological change and medium-educated individuals who seem to be harmed. Moreover, there might be other benefits to new technologies at work apart from wage or job mobility. These could be safer work environments, changes in the working time or the outsourcing of tedious or repetitive tasks. Antón et al. (2020) look at the effect of robots on nonmonetary working conditions and find an increase in the domain of work intensity and no impact on physical environment or skills. Since robots are a very special technology and measures for their adoption are only available on the industry level, there is still room for more micro-level research.

We acknowledge helpful comments and suggestions from Ronald Bachmann, Katja Görlitz, Eduard Storm, and David Zuchowski. We also thank participants at the Young Economists’ Meeting 2023 and participants in internal seminars at the RWI – Leibniz Institute for Economic Research.