Raising educational attainment is the canonical way skills keep pace with technology. This column argues that changes in learning content matter just as much, drawing on five decades of legally binding German vocational training curricula linked to patents and administrative worker records. Occupational curricula more exposed to digital technologies are rewritten faster, mainly by adding new skills rather than discarding old ones. Entrants trained under the new content earn higher wages; older incumbents, who had completed the older curricula, see their wages fall.
Generative AI has revived the familiar worry that technology will erode the labour market value of human expertise (Acemoglu et al. 2023). Policymakers are responding by promoting access to learning and seeking to improve its labour market returns. The European Commission’s 2025 union of skills aims to lift annual adult participation in learning, while the US Workforce Innovation and Opportunity Act links training-provider eligibility to employment outcomes (European Commission 2025, US Congress 2014, 2026). Both policies raise the same question: how does human expertise keep up as technology advances?
The canonical economic answer is to raise educational attainment, so that skill supply keeps up in the race between education and technology (Tinbergen 1975, Goldin and Katz 2008), since educated workers learn new technologies most readily (Hassan et al. 2026). Yet much of the change that technology brings to skill demand takes place within occupations rather than through shifts in employment between occupations (Spitz-Oener 2006, Atalay et al. 2020). In a new paper (Salomons et al. 2026), we study a complementary margin: changing what people learn within existing qualifications.
Whether educational content adapts to technology, and whether that reinstates the value of workers’ skills, is difficult to establish: in most systems, what is taught is decided locally and not systematically documented. German vocational training is an exception. Since the 1969 Vocational Training Act, it has been codified in legally binding national curricula, revised through negotiations between employer associations, unions, and the Federal Institute for Vocational Education and Training (BIBB).
We collect 756 curricula covering 492 training occupations over 1971–2021, the near universe of formal non-college training in a country where most workers are vocationally trained. Linking these to administrative employer–employee records reveals the version of the curriculum each worker completed and how the worker fared in the labour market.
Curricula describe skills in fine detail. We exploit that with word embeddings, measuring how closely each curriculum’s text resembles the text of US patents (Seegmiller et al. 2023). We focus on breakthroughs in digital technologies (Kelly et al. 2021), lagged 20 years (as in Autor et al. 2024), to allow for diffusion of those breakthrough technologies into production. Because breakthroughs are unexpected jumps in the frontier, they are unlikely to be a response to changes in training.
Curricula are revised unevenly. We find that among those ever revised, the average revision interval is 15.3 years; technology exposure compresses that interval. Fifteen years after a curriculum first appears, 75% of those with below-median exposure to digital technology are still unchanged, against 51% of highly exposed ones. Controlling for calendar year, a one-standard-deviation increase in exposure raises the annual revision probability by about a third and cuts the time to revision by 1.6 years, also within broad occupational groups.
Updates to the curricula substantially change training content, and in the expected direction. Digital technology codifies routine tasks and complements non-routine ones (Autor et al. 2003); scoring curricula against task descriptions, we find non-routine task intensity rising by about 0.04 standard deviations a year among revised curricula. It rises almost three times faster in occupations that are highly exposed to technology than in those that are less exposed.
Content can change along three margins: skills can be added, the weight on existing skills can shift, or skills can be removed. A curriculum that only sheds automated routine tasks narrows the trainees’ skill set; one that adds non-routine content gives them expertise that their predecessors lacked.
We find that addition dominates. Figure 1 compares, for each curriculum update, the non-routine intensity of added, reweighted, and removed content against the retained content. Points above the 45-degree line are more non-routine than that retained baseline. Newly added content lies further above it, while reweighting and removal play smaller roles. Curricula thus add new expertise rather than merely shedding automated tasks, in line with evidence that new technologies create new work (Acemoglu and Restrepo 2018, Autor et al. 2024, 2026).
Figure 1 Curricula become less routine mainly through newly added content
The new content emphasises social and communication skills such as teamwork and negotiating, followed by digital and management skills, while physical and manual skills lose importance. This mirrors demand: the return to social skills has risen (Deming 2017), and vacancies increasingly ask for technical and socio-emotional skills together (Borgonovi et al. 2023). Workers trained under more recently revised curricula also report finding their training more useful.
This margin is quantitatively important. Assigning trainees the task content of their curricula, within-occupation curriculum changes account for two-thirds of the total shift towards non-routine vocational skill supply in West Germany between 1976 and 2021; the remaining third reflects changes in trainees’ occupational choice.
Is this new expertise valuable in the labour market? We compare successive cohorts of labour market entrants trained before and after a curriculum update in the same occupation with corresponding cohorts in occupations without an update, stacking 357 update events in a difference-in-differences design. Identification rests on timing: skill demand moves gradually, whereas a new curriculum arrives on a documented date.
Entrants trained under updated curricula earn up to 4.1% higher daily wages over their first five years post-training, and the premium grows for successive post-update cohorts. The growth is concentrated in technology-exposed occupations, where it reaches 6.3% for the last cohort we consider, compared with at most 2.5% in less exposed occupations (Figure 2).
These gains appear to reflect new skills, as they are not explained by who enters training or which firms provide it, and updates in occupations with little apparent need for new skills generate no gains. Updated-curricula cohorts are also less likely to leave their trained occupation, consistent with a better skills match, and firms exposed to newly trained workers invest more in capital, consistent with new skills complementing new technologies (vom Baur 2026).
Figure 2 Updated training raises entrants’ wages, especially in technology-exposed occupations
The same revision that reinstates expertise for entrants erodes it for workers already in the occupation. Janssen and Mohrenweiser (2018) document this for one computer numerical control machine (CNC)-driven revision; our data let us ask whether it holds across the system. Following incumbents with at least five years of occupational tenure, we find wage losses concentrated among older workers. Five years from the curriculum update, wages are 3.4% lower for incumbents aged 50–59 and 1.9% lower for those aged 40–49 than in occupations without revisions. Incumbents aged 30–39 lose little in wages but leave their occupation markedly more often. These losses are consistent with skill obsolescence. Firms do not close the gap: most provide employer-paid training each year, but there is no additional response around curriculum updates.
Figure 3 Older occupational incumbents experience relative wage losses after curriculum updates
Three implications follow from our findings. First, what people learn deserves attention alongside how many people train, and that means educational content must be measured. Second, institutional design may matter: employer involvement keeps skills labour-market relevant; unions and the Federal Institute for Vocational Education and Training keep content general rather than firm-specific; and coordination relieves the poaching problem that otherwise leads firms to underinvest in skills upgrading. Third, renewing formal training is not a complete adjustment strategy. Updates signal where incumbents’ expertise is losing value, and the skills specified in the curriculum updates could guide continuing training for workers trained under earlier standards – which our results suggest will not emerge automatically within firms.
Our evidence does not identify the skills that generative AI will reward next, but it points to questions of educational design rather than attainment. How quickly the new expertise demanded by AI is codified into training will shape who can acquire it. These are good reasons to watch not only how many people are trained, but what is written into curricula.
Source : VOXeu
The theory of optimal currency hedging has been well established for decades. This column uses…
The human and economic costs of extreme weather events are already large and will grow…
Geopolitical tensions are changing the nature of supply-chain risk. This column presents new survey evidence…
Firms’ access to external finance differs widely across Europe. This column uses data from the…
Whether inflation can ease public debt burdens is an increasingly important question as debt sustainability…
Senders now have more ways than ever to move money across borders, yet costs remain…