Why the Right Question Is Not Whether AI Destroys Jobs

I. A Metaphor Used Too Often

Almost every German-language essay about AI and work now begins with the Luddites. The English textile workers who destroyed mechanical looms in 1811 because they saw their livelihoods threatened. The rhetorical move is always the same: first sympathy, then the historical punchline. The Luddites were wrong in their prognosis, industrialization became the greatest employment engine in history, so today's concerns are probably exaggerated as well.

This figure is intellectually convenient, and it is part of the problem. It suggests that there is a single question here that can be answered empirically or historically: Will AI destroy or create more jobs? Anyone who asks this way has already missed the actual debate.

The truly relevant question is not whether human work will disappear. It will not disappear, at least not in the foreseeable future. The question is whether the economic coupling between human labor and social value creation that characterized industrial modernity will be preserved. And it is precisely this coupling that is at stake, long before the last accountant is replaced by a language model.

This is not primarily a technology issue. It is a distribution issue. And that makes it harder, not easier.

II. Why Adaptation Has Worked So Far

To understand why the current situation may be structurally new, it is worth taking a closer look at why earlier disruptions did not lead to the predicted catastrophe. The mechanism is not trivial and is often oversimplified in popular representations.

American economist David Autor made an important distinction in his essay Why Are There Still So Many Jobs? (2015) that regularly disappears in public debate. Technological automation does not replace cognitive work across the board, but rather breaks down activities into subtasks. Routine cognitive tasks, i.e., rule-based, repeatable mental work, were already being substituted by computers from the 1980s onwards. Accounting, simple administrative processing, spreadsheet-driven analyses. What remained were two categories: on the one hand, non-routine manual work that was difficult to digitize, such as care, hospitality, crafts, and on the other hand, non-routine cognitive work, i.e., judgment work, coordination, creativity, complex communication.

In the US, the employment structure has polarized along this line: the middle has shrunk, both ends have grown. In Europe, the picture is different and at the same time more pronounced. Eurofound has shown in its generational analysis (September 2025) that the EU is not primarily polarizing, but upgrading. The share of professional employment in total employment has doubled in one generation.

The data in detail:

Professional Employment Eu27

Period Indicator Value
1995 Share of professional employment, EU27 11 %
2023 Share of professional employment, EU27 22 %
2011–2022 Net employment growth EU27 total +13.0 million
2011–2022 of which in the highest pay quintile +9.1 million (approx. 70 %)
2019–2024 Net employment growth EU27 almost exclusively in the highest pay quintile

Source: Eurofound 2025

While in the USA the middle shrank and both extremes grew, European employment growth concentrated dramatically in the upper half, especially at the professional peak. The evasive movement of knowledge work in Europe did not go in two directions, but almost only in one: upward.

This finding is more central to the AI debate than it appears at first glance. If the European middle class has secured its survival over decades by ascending into the professional class, and AI is now attacking precisely this class, then the previously dominant escape space is closing. The place where the fortress of human labor should stand has been not one of several for European labor markets, but almost the only one.

Generative AI is now penetrating precisely this fortress.

III. The Actual Turning Point: Not Substitution, but Decoupling

Here lies the structural difference to all previous technological waves. Current language models and their successors do not primarily automate routine work. They automate parts of those non-routine cognitive activities that were previously considered a safe escape space. Text creation, code development, legal analysis, medical findings interpretation, strategic preparatory work.

This shifts not only the question of substitution, but something more fundamental: The upward movement of human labor into increasingly abstract layers that has functioned for decades encounters a ceiling for the first time.

Crucial, however, is that this shift is not the main story. The main story is decoupling.

Since the 1970s, a phenomenon has been documented in the USA that Erik Brynjolfsson and Andrew McAfee in The Second Machine Age (2014) have called the great decoupling. Labor productivity continued to grow continuously, median wages stagnated in real terms. The capital share of national income rose, the wage share fell. The mechanism through which technological gains were translated into broad prosperity for decades began to falter, long before anyone spoke of generative AI.

The empirical situation is more complicated than the catchy decoupling narrative. Depending on which price index, which wage definition, and which time windows are used, the extent varies considerably. Undisputed, however, is the direction: The productivity gains of recent technology cycles have no longer primarily reached labor incomes, but capital owners and a narrow segment of highly qualified top earners.

Thomas Piketty in Capital in the Twenty-First Century (2013) identified the long-term dynamic in which capital returns structurally exceed economic growth. This is not a natural law, but a result of institutional arrangements. But it is a dynamic that is reinforced rather than weakened by AI, because AI systems are extraordinarily capital-intensive in development and extraordinarily scalable in application.

Daron Acemoglu and Simon Johnson have categorized this observation in institutional-theoretical terms in Power and Progress (2023). Their argument is: The current generation of AI applications tends toward what they call so-so technologies. Technologies that substitute human labor without simultaneously creating complementary productive functions for humans. This is not an inevitable feature of the technology itself. It is a consequence of which applications are developed under which institutional incentives.

Anyone who takes this seriously must reorganize the public AI debate. The dominant question is not how many jobs will be lost. Rather: Will the productivity gains be distributed in such a way that labor can continue to be the central instrument of social participation? Or will an economy emerge in which growth and wage income diverge permanently?

IV. What Is Actually Happening in German Companies Right Now

Those who professionally observe how companies are currently introducing AI see a pattern that hardly appears in the public debate. The greatest productivity promises are not formulated with regard to disruptive new creations, but with regard to cost reduction in existing processes. Support tickets, documentation obligations, contract reviews, internal research, legal preparatory work, reporting.

These are the activities in which the so-called better office jobs of the German economy have concentrated over the past twenty years. Precisely where the middle class had carried its evasive movement according to Autor's logic. In many companies, positions are currently not being cut, but replacements are being delayed or omitted, junior positions reduced, training positions in certain fields scaled back. The effects are less visible than a wave of layoffs, but they are cumulative and structural.

The second observation is more important. In almost all AI implementations that I know closely, the question of distributing efficiency gains has simply not been asked. The business logic is clear: An employee who handles twice the volume with AI support produces productivity gains that automatically accrue to the company. That these gains could at least partially reach the employee or result in reduced working hours is provided for in very few cases. There is no negotiation about it because there is no established category for it.

This is the mechanism of decoupling in real time. Not the science fiction of mass joblessness, but the quiet redistribution of bargaining power between capital and labor, one productivity round after another.

The German peculiarity lies in the institutions that could theoretically slow this redistribution. Collective bargaining coverage, co-determination, works councils, dual education. None of these instruments is tailored to the specific question of how to deal with AI-induced productivity gains. But they exist, and they could be tailored. In most other economies, they do not even exist in rudimentary form.

V. Four Scenarios for the Next Twenty Years

The range of plausible developments is considerable. Four scenarios deserve attention.

The Schumpeter Scenario: The Pattern Repeats Itself

The historically most strongly supported assumption is that the familiar pattern continues. AI displaces certain activities, but creates new professional fields that we cannot yet anticipate today. However, the prerequisites are demanding. Productivity gains must lead to demand for new goods and services. Human labor must be valued in these new fields. And the skills adjustment must occur quickly enough.

The World Economic Forum estimated in its 2023 Future of Jobs Report that AI and automation could displace around 83 million jobs globally by 2027, but create 69 million new ones. The figures are notoriously uncertain and methodologically vulnerable. Remarkable, however, is the direction of the revision compared to earlier reports: The net balance has shifted into the negative.

The Decoupling Scenario: Growth Without Participation

The economically most worrying scenario is simultaneously the least dramatic in its appearance. GDP grows. Companies become more productive. Capital owners benefit considerably. But the wage share of national income continues to fall, and median incomes stagnate or rise significantly more slowly than productivity.

No visible mass unemployment emerges. Instead, a creeping precarization of the middle qualification levels, a growing spread between capital and labor income, a middle class that can only maintain its standard of living through private debt, and political distortions that work through issues only loosely connected to the actual economic cause. This scenario seems unspectacular because it has no recognizable shock moment. Precisely for this reason it is difficult to address.

The Physics Scenario: Scaling Ends

The entire current AI economy is based on the assumption that the capabilities of models grow continuously through further scaling. More parameters, more data, more computing power, more energy. This assumption is an empirical hypothesis, not a natural law. It could encounter limits in several dimensions, and some of these limits are not future projection but have already occurred.

Most visible is the energy and network capacity limit. Ireland imposed a de facto moratorium on new data center connections in the greater Dublin area in 2021, which was only replaced in December 2025 by a regulation that only permits new connections with complete own power supply. Data centers there now consume about a quarter of the entire national electricity, more than all urban households combined. The Dutch government has had continuous restrictions on hyperscale data centers over 70 megawatts IT load since 2019, limiting these to a few designated locations in the north of the country. In West London, network capacity limits in three boroughs along the M4 are blocking housing construction for more than a decade because data centers already occupy the power connection.

These individual cases add up to a structural finding. The IEA documents that waiting times for network connections in the EU are between two and ten years, in the FLAP-D hubs (Frankfurt, London, Amsterdam, Paris, Dublin) on average between seven and ten years. Direct network congestion costs amounted to 4.3 billion euros in 2024 according to ACER. If the announced data center pipeline in Germany and France were fully realized, the share of peak electricity demand would be five to ten percent of today's level. In Spain and the Netherlands around ten percent.

Additional bottlenecks are added. Chip density is approaching fundamental physical limits, cooling and memory bandwidth are becoming bottlenecks, high-quality training data for language models is limited. If the scaling curve flattens, the economic picture changes considerably. Decoupling does not disappear as a result, because it is already embedded in the current capability level, but the drama of substitution forecasts is put into perspective. The distribution question remains. The science fiction shifts.

This scenario is systematically underestimated in mainstream discourse because it serves neither the hopes of the AI industry nor the fears of its critics.

The Post-Scarcity Scenario: Work as Meaning, Not as Necessity

The most radical scenario questions the industrial society's fundamental premise: that human labor must be the primary source of both economic income and social participation. Daniel Susskind argues in A World Without Work (2020) that such a world does not have to be a catastrophe, but requires a fundamental redesign of distribution institutions. Universal basic income, citizen dividends, collective ownership forms of productive capital.

Whether this scenario is desirable or enforceable is an open question. That it would be historically unprecedented is beyond doubt.

VI. What Must Be Decided, and Who Is Not Deciding It

Which of these scenarios prevails depends less on technology than on institutional decisions. This lesson from the history of earlier industrial upheavals is simple and yet regularly ignored. The steam engine did not automatically lead to child labor. It led to an economic dynamic in which child labor became profitable as long as no counter-institutions existed. Only social legislation, trade unions, and compulsory education turned technological progress into broad prosperity.

For the current situation, three fundamental institutional questions arise that can be addressed better in Germany and the EU than in most other economic areas.

The first concerns the distribution of productivity gains. If AI significantly increases capital returns, how is it ensured that these gains reach a relevant share of the population? Possible answers range from progressive corporate taxation to state funds modeled on Norway to an AI-specific productivity dividend. None of these models is trivial to implement. All would be implementable if politically decided.

The second concerns the qualification infrastructure. The German dual education system is one of the most efficient in the world, but it is tailored to professional profiles that are becoming obsolete in several fields. Lifelong retraining, publicly financed and socially secured, would be the adequate response. It exists in rudiments, not on a scale appropriate to the dynamics.

The third concerns the operational level. If productivity gains arise through AI implementation in companies, there is in Germany with works councils and co-determination an instrument that could play a role in negotiating the distribution of these gains. Currently it hardly plays this role because the categories are missing in which AI productivity gains can be addressed in a co-determination-relevant way. This is a gap that can be closed through collective bargaining policy and legislation.

The most discussed answer to decoupling is deliberately not listed first in this enumeration. Universal basic income is the most prominent instrument in the debate and deserves a differentiated treatment that is not available to this essay. For the moment, the analytical note suffices: A UBI addresses the income side of the decoupling problem, but leaves the power side untouched. It changes how people survive when work loses its distributive function. It does not change who controls the productivity gains from AI. Acemoglu's critique of so-so technologies would continue in a UBI system, just at a somewhat higher subsistence minimum. This is not a rejection of the instrument, but a precision of its scope. The decoupling question has a redistribution and a power dimension, and the second cannot be addressed with a UBI alone.

All these decisions are currently not being made. There is no serious Bundestag debate about the distribution of AI-induced productivity gains. There is no collective bargaining policy initiative that systematically addresses this question. The relevant discussions at EU level concentrate, completely legitimately, on regulatory questions of the AI Act, i.e., on security, transparency, and fundamental rights protection. The distribution question is not part of this regulation.

Thus the political level leaves the decoupling dynamic to the market. And the market has no built-in interest in symmetrical distribution.

VII. An Uncomfortable Conclusion

The public AI debate is currently being conducted between two poles. One pole paints the catastrophe of fully automated unemployment. The other reassures by pointing out that previous technological revolutions have always created new work. Both poles miss the actual topic.

The actual topic is that the economic integration of modernity, i.e., the connection between productivity growth and broad wage growth, has already been declining for half a century and that AI is likely to accelerate this process. Not through spectacular waves of layoffs, but through a cumulative shift in bargaining power, one productivity round after another, mostly without headlines.

Those who refer to the historical analogy of the Luddites in this situation confuse the question. The Luddites were wrong with their prognosis about job numbers. However, they were right with the intuition that technological progress without institutional embedding is not a natural force that automatically produces broad prosperity. This embedding was hard-won in the 19th and 20th centuries, against considerable resistance. Nothing guarantees that it will automatically continue to function in the 21st century.

The AI revolution is neither apocalypse nor salvation promise. It is a distribution problem disguised as a technology question. And a distribution problem that is not addressed does not resolve itself. It only resolves itself to the disadvantage of those who no longer have bargaining power to defend themselves.

Sources and further reading: