Demand externalities, institutional choice and the AI layoff trap.
What the textbooks leave out
The history of capitalism is written by its winners, which is why the Industrial Revolution appears in most textbooks as a story of triumph. Steam, steel and the assembly line lifted much of humanity out of agrarian subsistence into an age of mass prosperity. What the textbooks tend to omit is what happened to the people in between — those who lived through the transition rather than reading about it afterwards.
At its peak in the mid-1820s, Britain's handloom weaving industry employed roughly four hundred thousand workers across the United Kingdom. By 1860, that number had collapsed to a fraction of its former size. The great majority of a profession — and with it a way of life, a household economy and a local social fabric — was annihilated within a single lifetime. At the start of the nineteenth century, roughly a third of the British workforce still drew its livelihood from agriculture; by the early twentieth century, that share had fallen below ten per cent. The great migration from field to factory was not a gentle optimisation. It was a rupture, and its costs were borne overwhelmingly by the people least able to absorb them.
What we now remember as a natural arc of progress was, at the time, a civilisational emergency. The response was not automatic. It took roughly a century of strikes, riots, parliamentary commissions, revolutions and ideological convulsions — Chartism, socialism, Bismarckian social insurance, the French Front Populaire, the New Deal — before the new industrial order stabilised into something a modern worker would recognise as humane. The eight-hour day and the five-day workweek, often imagined as natural entitlements, were in fact social inventions. Henry Ford, who announced the five-dollar, eight-hour day at his Highland Park plant in January 1914 and the five-day workweek in 1926, did not do so out of generosity. He did so because he understood something the classical economists had missed: that the men who built the Model T also had to be able to afford to drive one. Without leisure, there is no consumption. Without consumption, there is no mass market. Without a mass market, the efficiencies of the assembly line become a curse instead of a blessing.
This is the lesson that contemporary discussions of artificial intelligence keep rediscovering, usually too late. We are living through a technological transformation that future textbooks will almost certainly compare to the steam engine and the dynamo. As before, the narrative threatens to skip directly from problem to resolution without dwelling on the decades of dislocation in between. The question is whether we can do better this time — not because we are smarter or kinder, but because we can read the history our ancestors were busy making.
What economic historians know
It is a curious feature of the economics discipline that its deepest empirical insights tend to come not from the theorists but from the historians among them. Ronald Coase's theory of transaction costs, which eventually matured into the New Institutional Economics of Oliver Williamson and Elinor Ostrom, was grounded in careful archival work about how firms actually emerged in industrialising Britain. Ostrom, who shared the 2009 Nobel Prize with Williamson, built her entire framework for governing common-pool resources on decades of fieldwork in Swiss alpine villages, Japanese fisheries and Spanish irrigation communes. And in 2024, the Nobel Prize in Economics was awarded to Daron Acemoglu, Simon Johnson and James Robinson for a body of work that is, at its heart, a historical argument: that the long-run prosperity of nations is determined not by geography or culture but by institutions, and that institutions are made and unmade by political choices under duress.
This is the intellectual tradition into which any serious discussion of AI has to be placed. The question is not whether artificial intelligence will transform the economy — it will, and the transformation is already under way — but under what institutional configuration the transformation will occur. The Industrial Revolution produced both the factory and the trade union, both the railway and the antitrust statute, both the chemical industry and the Clean Air Act. The outcome was not technologically determined. It was politically constructed, usually in the aftermath of crises that institutional foresight might have prevented.
The romantic reading of that period — a heroic engineering achievement that eventually made everyone better off — is true only if one averages across a century and across generations. Inside any given decade of the nineteenth century, the distributional story was brutal. Wages for unskilled labour in Manchester, Leeds, Lille, Mulhouse and the Ruhr stagnated or declined through the first half of the 1800s even as industrial output multiplied. Life expectancy in the industrial cities fell. Child labour spread. Skilled artisans — the most capable and productive workers of their era — were systematically impoverished as their craft knowledge was embedded first in machines and then in process manuals written by men who had never touched a loom. The story that the Industrial Revolution was good for humanity is retrospective. The story that it was terrible for most of the humans who lived through it is contemporaneous. Both are true. Economic historians know this. Macroeconomic models, trained on post-war time series, tend to forget it.
The trap
A recent theoretical paper by Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University, The AI Layoff Trap, sharpens the present problem into its purest analytical form. Using a competitive task-based model of firm behaviour, the authors show that rational, profit-maximising firms face a dominant strategy to automate human labour out of existence, even when every firm in the economy understands that doing so will erode the consumer demand on which their own revenues depend.
The mechanism is a demand externality. When a single firm replaces a worker with AI, it captures the full productivity gain while distributing the income loss — and therefore the demand loss — across the entire economy. From the firm's point of view, automation is privately profitable even when it is collectively disastrous. The logic is the same as a classic tragedy of the commons: each actor, optimising locally, helps destroy the very resource on which all of them depend. Here the commons is aggregate demand, and the resource being depleted is the purchasing power of the working population.
What makes the Falk–Tsoukalas result especially uncomfortable is the list of remedies it rules out. More market competition, which economists typically prescribe for inefficiency, actually amplifies the excess: when firms compete harder on cost, they automate faster, and the downward pressure on demand intensifies. Free entry does not solve the problem, because entrants face the same individual incentives as incumbents. Wage adjustments cannot restore equilibrium once the marginal product of labour has fallen below subsistence; the market simply clears at zero hours. Capital income taxes, which redistribute rents after the fact, fail to alter the behaviour that produced the distortion in the first place — they merely redistribute a shrinking pie. Worker equity participation, a favourite of progressive commentators, does not work either: giving employees shares in the firm that is automating them does not change the firm's dominant strategy. Universal basic income relieves the symptom but leaves the underlying competitive dynamic untouched, and indeed makes it cheaper for firms to displace workers, because the political cost of lay-offs falls. Upskilling — the perennial consolation offered to displaced workers — presumes that the set of human-exclusive tasks is not itself shrinking, which is precisely the assumption at stake in an AI transition.
Coasian bargaining, the standard invocation of property-rights economics, collapses for reasons that Oliver Williamson made canonical in his refinement of Coase. It is not enough to say the externality is diffuse; the deeper frictions are three. Bounded rationality prevents firms from enumerating the relevant externalities across a continuously changing task space. Opportunism ensures that those who would benefit from defecting on any voluntary automation pact have every incentive to do so. And the asset specificity of the tasks at stake is asymmetric in a way that dooms private negotiation: the workers being displaced are often specific to their firms, but the demand loss their displacement creates is diffused across the entire consuming economy, and there is no party with standing to bargain on behalf of the latter.
What remains, in the Falk–Tsoukalas model, is a single instrument: a Pigouvian tax on automation itself, set to internalise the demand externality and bring firm-level incentives back into alignment with collective welfare. One need not accept every modelling choice to take the structural point seriously. Falk and Tsoukalas are not the first to suggest an automation tax — Bill Gates floated the idea publicly in 2017, and Daron Acemoglu has sketched versions of it for years — but they are among the first to show, within a tractable microfoundation, that most of the alternatives currently on the public menu do not actually fix the problem they are supposed to fix. They address aftermath, not cause.
There is, however, a further wrinkle that the symmetric-firm assumption of the model quietly conceals. AI is not a monolithic technology but a stack — foundation models, inference infrastructure, deployment tooling, vertical applications, proprietary data, distribution networks. David Teece argued in his 1986 paper Profiting from Technological Innovation that in weak appropriability regimes, the innovator rarely captures the value; owners of complementary assets do. Applied to the present stack, the question becomes: who actually banks the automation rents? If foundation-model providers operate as bottleneck monopolies while deploying firms compete the surplus away, the demand-erosion dynamic intensifies, because both the displaced workers and the operating firms lose, and a concentrated upstream layer consumes little of what it extracts. If, conversely, deployment know-how and proprietary data are the binding complementary assets, the rent distribution is flatter and the dystopian scenario milder. The essay will not pretend to settle this empirically. It is noted here because any policy response must contend with it: a Pigouvian tax levied on the wrong margin, or in ignorance of where rents actually accrue, will fail to internalise the externality it is meant to address.
Work is not only income
There is a register in which the Falk–Tsoukalas framing is already too narrow. The paper describes the cost of displacement as lost aggregate demand, which is an accurate description of the macroeconomic mechanism but an incomplete description of the human stake. Work, in every society that has organised itself around it, is more than a source of income. It is a site of biographical integration and intersubjective recognition: the context in which ordinary people come to understand themselves as competent members of a community, acquire roles that others acknowledge and locate themselves in a narrative of contribution. Jürgen Habermas's distinction between system and lifeworld captures the point with precision. The displacement under discussion operates on both registers simultaneously. When workers are told that their tasks can be done by a machine, what collapses is not only their purchasing power but also the social architecture through which they understand who they are.
The distinction matters politically as well as analytically, because it opens onto a failure mode that a purely economic model cannot see. If work is merely an income source, its loss can in principle be compensated with a transfer payment — universal basic income, for example. If work is also a site of identity formation and social integration, transfer payments solve only half the problem, and a population that receives the cheque but loses the role is not thereby reconciled to the new order. This is the territory Habermas mapped in Legitimationsprobleme im Spätkapitalismus (1973): when an economic system cannot deliver the mass employment on which its political legitimacy tacitly depends, a legitimation deficit opens that does not close by itself. The empirical signature of such a deficit is familiar — eroding turnout, institutional distrust, the rise of movements that offer identity where the economy no longer offers roles — and its relationship to the coming AI transition is direct.
Three futures
One can sketch the space of outcomes along two dimensions: whether firms use AI primarily to substitute capital for labour or to expand the frontier of what labour can accomplish, and whether institutions respond early enough to shape the competitive equilibrium or only after the damage has been done. Four combinations are logically possible, but only three are analytically interesting.
The first is the dystopian scenario Keynes warned about in 1930, under a different name. If firms automate aggressively while institutions lag, the economy enters a deflationary spiral with a negative feedback loop. Workers lose employment and income. Households, uncertain about the future, increase precautionary saving and reduce consumption. Aggregate demand contracts. Firms see their revenues fall and respond by cutting costs further — usually by automating more, because labour is the easiest line item to eliminate. Prices drift downward, real debt burdens rise, investment stalls and the economy slides into what Irving Fisher called debt-deflation in 1933 and what Japan has been living through, in slow motion, since 1991. The paradox of thrift — that behaviour rational for the individual household becomes destructive when practised by all households simultaneously — reappears in a new register as the paradox of automation. Every participant is playing the game correctly. The game itself is misaligned.
The second trajectory is less often discussed in economic commentary but, historically, more frequent. It is the legitimation path, in which the macroeconomic aggregates do not necessarily collapse but the political and social fabric does. A sufficient share of the population is displaced visibly and quickly enough that the losers cease to consent to the order that displaces them. Turnout erodes among the disaffected. Anti-system movements capture their attention, offering identity and meaning where the economy no longer offers roles. Institutional trust falls below the level needed to enact the policy responses that might otherwise cushion the transition. The historical signatures are familiar: the 1930s on one register, the 2010s resurgence of authoritarian and nativist politics across Europe and North America on another. This trajectory does not require a deflationary spiral. It requires only that the loss of work be concentrated, visible and institutionally unaddressed. It can arrive even if GDP looks acceptable — indeed, precisely because elites read the aggregates and miss the distribution.
The two trajectories are not mutually exclusive; a moderate economic shock can still produce a severe political reaction if it is felt as concentrated and unjust.
The third trajectory depends on a different set of choices. If firms use AI primarily to expand the scope and quality of what workers can do, rather than to replace them, the productivity gain is shared, demand is preserved, and a positive feedback loop emerges: better products, faster time-to-market, rising wages, expanding consumption and further innovation. Acemoglu and Pascual Restrepo have called this the reinstatement effect — the way technological change historically creates new tasks even as it destroys old ones — and they have emphasised that reinstatement is not automatic. It happened in the twentieth century because institutions, unions and public investment aligned the incentives towards human-complementary rather than human-substituting technologies. The computer revolution of the 1980s and 1990s, for all its disruption, produced as many new occupations as it eliminated, in part because complementary investments in education and public infrastructure kept the returns to labour rising alongside the returns to capital. It could fail to happen in the twenty-first century if those institutional alignments are absent or dismantled.
Which of these three paths is followed is not a matter of prediction but of choice. And the choice is being made now, inside firms and ministries and parliaments that largely do not realise they are making it.
Polanyi's double movement
Karl Polanyi, writing in 1944 at the close of the most destructive decade in human history, offered a framework for understanding why market societies periodically convulse. His central concept was the fictitious commodity: labour, land and money are treated by the market as though they were produced for sale, but they are not. Labour in particular is inseparable from the human beings who perform it, and a social order that pretends otherwise sets itself on a collision course with the lives it is organising. Every expansion of unregulated market logic, he argued, eventually provokes a political counter-movement aimed at protecting society from the market's consequences. The counter-movements have sometimes been constructive — Bismarck's social insurance laws, British labour law, the regulatory apparatus of the American New Deal, the post-war European social market economy — and sometimes catastrophic — fascism, autarky, revolutionary violence, ethnic nationalism. The form the counter-movement takes depends, to a considerable degree, on what institutional options are available when the pressure becomes unbearable. A society that has pre-built the shock absorbers tends to produce Bismarck and Roosevelt. A society that has not tends to produce something worse.
It is worth recalling what the unmanaged versions have actually looked like. The Luddites of 1811 to 1816 — not the credulous machine-smashers of romantic imagination but the skilled Nottinghamshire weavers and Yorkshire croppers whose entire craft was being obliterated by the power loom — did not smash machines out of stupidity. They smashed them because Parliament had repealed the labour protections that had previously moderated technological change, and because the Napoleonic Wars had pushed food prices to famine levels while mill owners captured the surplus. They were not wrong about what was happening to them, only about what could stop it. A generation later, the Manchester model of industrial capitalism produced the conditions that Friedrich Engels catalogued in The Condition of the Working Class in England (1845) — twelve-hour days, child labour, unsanitary housing, routine industrial accidents — and those conditions produced, within three years, the Communist Manifesto. Within seventy years, they had organised the Russian Revolution, the Chinese Revolution and a global ideological contest that ordered international politics for most of the twentieth century. The question the communists posed — what happens to a society that treats human beings as an input to be minimised — was not the wrong question. The answer they proposed was, in most places where it was tried, worse than the problem. But the problem was real, and where it went unaddressed, it produced answers worse still.
This is the analytical frame within which the AI transition needs to be placed. It is not enough to describe the technology, or even to describe the economic mechanism by which the technology produces unemployment. One has to ask, additionally, what kind of counter-movement is being prepared. The institutional architecture of the mid-twentieth-century welfare state was not built in the 1930s. It was built in the decades before, during the Progressive Era, the German social reforms of the 1880s, the British Liberal reforms of 1906 through 1911, the municipal socialism of fin-de-siècle Europe. By the time the Great Depression arrived, the conceptual toolkit was already available. The same will be true of the AI transition: the instruments adopted during the eventual crisis will be the instruments that were built, theorised and trialled in the years before it. The present decade is therefore not a time for waiting. It is the time when the toolkit for the next forty years is assembled.
The institutional question
If the Falk–Tsoukalas model is correct that firm-level incentives under AI drive towards an outcome bad for firms themselves, the practical question is what institutional configuration can restore the alignment. The menu is larger than the current policy discourse suggests, and continental Europe has already trialled more of it than the Anglo-American debate typically acknowledges.
A Pigouvian automation tax is one lever, and a serious one: it operates on the margin at which the distortion is actually produced, rather than on the downstream consequences. The administrative design is not trivial — defining what counts as automation, distinguishing productivity-enhancing from labour-displacing uses of AI, avoiding capital-flight effects in open economies — but none of these problems is categorically different from the problems that already attend corporate or carbon taxation. A decade of carbon-pricing design experience transfers directly.
A second lever is the institutionalisation of voice inside the firm. German co-determination (Mitbestimmung), codified in the Betriebsverfassungsgesetz of 1952 and substantially extended by the Mitbestimmungsgesetz of 1976, gives works councils information and consultation rights over significant operational changes — including, in practice, the introduction of new technologies that alter workflows, workloads or employment levels (§§ 90 and 91 BetrVG). In firms with more than two thousand employees, workers hold near-parity representation on the supervisory board. Williamson's New Institutional Economics would classify this as a hybrid governance mode, sitting between market contract and hierarchical command, and the point is precisely that it operates endogenously: it shapes the firm's automation decisions before a regulator has to. From a Habermasian angle, it is also a deliberative institution, giving those affected by automation a procedural claim to participate in the decision rather than merely to be compensated for its consequences. Either reading suggests that the present debate's neglect of co-determination is an analytical loss.
A third lever is the preservation of the employer–employee relationship through structural transitions. Germany's Kurzarbeitergeld, with roots in the 1910 Kali-Gesetz and formalised in the 1920s, replaces a portion of workers' lost wages when firms reduce hours instead of headcount. It was deployed at unprecedented scale during the 2008 financial crisis (around 1.4 million workers) and the 2020 pandemic (peaking near 6.7 million in May). Italy's Cassa Integrazione Guadagni, dating from 1941 and substantially expanded after the war, performs the analogous function, replacing fifty to eighty per cent of lost earnings for workers idled by economic or technical disruption. Both institutions were designed for cyclical shocks, but their structural logic — distributing the cost of transition across the public and the firm rather than concentrating it on the displaced worker — is precisely the logic an AI transition requires. They are, in the Falk–Tsoukalas frame, partial substitutes for the automation tax, because they maintain household demand and employer-employee matches through the adjustment window.
A fourth lever is working-time reduction, the instrument Ford invented in 1926 and Europe has repeatedly updated. The French Front Populaire, under Léon Blum, legislated the forty-hour week and two weeks of paid holiday in June 1936, decades before the idea became normal on the other side of the Channel. The measure was later partially rolled back under the Third Republic's final governments, a useful reminder that institutional gains are reversible; but it established the template, and the French Loi Aubry of 1998–2000, reducing the statutory week to thirty-five hours, was its direct descendant. The contemporary four-day-week experiments under way in Iceland, the United Kingdom, Portugal and Spain belong to the same lineage. In the Falk–Tsoukalas framework, passing productivity gains through partly as leisure rather than entirely as employment reduction distributes the dividend in a way that maintains aggregate demand without requiring an explicit redistribution scheme.
A fifth lever is the pre-emptive construction of welfare and skills infrastructure. Bismarck's compulsory health insurance (1883), accident insurance (1884), and old-age and invalidity pensions (1889) were not enacted from generosity; they were enacted, as Bismarck told the Reichstag, to defang socialist agitation amid the industrial dislocation of the 1870s and 1880s. They worked: the most conservative German chancellor of the century constructed the first modern welfare state and, in doing so, pre-built the shock absorbers that carried Germany through later transitions. The American parallel is the Morrill Act of 1862, which funded the land-grant universities; the GI Bill of 1944, which underwrote a generation of technically skilled veterans; and the National Defense Education Act of 1958, which shaped the post-war technological frontier in pro-worker directions. In each case, public institutions shaped the composition of investment in ways that private capital, left to itself, would not have chosen. The research priorities of the National Science Foundation, the European Research Council and their equivalents are therefore not side-issues in the AI debate. They are levers of first-order importance.
A sixth lever, visible only in Europe at the time of writing, is pre-emptive regulation of the technology itself. The EU AI Act, which entered into force on 1 August 2024, with the bulk of its high-risk obligations applicable from August 2026, is the first comprehensive risk-based legal framework for AI enacted anywhere in the world. Its high-risk category explicitly includes AI systems used in employment and in the management of workers. Whatever one thinks of its particular provisions, the Act represents a categorical break from the reactive pattern of the nineteenth century: an institutional pre-build legislated before, rather than after, mass displacement. The German and French industrial models were constructed mostly after the disruptions they were responding to. The EU AI Act is an attempt to construct the institution ahead of the disruption. It will succeed or fail in proportion to the other five levers.
None of these instruments is sufficient on its own. All of them are easier to enact before the crisis than during it.
At least one major European economy has recently illustrated the reverse case. The German Hartz reforms of 2003 to 2005 flexibilised the labour market, deregulated temporary work and slashed long-term unemployment benefits; they did reduce headline unemployment, but they also expanded the low-wage sector and weakened the very shock absorbers the post-war German model had been built around. The lesson is not that flexibility is always wrong; it is that institutional architecture is load-bearing, and that dismantling it in ordinary times exposes the society that did so to risks it has lost the capacity to manage.
This is the essential insight an economic historian would bring to the present moment: the institutional choices that look most radical in ordinary times look, in retrospect, like the minimum conditions for avoiding catastrophe. The eight-hour day was dismissed as utopian in 1866 and taken for granted in 1936. The welfare state was denounced as socialist adventurism in the 1880s and accepted as the foundation of political stability in the 1950s. Universal healthcare was a fringe proposal in 1910 and a settled norm in most of the industrial world by 1970. The counter-movements happened. The question was always only whether they happened in time, and at what cost.
The decision being made now
The AI transition is not an event that will happen to us. It is a sequence of decisions being made, right now, inside firms, governments, investment committees and regulatory agencies. Each of those decisions is being made by people who are individually rational, institutionally constrained and largely unaware that they are collectively writing the first chapter of an economic history that their grandchildren will read.
The Falk–Tsoukalas paper is important not because its specific policy recommendation will or will not be adopted but because it clarifies the structure of the problem. The problem is not that firms are greedy, or that workers are unprepared, or that technology is moving too fast. The problem is that the competitive equilibrium under current institutions points towards an outcome that harms firm owners and workers alike. The way out is not through individual virtue or better training programmes. The way out is through the deliberate construction of institutions that change the pay-offs firms face when they make the marginal decision to automate one more task — and through the deliberate construction of procedures, deliberative and democratic, by which the people most affected have a standing claim to participate in the decision.
This is what every successful economy did during the Industrial Revolution, eventually, and at terrible cost. The question posed by the current moment is whether the cost has to be paid again, in full, or whether a generation that can read the history might choose, for once, to act on it.
The decision is not the machine's. It is ours — if we can still answer to that pronoun before we join the ghosts of earlier transitions.
Sources
Acemoglu, D., & Johnson, S. (2023). Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs.
Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2).
Coase, R. H. (1937). The Nature of the Firm. Economica, 4(16).
Falk, B. H., & Tsoukalas, G. (2026). The AI Layoff Trap. arXiv:2603.20617 [econ.TH].
Fisher, I. (1933). The Debt-Deflation Theory of Great Depressions. Econometrica, 1(4), 337–357.
Habermas, J. (1973). Legitimationsprobleme im Spätkapitalismus. Suhrkamp.
Habermas, J. (1981). Theorie des kommunikativen Handelns. Suhrkamp.
Keynes, J. M. (1930). Economic Possibilities for our Grandchildren. The Nation and Athenaeum, October 1930; later in Essays in Persuasion (1931).
Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.
Polanyi, K. (1944). The Great Transformation: The Political and Economic Origins of Our Time.
Regulation (EU) 2024/1689 of the European Parliament and of the Council (EU AI Act), OJ L 2024/1689, 12 July 2024.
Teece, D. J. (1986). Profiting from Technological Innovation. Research Policy, 15(6).
Thompson, E. P. (1963). The Making of the English Working Class. Victor Gollancz.
Williamson, O. E. (1985). The Economic Institutions of Capitalism. Free Press.
Nobel Prize in Economic Sciences 2024: Daron Acemoglu, Simon Johnson, James A. Robinson — "for studies of how institutions are formed and affect prosperity."