Over the past year there has been a significant wave of layoffs linked to the rise of AI. Major companies such as Amazon, Oracle, and Meta are among those affected. According to The Guardian, Microsoft cut around 15,000 jobs last year. Amazon reportedly laid off 30,000 employees over the past six months, while Meta reduced its workforce by more than 1,000 during the same period. More recently, Oracle, a major software company, has also reportedly laid off thousands of employees.
Researchers Brett Hemenway Falk from the University of Pennsylvania and Gerry Tsoukalas from Boston University have identified an “automation trap” using a mathematical model. Each individual company benefits from replacing workers with AI due to lower costs, but collectively, firms lose because overall demand declines. The researchers note that CEOs are aware of this dynamic – yet continue down the same path.
In their paper, Hemenway and Tsoukalas explain that workers are not just labor – they are also consumers. When they are laid off, their income falls, and so does demand. As demand declines, company revenues shrink as well.
“Even if reinstatement eventually occurs, a problem arises along the way: displaced workers are also consumers, and when their lost income is not replaced, each round of layoffs erodes the purchasing power all firms depend on. At the limit, this becomes self-destructive: firms automate their way to boundless productivity and zero demand.”
The authors argue that even when firms understand the consequences, each company captures the full benefit of cost-cutting but bears only a fraction of the damage caused by declining aggregate demand – with the rest spilling over onto competitors. They describe automation as a dominant strategy, where firms feel compelled to adopt AI to avoid losing ground.
However, the researchers point out that when all tasks become equally easy to automate, the situation turns into a classic Prisoner’s Dilemma where every firm replaces workers with AI, even though collective restraint would increase profits for all.
The resulting losses are not a redistribution from workers to firm owners, but a net loss of welfare that harms both sides:
“The resulting surplus loss is not a transfer from workers to firm owners; it is a deadweight loss that harms both.”
The Model behind the findings
The researchers develop a task-based automation model inspired by the work of Daron Acemoglu and Pascual Restrepo, but shift the focus from the labor market to the product market.
The mechanism works as follows:
“…When automation displaces workers, their forgone spending reduces every firm’s revenue. Each of several symmetric firms chooses what fraction of its workforce to replace with AI. Automated tasks are performed at lower cost, but integration frictions make each successive task harder to automate. On the demand side, workers spend a fraction of their income on the sector’s output; firm owners spend less, normalized to zero in the baseline. Some displaced wage income is recovered through reemployment or transfers, but the remainder is lost to the sector.”
In the baseline version of the model, wages are held fixed and capital-income recycling is shut down. The authors note that these assumptions are relaxed in extended versions of the model. Despite its simplicity, the framework allows for the analysis of a wide range of policy tools and robustness checks.
Their results show that the optimal level of automation for an individual firm is higher than what is optimal for the economy as a whole. In other words, firms’ decisions diverge from the social optimum, leading to systematic over-automation.
What solutions do they propose?
American researchers evaluated several policy tools that could potentially address this problem. Among them were measures such as upskilling and worker participation in capital ownership. They concluded that while these approaches can reduce the gap, they do not eliminate it. Coasian bargaining also fails – the researchers explain this by noting that, due to automation being a dominant strategy, voluntary agreements are not self-enforcing. Taxes on capital income do not change the level of automation either, since they affect profits rather than the marginal decision. The same applies to universal basic income – it raises the minimum standard of living but does not alter the incentives to automate.
The only effective solution identified by the researchers is a Pigouvian tax on automation, set equal to the unaccounted demand loss per automated task. This policy can implement the cooperative optimum.
According to the authors, tax revenues could be used to fund retraining programs, increasing income recovery and gradually reducing the externality over time.
What is a Pigouvian tax?
A Pigouvian tax is an economic instrument designed to offset the harm caused to society or the environment by business activity. The concept was introduced by British economist Arthur Cecil Pigou in The Economics of Welfare (1920), where he defined it as a per-unit tax equal to the marginal external cost, aligning private incentives with social costs.
The researchers argue that this approach is particularly well-suited to AI-driven automation:
“In contrast to many textbook externalities, where the harmed parties are outside the firms’ market (e.g., pollution), here the harmed parties are workers whose income constitutes the firms’ own demand. This means the tax rate, its revenue, and its incidence all interact through the same labor-market channel, creating richer policy design questions than the standard case.”
At the core of the paper is a demand-side externality – firms account only for their private gains from automation while ignoring the broader damage to demand. By taxing automation, policymakers effectively make it more expensive – internalizing the harm it creates.
As a result, a Pigouvian tax aligns private and social incentives, restoring market equilibrium and leading to a socially optimal level of automation.
The authors also compare their findings with prior research, highlighting key differences in mechanisms and assumptions. The full paper, including formal modeling and mathematical derivations, is available on arXiv.