Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University have a fascinating new paper, The AI Layoff Trap, which identifies what may become one of the central macroeconomic problems of the AI revolution: what happens if firms start replacing workers with AI faster than the economy can create new jobs and incomes for them? Their answer is unsettling. Workers are also consumers. By replacing them, firms may collectively erode the demand on which their own profits depend.

The mechanism is a classic negative externality with an unusual twist. A company replacing a worker with AI captures the entire cost saving, while bearing only a fraction of the resulting loss in consumer demand. The rest is imposed on its competitors. Each company therefore has a perfectly rational incentive to automate even when all companies would ultimately be better off if they automated less aggressively. In the extreme, this becomes a prisoners’ dilemma: everyone races to cut labour costs, only to discover that they have collectively destroyed part of their customer base.

Consider an industry with 20 similarly sized firms. If one company replaces workers with AI, it captures 100 per cent of the labour-cost saving but experiences only roughly one-twentieth of the resulting decline in demand through its own sales. The remaining 19/20 is dispersed among its competitors. From the perspective of the individual company, automation therefore looks overwhelmingly attractive.

But suppose all 20 firms make exactly the same calculation. What was negligible at the company level becomes substantial at the aggregate level. Employment and household incomes fall, consumption declines and all 20 companies confront a shrinking market. This is a textbook fallacy of composition: what is rational for one company becomes collectively destructive when everybody does it.

This produces one of the paper’s more counter-intuitive results: more competition can make the problem worse. A monopolist internalises the demand destruction caused by its own layoffs because virtually all of the lost spending eventually comes back to hurt its own revenues. In a fragmented market, each firm bears only a small fraction of that cost. Competition, normally regarded as a mechanism that disciplines firms and improves efficiency, can therefore produce an automation arms race and push the economy further away from the collective optimum.

Better AI does not necessarily solve the problem either. Higher AI productivity increases the private return from replacing workers and intensifies the race. Each firm expects to gain productivity and market share, but when all competitors adopt the same technology, the relative competitive advantage disappears. What remains is higher automation, fewer workers and weaker demand — a version of the Red Queen effect in which everyone has to run faster simply to remain in the same place.

This is where the policy simulations become particularly interesting. Retraining displaced workers works, but only if it works fast enough. If workers quickly move into new jobs and recover most of their previous income, the demand externality becomes smaller. If their income is fully restored, it disappears. If they move into better-paid jobs, the problem can even reverse. The difficulty is that historical experience suggests displaced workers frequently suffer persistent earnings losses. Retraining is therefore essential, but it may not be fast enough to deal with a rapid AI shock.

Universal basic income is less effective than its advocates might expect. It supports household incomes and aggregate demand, but does not change the marginal calculation facing a company deciding whether to replace another worker with AI. The cost saving remains. So does the incentive to automate. UBI can cushion the social consequences and sustain demand, but it does not stop the automation race itself.

A conventional tax on profits or capital income has a similar problem. It can redistribute part of the gains from capital owners to households, but it does not fundamentally change the relative profitability of substituting AI for labour. Worker ownership and profit-sharing perform somewhat better because they recycle more capital income into consumption, but they still do not eliminate the underlying externality. Nor can firms simply agree to automate less: every individual company has an incentive to defect from such an agreement and free-ride on the demand maintained by competitors that continue employing workers.

The only instrument in Falk and Tsoukalas’s model that directly corrects the distortion is a Pigouvian tax on automation. Its purpose is not to punish technology. It is analogous to carbon pricing: make the company internalise a cost that its decision imposes on others. The tax should correspond to the portion of demand destruction generated by automation that the individual firm does not itself bear. Once this external cost enters the company’s calculation, private and collective incentives become better aligned.

The optimal policy is therefore dynamic rather than a permanent “robot tax”. During a period of exceptionally rapid labour displacement, an automation tax could slow excessive substitution while its revenues finance wage insurance, income support and, above all, retraining. As displaced workers move into new jobs and recover their incomes, the demand externality declines and the tax can gradually be reduced. The objective is not to stop technological progress but to manage the speed of transition until labour markets catch up.

This is remarkably close to an argument my co-authors and I made in Tax on Robots: Whether and How Much. We argued against indiscriminately taxing technology because doing so risks discouraging investment, innovation and productivity growth — and may simply shift automation abroad. The case is much stronger for taxing the use of automation where it generates identifiable negative externalities, rather than taxing robots or AI as such.

Falk and Tsoukalas add an important macroeconomic dimension to this argument. The externality from automation is not merely lost tax revenue or greater inequality as income shifts from labour to capital. There is a third channel: aggregate demand. A company reducing employment lowers its own costs but shifts part of the resulting demand loss onto everybody else. With sufficiently rapid AI adoption, microeconomic rationality can therefore produce macroeconomic irrationality.

The policy conclusion should not therefore be “tax AI”. It should be to tax the social cost of substituting AI for labour when that substitution occurs faster than the economy can generate replacement jobs and incomes. AI that complements workers, raises productivity and creates new activities should not be penalised. But automation whose private cost saving exceeds its social benefit may justify a temporary corrective tax, with the proceeds used to accelerate workers’ return to productive employment. Once new jobs and incomes restore the lost demand, the rationale for the tax disappears.

That distinction matters. The real danger is not that AI becomes too productive. It is that firms become extraordinarily productive at the same time as the economy becomes progressively less capable of buying what they produce.