In a new working paper, Magnus Lodefalk, Lydia Löthman, Michael Koch, and Erik Engberg examine how generative AI is reshaping the labor market. They find little evidence that AI has cut the total number of jobs, but show that it has slowed hiring for the youngest workers, especially in the AI-exposed occupations where young women are concentrated. Over time, AI’s effect on entry-level roles risks thinning the next generation’s ability to build the skills and networks that careers are made of.
Since late 2022, stock prices have risen sharply while job openings have fallen almost as steeply. In the United States, equity prices have risen by more than 70 percent while job openings have fallen by over 30 percent. The divergence has become one of the most-cited statistics in discussions of artificial intelligence and the labor market, and for many observers the interpretation is straightforward: generative AI is replacing workers, depressing labor demand while lifting corporate profits. Our research, using full-population register data for Sweden and 4.6 million job advertisements, finds that the evidence does not support this reading of the aggregate picture. The underlying concern, however, is well founded. Generative AI is not reducing employment, but it is changing who gets hired.
The aggregate decline is a monetary policy story
Some commentators attribute the divergence between stock prices and job openings to the rapid adoption of generative AI, while others emphasize macroeconomic forces, particularly the tightening of monetary policy. We examine these competing explanations using Swedish data and a useful timing gap. Sweden’s central bank, the Riksbank, began raising interest rates in April 2022, seven months before the launch of ChatGPT in November 2022. This sequence lets us separate the effects: if monetary tightening is the primary driver of declining job postings, the initial downturn should coincide with the rate hike. If generative AI is responsible, the effect should emerge only after late 2022 and be concentrated in occupations with high exposure to the technology.
Using the full population of advertisements published on Platsbanken, Sweden’s public employment service vacancy portal, and matching these data to an occupation-level measure of generative-AI exposure (the DAIOE index, Engberg et al. 2024), we find little evidence that ChatGPT triggered the broad decline in vacancies. Postings began falling across AI-exposure groups from April 2022 onward, immediately after the Riksbank’s first rate hike. Among high-AI-exposure occupations, the relative decline after the rate hike was sizable, around 12 percent, while the additional decline after the November release of ChatGPT was statistically imprecise and not robustly different from zero. An occupation’s exposure to AI is also uncorrelated with its sensitivity to interest rates, which suggests that monetary tightening and generative AI represent distinct channels. This echoes a point Derek Thompson has made for the United States: the decline in job postings appears to track the Federal Reserve’s tightening cycle more closely than the launch of ChatGPT.
Within employers, the age gradient is large and accelerating
The aggregate posting data conceal a different pattern inside individual employers. We draw on monthly employer declarations to the tax authority, which list everyone on each employer’s payroll, month by month, across the entire Swedish workforce. Linked population registers supply each worker’s age, sex, and occupation. Because we follow the same people from month to month, we can see not just how many workers an employer has but who joins and who leaves: someone newly on a payroll is a hire, and someone who stops appearing is a departure, whether a layoff, a quit, or a move elsewhere. We then compare how employment in the most AI-exposed occupations, such as software development, payroll and human-resources work, and customer service, evolved against less-exposed roles within the same employer, before and after ChatGPT. Holding the employer fixed nets out company-wide hiring freezes or expansions, so what remains is the shift across occupations of differing AI exposure.
By the first half of 2025, employment of workers aged 22-25 in the most AI-exposed occupations had fallen by 5.5 percent relative to less-exposed occupations within the same employers, with a tight confidence interval running from 4.9 to 5.8 percent. By contrast, workers aged 31-49 were essentially unaffected. Workers over 50 showed a small positive gain, although that result is more sensitive to how a role’s exposure to AI is measured.
The adjustment occurs almost entirely through reduced hiring. For workers aged 22-25, the decline in hires is large and precisely estimated, while the corresponding change in separations (layoffs and quits) is only about one-quarter as large. This distinction matters in the Swedish context. Employment protection is organized around a last-in-first-out principle, which constrains dismissals of incumbent workers. High-AI-exposure employers therefore do not appear to be laying off junior workers in large numbers. Rather, they are slowing entry into these jobs: not replacing departing workers and not bringing in new cohorts at the same rate as before. A similar pattern appears in recent U.S. evidence. Seyed Hosseini and Guy Lichtinger use resume and job-posting data covering roughly 62 million workers across 285,000 firms and find that junior employment falls sharply in generative-AI-adopting firms, while senior employment is largely unchanged. As in Sweden, the decline is concentrated in AI-exposed jobs and operates mainly through slower hiring rather than increased separations. The hiring-freeze pattern thus appears to be emerging across institutional settings.
Young women face a larger adjustment
The most under-examined finding concerns young women. Women are overrepresented among 22-25 year olds in Q4, the segment experiencing the largest relative decline in employment. This pattern is driven by occupational sorting, as young women are disproportionately employed in payroll, administrative, and customer service roles that are concentrated in Q4. For example, payroll administrators are 82 percent female and among the most AI exposed occupations in Sweden, while customer service agents and receptionists are 66 percent female. Men are also overrepresented in some highly exposed occupations, such as software development, which is around 80 percent male, but these occupations make up a smaller share of the 22-25 Q4 group. Overall, women make up about 54 percent of this group, a share that remained largely unchanged during our sample period.
The numbers track this concentration. Among workers aged 22-25, employment in high-exposure occupations fell about 1.6 percent for women, roughly twice the 0.7 percent decline for young men. About two-fifths of that gap reflects the occupational sorting: young women are simply more concentrated in the most exposed jobs. The remaining gap is harder to interpret, and we are deliberately cautious about it. Young men’s comparison group, those in less-exposed roles, was itself shrinking over this period, so the smaller decline measured for men may rest on an unstable benchmark rather than reflect genuinely gentler treatment. We therefore treat the occupational-concentration channel as the firmer of the two findings: the clearest reason young women are more affected is that they are more often in the jobs AI reaches first.
This is consistent with simulation evidence for Sweden from Malin Gardberg et al., who show that pre-existing occupational sorting by gender can widen the post-AI wage gap. The International Labour Organization similarly estimates that female-dominated occupations globally face almost twice the generative-AI exposure of male-dominated ones. Our results suggest these risks are already visible in Swedish employer hiring decisions.
Why outcomes differ across countries
The emerging evidence points to substantial cross-country heterogeneity. For the U.S., Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen document a sizable employment decline for young workers in AI-exposed occupations, and we find a similarly sizable decline in Sweden. By contrast, Antti Kauhanen and Petri Rouvinen find no comparable effect in Finland, using population-level data and a closely related research design. The Swedish result is not an artifact of method: when we reweight our sample to match Finland’s industry and occupation mix and apply Finland’s preferred exposure measure, the decline barely moves.
What explains the divergence remains uncertain. Plausible candidates include differences in the intensity of AI adoption, the occupational structure of youth employment, employer hiring practices, and vocational-training pathways that may channel young Finns into less-exposed entry-level jobs. The cross-country heterogeneity is therefore not a nuisance result; it is part of the substantive finding. Institutions, practices, and norms appear to shape how the same underlying technology is translated into labor-market outcomes.
A skill-formation concern
Job counts alone understate what is at stake. Entry-level positions in AI-exposed occupations are not just jobs; they are training grounds. They are where young workers accumulate applied skills, professional networks, institutional knowledge, and the tacit know-how that later becomes mid-career capability. When employers stop bringing in new cohorts, they are not only reducing near-term labor costs. They are also weakening the pipeline of experienced workers they will need in the future.
Brynjolfsson, Chandar, and Chen describe young workers as the “canaries in the coal mine,” early signals of a broader labor-market adjustment. If the entry-level hiring slowdown persists, the result will be more than a temporary employment dip. It risks becoming a generational skill-formation gap: fewer workers learning by doing, fewer junior employees moving into senior roles, and a thinner future stock of occupational expertise. Such a gap may show up first in career trajectories, firm capabilities, and productivity growth, long before it appears in headline unemployment figures.
Implications
Three practical implications follow. First, monitoring needs to become more granular. Aggregate unemployment rates and headline job-posting indices are unlikely to detect this adjustment early. What makes it visible is high-frequency, employer-level data that track hiring composition by age, occupation, and exposure to new technologies. Countries with employer-declaration registers of the kind Sweden maintains therefore have a significant informational advantage; the U.S., lacking a comparable register (to the best of our knowledge), must rely on private payroll and resume data. Second, education and training systems need to adapt. If entry-level positions in AI-exposed fields contract, universities and vocational programs cannot assume that firms will keep providing the same amount of early-career learning by doing. Applied learning, work-integrated training, and structured transition pathways will need to play a larger role in supplying the experiential component that employment has historically provided. Third, policymakers should watch for coverage gaps. Recent graduates who cannot find work, and who have not yet accrued eligibility for unemployment-insurance benefits, may be largely invisible to the social-insurance system. Monitoring whether this group is growing disproportionately in AI-exposed fields should be a priority. The policy challenge is not only to measure job loss after it occurs, but to detect blocked entry before it hardens into a persistent career disadvantage.
Author disclosures: Funded by the Torsten Söderberg Foundation (grants E46/21, ET3/23) and WASP-HS (grant 805). Employment data from Statistics Sweden via the MONA platform under ethical approvals detailed in the working paper. No conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
