Yves here. It may seem a bit dog-bites-man-ish to have a study verify that AI is working as intended, as in reducing pay to workers to the benefit of enterprises and tech titans. But this finding is important from a policy perspective. Many AI touts contend that AI will still create more jobs through some magical not-well-explained process. So debunking those claims is important.
By Antonio Minniti, Associate Professor of Economics University Of Bologna, Klaus Prettner, Professor of Economics Wu Vienna University of Economics And Business, Francesco Venturini, Permanent Visiting Fellow National Institute Of Economic And Social Research; Assistant Professor in Economics University of Perugia, and David Bloom, Clarence James Gamble Professor of Economics and Demography Harvard University. Originally published at VoxEU
The swift rise of artificial intelligence is raising fundamental concerns about the future of work. This column uses data from 238 regions across 21 European countries to examine how AI-related innovation influences the distribution of income between labour and capital, and among different skill classes of labour. Regions with more intense AI patenting tend to experience a decline in the labour share of income, especially in areas with a strong industrial base, indicating that AI acts as a capital-biased innovation. Without appropriate policy intervention, this trend could exacerbate existing inequalities.
The swift rise of artificial intelligence (AI) is reshaping production systems and raising fundamental concerns about the future of work (e.g. Korinek and Stiglitz 2019, Webb 2019, Prytkova et al. 2024). Will AI mainly support human labour or render it redundant? Who will capture the benefits of this transformation – workers or capital owners? And among workers, who will fare relatively better – the more- or less-skilled?
Although these questions dominate much of the public debate on AI, robust empirical evidence remains limited. In a recent article (Minniti et al. 2025) we offer new insights into how AI-related innovation influences the distribution of income between labour and capital, and among different skill classes of labour, in European regions.
The results show that regions with more intense AI patenting tend to experience a decline in the labour share of income, especially in areas with a strong industrial base. This pattern indicates that AI acts as a capital-biased innovation, shifting the returns from technological progress increasingly towards capital.
If not addressed through targeted policies, this trend could exacerbate existing inequalities and pose long-term challenges to social cohesion in advanced economies.
AI Innovation, Skills, and the Changing Distribution of Income
We examine the relationship between AI-driven innovation and the labour share using data from 238 regions across 21 European countries, spanning the period 2000–2017. 1 This regional lens is essential, as differences in industrial structure, educational attainment, and local innovation dynamics strongly mediate the effects of technological change.
We measure AI innovation through a newly constructed dataset of AI-related patents, including technologies such as machine learning, natural language processing, image-processing methods, and so on. These data are linked to detailed regional indicators on employment, wages, and value added, disaggregated by skill level.
The findings reveal a robust and statistically significant negative association between AI patenting intensity and the labour income share. A doubling in AI patent intensity is associated with a 0.5 to 1.6 percentage point reduction in the labour share. Figure 1 illustrates this development, displaying AI intensity across European regions (left panel) and the cumulative change in the labour share between 2000 and 2017 (right panel). In regions with higher AI intensity, the labour income share tended to decrease during the overall time span.
Figure 1
Note: The left-hand side of the graph reports regional AI patenting intensity, defined as technologically revealed comparative advantage, calculated as the average count of AI patents relative to that of other patents filed in the region between 2000 and 2017. The resulting value is then expressed as a ratio to the same measure for the other regions. If the index is greater than 1, the region is specialized in AI patenting. The right-hand side shows the cumulative change in the labour share between 2000 and 2017, expressed in percentage points.
Crucially, the decline in the labour share is not uniform across the workforce. We find that medium- and high-skilled workers experience the strongest reductions in their income share, primarily driven by wage compression as opposed to changes in employment. 2 By contrast, the labour share of low-skilled workers declines less – largely because of mild employment growth in this segment, which partly offsets stagnant or falling wages. These patterns suggest that AI is reshaping the labour market not only between labour and capital, but also within labor, by redistributing returns across skill groups (see also Bloom et al. 2024 and 2025 for a theoretical framework that is consistent with these results).
Importantly, we do not observe proportional gains in labour productivity that might offset these distributional shifts. In this context, AI not only raises efficiency but also reallocates the gains from innovation, amplifying the share of income accruing to capital.
From Routine Substitution to Skill Compression: Rethinking Polarisation
While the decline in the labour share began long before the emergence of AI, the evidence suggests that AI innovation is reinforcing and accelerating this trend. Historically, technological progress has often complemented skilled labour while replacing routine, low-skill tasks. By contrast, modern AI is increasingly able to replicate cognitive work, expanding the scope of automation into high- and medium-skill occupations.
This shift is particularly visible in the service economy. AI applications are now used in law, finance, logistics, public administration, and other white-collar domains previously considered resistant to automation. As a result, the current wave of innovation does not conform to the classical narrative of low-skill job displacement followed by skill upgrading. Instead, we observe signs of wage compression at the top and middle of the skill distribution.
The net effect is a new form of labour market polarisation, one that does not necessarily benefit high-skilled workers as in past waves of technological change. Rather, the evidence points to a redistribution of economic value away from labour across the board, with particularly adverse effects for those previously insulated from technological disruption. This occurs even against the backdrop that AI raises average productivity and economic growth (e.g., Acemoglu 2025).
In the European context, where concerns over wage stagnation, inequality, and political discontent are mounting, this evolving dynamic demands close attention.
Policy Implications
The economic effects of AI diffusion are not deterministic – the distributional outcomes will depend on how policymakers respond to the deployment and increasing use of this technology. To mitigate the adverse labour effects, investing in human capital is crucial, particularly through the development of lifelong learning systems and vocational programmes that equip workers with the skills to transition into AI-complementary roles.
In tandem, tax systems must adapt to the shifting balance between labour and capital. As capital increasingly captures the returns from production, fiscal frameworks could be developed to ensure equity and revenue sustainability. Potential policies could reduce labour taxation and move to a (higher) taxation of pollution, which has negative externalities that would be efficient to reduce anyway; a (higher) taxation of land, which is fixed in supply so that its taxation is non-distortionary; and a (higher) taxation of consumption, which will continue to grow in the presence of AI-driven economic growth (for a detailed discussion of such policies see Prettner and Bloom 2020, especially Chapter 7). Moreover, to avoid reinforcing geographic disparities, economic policy could actively promote the diffusion of AI technologies across regions and sectors. Broadening AI adoption can help ensure that productivity gains are not confined to a few innovation hubs but instead contribute to inclusive and regionally balanced growth.
If governed effectively, AI could appreciably raise aggregate economic wellbeing. Without appropriate policy intervention, however, it may reinforce structural divides – between labour and capital and between core and peripheral regions.
_______
- According to the Nomenclature of Territorial Units for Statistics (NUTS) of the EU, the data refer to the NUTS-2 level (basic regions).
- The skill types are defined using the ISCED classification system, which segments educational attainment into three categories: Low (early childhood education; primary education. lower secondary education), Medium (upper secondary education. post-secondary non-tertiary education), and High (short-cycle tertiary education; bachelor’s degree or equivalent tertiary education level).
See original post for references
24 comments
Comments are closed.
I saw Microsoft’s AI CEO say that most white collar jobs will be automated in 12-18 months so I then ask the question:
There are estimated 90m people employed in white collar jobs in the US, how will the economy (which is based on consumerism and debt) be able to function with 90m people out of work?
I think you need to extend the question not just to white collar jobs but all jobs in which a computer is the main tool.
The people whose job consists of going to meetings might be in trouble. In the short to medium term, it’s also possible that IT positions will increase because someone needs to monitor all the janky tech that doesn’t work properly but was rolled out anyway.
But given that so many companies are bringing in “AI” without having any defined use for it ahead of time, I don’t see any huge job loss right away. In fact there will probably be increasing demand for more meetings about what to do with the “AI”. That has been the case at a workplace I’m familiar with.
Microsoft is involved in those massive circular financing agreements between all the tech companies and stands to lose bigly if “AI” doesn’t pan out. I suspect this “AI” CEO is talking his own book .
Some time ago, NC linked to an article where it appeared that nowadays the top 10% of income earners account for 50% of the consumption in the USA.
My guess is that those AI evangelists are not distraught by the perspective of the economy losing a large share of its consumers, because they assume the losses will be amongst people who consume comparatively little anyway, and hence are dispensable.
In short, they are comfortable with a progression from the current situation to one where the top 10% (i.e. they themselves), with incomes boosted by the profits generated through the application of AI, represent 90% of the consumption. If this entails culling the herd, so what? The resulting economy will be sustainable and that is what counts.
I do not think it will work out like that.
if governed effectively, AI could appreciably raise aggregate economic wellbeing. Without appropriate policy intervention, however, it may reinforce structural divides – between labour and capital and between core and peripheral regions.
Human translate:
A paticular area of computing, if accepted without scructiny, might serve the present oligarchy, or it best, it will f everyone outside that golden vale.
This is a new variant of the justification for just about every policy or “innovation” like globalization or offshoring, that increases GDP but creates a lot of losers. All we need to do is redistribute the gains. Help me.
Removal of bullshit jobs is a good thing but the fallacy of composition comes into play when it’s about how/if the financial gains to capital are to be distributed to ensure the continuity of private businesses. Noone or an increasingly insufficient number of folks with the readies to pay for what you’re selling ensures that business won’t continue. One neoliberal play would be to implement a UBI to provide a flow through of public money into the pockets of private rentier extractive capital.
Private capital has for too long decided what is a paid work job (one from which it can extract a private profit). The ownership of the idea of what is to be paid work should transfer to the people via control of public money (one from which society creates beneficial capital for its survival). Under capitalism this won’t happen and while there is a continuity of pernicious and frivolous profit extracting activities.
I do not agree with your assumption. Bullshit jobs are how you learn the trade in many many white collar professions, as in the scut work that you need to do to master understanding. When I started at Goldman, I prepared spreadsheets by hand on green ledger paper, with a calculator, having to find and enter details from hard copies of SEC filings. Those in the class right after me got to use PCs and down load data from Compuserve. Their understanding of accounting and SEC disclosures was soon visibly less good than those of the elder cohorts who had done the nasty grunt work.
And it was not as if their jobs got better. They were expected to do more of different sorts of scut work instead.
Perhaps we have a difference in understanding regarding bullshit jobs.
I see your “I prepared spreadsheets by hand on green ledger paper, with a calculator, having to find and enter details from hard copies of SEC filings.” not as a bullshit job but as apprenticeship training to learn by following first principles in getting from A to B.
A step along the way so as to be able to fully comprehend the process that got from A to B as a part of learning a profession ain’t bullshit. The articled clerk to a solicitor, as it used to be, is analogous as is a trade apprenticeship. If your job had stayed at the preparing spreadsheets stage and that process could be automated then to an employer it would make sense to do that. Which is not saying that apprenticeships are unnecessary, just that as an end in itself it would be not very satisfying.
The freeing of a workforce to do re-defined meaningful and socially satisfying, meaningful public good jobs would be the goal. That would require a relegation of capitalism and as we approach end stage financialised rentier capitalism such things should be considered. It will require a revolution in thinking as it’s easier to imagine the end of the world than the end of capitalism for those habituated to TINA.
As I indicated to another poster, the sort of work I described is OVERWHELMINGLY the sort eliminated by AI. It has been widely described for over a decade how lesser knowledge tools have greatly reduced the number of entry level white collar jobs. Why do you think unemployment among college grads is so high, FFS?
That’s not a good example of a bullshit job as Graeber defined it. Hard grunt work may still be socially useful (and many socially useful jobs are that way).
David Graeber’s description of bullshit jobs is that the world would somehow be better off or really unchanged if the work weren’t done. People who labor endlessly producing corporate reports they know won’t be read or acted on, or spend many hours means-testing gov’t benefits in ways that humiliate recipients when it really is immediately clear who should get the money, and it would actually be cheaper to just give it away without means-testing in the first place.
He had lots of interviews with people who knew their jobs were really meaningless or counterproductive but needed a job so they kept doing it. Whether there was grunt work involved really wasn’t a defining quality for a bs job. And as you note, the physical labor often is a key part of the learning.
The jobs that are being eliminated by AI are the sort I described. So the person above effectively said AI was eliminating those jobs.
Moreover, producing corporate reports IS valuable. They are low stress indoor work, not mechanical, and nearly always required for regulatory or other compliance reasons. If Graeber calls that a bullshit job, he reveals he is an elitist snob who has covert contempt for ordinary people.
Yves,
I would venture to say that whoever developed and coined the central use of GDP as a metric for “wealth” was really interested in extractable wealth in gross terms, to present to a growing investor class/nations, (perhaps in the 18th and 19th centuries via colonialism), but remains a metric past those ‘times’ because it can now be used to estimate how much can also be internally taxed, or shucked from the bank accounts, and everyday transactions, of ordinary folk, (also think eternal recertification fees/requirements, and 3x medical charges/costs for less services, as 2 examples that come to mind.).
Does that help?
No, the concept was developed in the 1930s when the government started trying to define and collect economic data, in the FDR era. The investor class has no such power then. Stocks were in the toilet till the early 1950s. Wall Street firm partners (famous name bulge bracket firms) made more from military pay went they went to war in WWII than they did from their partnerships.Capitalism so discredited that the US suffered net emigration in 1935 with the USSR the top destination.
Sorry that I can’t help you; but perhaps you can ask the AI plagiarism machine for assistance?
SC declined to review whether the lower courts have correctly ruled AI generated art cannot be intellectual property.
https://www.theverge.com/policy/887678/supreme-court-ai-art-copyright
If this applies to general IP, would this make corporations reluctant to use AI?
I’ve seen concrete examples of LLM being used to help victims of domestic abuse navigate the bureaucracy around dealing with the police and others that are involved with domestic abuse (medics, local government, the law)..
I have met and talked with the person behind this:
LINDA
This is the talk he gave back in november: AI Against the Odds: Building Tools for the Forgotten
He built it off the experience of his mum facing stage 4 cancer and coming out of 30 years domestic abuse at the same time.
essentially helps you translate your experience into the language and process of the various people you have to deal with. So an aid.
I think its a great example. and very much doesnt serve the oligarchy
ones like it are few and far between though.
and you dont need “agi” for it. the current models are good enough
Researchers might disagree, e.g.:
https://dl.acm.org/doi/10.1145/3715275.3732039
Human life is the actual currency that is traded for goods and services. We each spend part of our lives working for somebody else’s benefit, and we get paid some amount so that we can purchase what we need to keep ourselves and our families alive. We are expendable and replaceable – fungible. And we wear out and are replaced. The increasing population is a means of increasing potential wealth, since more people means more purchases of goods and services from the owner class, and more profit to the owner class.
AI is a means for the owner class to further lower the number of members of the working class that it needs to pay to produce the products and services that the working class has to purchase from the owner class. Along with prior advances in production, what we get is less people needed to work, more people wanting a job, and a working population that is cannot find enough employment for all of those needing to work. Of course the owner class never lowers its prices in a meaningful way, even as production costs are lowered by production improvements including AI, so their profits increase. Moreover, their profits are never shared with the working class, so that the pool of income for the working class shrinks relative to the population of the working class. The result is increasing poverty.
Interestingly, the owner class has called for an increase in the human population. More lives means more sales, although as poverty spreads, all that really increases is misery for the working class, as less people are able to afford to pay for all the necessities. (Luxuries are only for the owner class, and a small sliver of the upper working class, a situation designed to stimulate desire in the working class for goods it can never afford – like coffee and beef and automobiles that can reliably get them to their jobs. Luxuries now include homes that the working class can afford to own.)
Of course there are calls by the working class to boycott or strike against the owner class, but those efforts have virtually no effect. That’s because the members of the working class are desperate to survive and have their children survive, of course.
But instead of strikes and boycotts and other similar actions, what would happen if the working class simply stopped having enough children to keep the population growing, and if the population were to shrink enough to ensure that there was no “excess” population to keep wages low for the owner class? What if there came to be an excess of goods and services? Is it possible that there could be a balance between the population and the amount of goods and services available to them?
While you opening statement is fair, you forget that the alternative pre-capitalism was (largely) subsistence farming, which was also “trading” labor for goods, as in your output and often done as a peasant in a feudal system, as in you did not your land.
I’m not sure I understand your comment relative to mine.
Maybe I should have clearly stated that I don’t believe it is capitalism per se that is the problem, but it is the historic emphasis on the owner’s well being relative to the well being of the workers that is the problem.
I do believe that capitalism must grapple with the finiteness of everything including space on the planet, and that growth at some point (I believe we’ve reached that point) must cease, and some contraction of the population is necessary or neither the human population nor any other living thing will survive.
Your opener, which governed your statement, was: “Human life is the actual currency that is traded for goods and services.”
Human life, as in labor has ALWAYS been “traded” for survival, if not in the modern method of exchange denominated in money terms. Humans work to live. Even in the feudal days, nobles worked, albeit far less regularly and intensely, by being expected to go into combat to protect their serfs/peasants.
There is a joke in the New Age, of monks saying “Before I was enlightened, I hauled water and cut wood. Now that I am enlightened, I haul water and cut wood.”
Reminds me of Deirdre N. McCloskey saying children are fungible … yet she presents a “behavioral” approach to economic development via Humanomics[tm] now. I mean behavioral is just a post marketing study in how successive generations have accepted orthodox grooming. Sad part is 90+% of economics for yonks had a erroneous baseline skewed by ideology and region and religion where elites pay a agenda forward.
Why some slave their minds to these optics is beyond me … look at the results …
Interesting analysis. The discussion about AI shifting income toward capital owners is something we are already seeing in many digital industries.
One small usability point: long analytical articles like this are easier to engage with when typography and font readability are optimized, especially for mobile readers. Clear fonts and spacing can make complex economic discussions more accessible to a wider audience.