In the United States, income concentration has increased significantly, according to US Bank Economics in 2026. Automation has produced real productivity gains over forty years. But these gains have often translated into dividends rather than wage increases. The question that imposes itself is not whether robots create value: it is understanding why income inequality has increased more in the United States than in several Nordic countries.
The essentials
- A 2023 study argues that a framework including robots can explain a significant portion of the American gap between wages and productivity.
- The relationship between robot adoption and working conditions follows an inverted U-curve: marginal improvement in the short term for low-skilled workers, displacement in the long term.
- The IMF estimates that approximately 40% of global employment is exposed to AI, with risk distributed unequally across countries and skill levels.
- Scandinavian countries present levels of automation and inequality trajectories different from those of the United States, influenced by deliberate choices in training, unionization, and taxation.
- Daron Acemoglu estimates an average annual productivity gain of approximately 0.05% over ten years in his central scenario.
Forty years of data tell a precise story
Fixed-effects studies allow comparison of factories, sectors, or regions that have adopted robots with comparable ones that have not. Fixed-effects studies show that in the short term, the introduction of industrial robots marginally improves conditions for low-skilled workers. Workplace accidents decline. The most arduous tasks are automated. Some wages progress slightly, driven by increased productivity.
Then the curve reverses. Automation can eliminate certain positions and alter working conditions; its overall effect depends in particular on productivity gains and the creation of new tasks. This trend is documented across several economic cycles. Automation can first penalize directly replaced labor; the potential productivity effect favorable to factor prices occurs later.
Gains may initially increase the capital share in automated sectors, but their final distribution is not exclusively reserved for capital owners. The share of labor income in American GDP has declined since the early 1980s. In sectors where it occurs, automation has been associated with a decline in labor’s share of added value, which is not sufficient to establish that profit growth is faster than wage growth. US Bank Economics points to income inequality.
The mechanism Johnson identifies
Simon Johnson, economist at MIT and co-author of Power and Progress, frames the issue in precise terms in his article The Simple Macroeconomics of AI published in 2025. According to Acemoglu, the central average annual effect of AI on productivity would be approximately 0.05% over ten years. Conversely, if AI is oriented toward complementarity with workers—making them more efficient, more capable, better informed—the gains are structurally higher and better distributed.
The distinction is not philosophical. It is measurable in companies’ investment choices. Automating a warehouse to eliminate warehouse worker positions generates a net gain for the shareholder and a net cost for the worker. Equipping that same warehouse worker with an intelligent logistics assistant that allows him to manage twice as much flow with fewer errors creates value in both cases and leaves room for wage negotiation.
Johnson argues that this bifurcation does not resolve itself spontaneously through the market. Companies maximize their individual profit. Orientation toward complementarity requires tax incentives, regulation of AI use, and institutions capable of weighing in on technological deployment. Without regulation or incentives, technology tends to reinforce existing power relationships.
The authors estimate that recent experience contradicts the idea of an automatic and widespread diffusion of automation gains.
The competing argument deserves to be taken seriously
Tyler Cowen, economist at George Mason and careful observer of innovation dynamics, defends a different reading. His thesis: the productivity gains from automation, even if initially captured by capital, ultimately diffuse through the decline in prices of goods and services. The ordinary consumer benefits from lower consumption prices. Electronics, logistics, industrial food production—all these intensely automated sectors have seen their prices fall in real terms over forty years.
This point is empirically defensible. The contestation concerns its relative magnitude. Consumer prices have indeed fallen on certain items. But constrained expenditures can weigh heavily on modest households, though recent data do not show overall that they have grown faster than their incomes. The gain in purchasing power on a television does not offset the rise in rent.
Diffusion through prices is real but partial, and it operates on a time horizon far longer than job losses.
Automation widens inequality along lines that are not merely salary-based. American data indicate that displaced workers are unevenly distributed across regions, demographic groups, and skill levels. The diffusion of gains through prices, when it exists, benefits all consumers. Job losses strike selectively.
The distinct choices of Scandinavian countries
International comparison is the most robust test available. Levels of automation differ between Denmark, Sweden, Finland, and the United States. Denmark, Sweden, and Finland have relatively low inequality, but their trajectories are not all stable.
Three mechanisms explain the gap.
Training first. Scandinavian systems invest massively in active professional retraining—long, funded pathways that facilitate occupational transitions without guaranteeing full maintenance of previous income. Denmark devotes around 2% of its GDP to active labor market policies, compared to less than 0.1% in the United States.
Collective bargaining coverage is very high in Nordic countries. In Nordic countries, a large share of wages and working conditions is determined by collective negotiation, which can influence the sharing of productivity gains. Workers structurally participate in sharing these gains.
And finally, taxation. High marginal rates on capital income, combined with robust redistribution mechanisms, reinject into the system a fraction of gains captured by owners. Public subsidies to companies without wage requirements, a well-documented mechanism in the United States, do not exist in this form in Scandinavia.
These three choices result from explicit political decisions, maintained over several decades, sometimes at the cost of major social conflicts. The Scandinavian model is constructed, not spontaneous.
IMF signals for the next decade
The IMF estimates in its note Gen-AI: Artificial Intelligence and the Future of Work that nearly 40% of global employment is exposed to AI. Approximately 40% of global employment is potentially exposed to AI, with possible effects of complementarity or substitution.
The distribution of this risk is radically unequal. In high-income countries, exposed jobs are often skilled jobs—analysts, lawyers, accountants, writers—for which complementarity with AI is technically possible. In middle- or low-income countries, they are unskilled manufacturing and service jobs, for which the alternative does not yet exist.
In the United States, exposure to AI extends across multiple sectors. White-collar workers in the financial and legal sectors and low-skilled workers in warehouses, food service, and transportation are among the most exposed. The previous shock wave—industrial robotics in the 1990s-2010s—had struck factories. The next wave of automation will affect a much more diverse set of jobs.
What does sharing gains look like in 2035
Economists pose the question of the trade-off between mass automation and increasing worker incomes. Automation has contributed to inequality and a relative decline in certain wages, among other economic and institutional mechanisms. The Scandinavian model constitutes another option, tested at large scale. Neither mechanically transfers to the American context of 2025.
Several trajectories appear possible by 2035, depending on the institutional choices that will or will not be made in the years to come.
The United States has a low unionization rate and significant robot adoption; if this trajectory persists, income inequality could continue to increase. A significant share of AI gains could accumulate among owners of the models and infrastructure. The qualified middle class will see part of its work augmented, another part displaced. Low-skilled workers will suffer displacement without sufficient retraining safety nets. This scenario does not necessarily produce a global economic catastrophe: GDP indicators can remain solid while inequality progresses.
Data from the last four decades show that a significant portion of changes in American wage structure are linked to the displacement of routine tasks by automation.
A favorable bifurcation requires active choices. Taxation on capital gains generated by AI that would finance retraining funds. Transparency requirements on automation decisions, allowing worker representatives to anticipate and negotiate. Public investments in continuous vocational education, calibrated to expanding trades rather than declining ones. Consultation and transparency instruments exist and have been tested, while the existence of taxation on AI capital gains financing retraining remains to be established.
They are not hypothetical.
The signal to watch is less the speed of AI adoption than the evolution of the ratio between labor income share and capital income share in the sectors automating fastest. If this ratio stabilizes or improves in the next two or three years, this would indicate that redistribution mechanisms, negotiated or regulatory, are beginning to function. If it continues to deteriorate at the same pace as since 2015, the concentration scenario becomes structurally entrenched.
The construction of sharing institutions in Scandinavia extended over several decades. AI is already transforming certain tasks in the short term, but generalized effects on employment and incomes remain so far limited and uncertain. But faster institutional changes are possible; the New Deal was built in less than a decade. The trade-off between the two trajectories depends on the design and deployment decisions of companies, public rules, and collective bargaining that determine whether technology works for whom.
Sources
- ScienceDirect (2024), Fixed-effects studies on industrial robot adoption and working conditions over 40 years: https://www.sciencedirect.com/science/article/abs/pii/S0160791X24001696
- Simon Johnson, The Simple Macroeconomics of AI (2025), Economic Policy: https://doi.org/10.1093/epolic/eiae042
- US Bank Economics (2026), Income Concentration in the United States, record level in 60 years (US Bank Economics report, 2026)
- IMF, World Employment Outlook: 40% of global jobs exposed to AI and robots (International Monetary Fund)
- World Bank, Share of labor income in American GDP, long-term data (World Bank Open Data)
