Artificial intelligence is usually framed as a productivity revolution: lower costs, faster output, new products. But a 2026 working paper by Brett Hemenway Falk (University of Pennsylvania) and Gerry Tsoukalas (Boston University) asks a more uncomfortable question — what happens if firms automate faster than displaced workers can be reabsorbed into new work?
01The problem is not irrational CEOs. It is rational incentives.
Imagine a competitive industry with several firms. Each firm can replace a human worker with an AI system. The firm making the decision receives the direct saving — lower wages or labour costs. But the worker who disappears from the payroll is also a consumer. If that income is not quickly replaced, some of the worker's spending disappears.
The crucial asymmetry is that the firm captures the private benefit of automation, while the reduction in demand is distributed across the broader market. In a simple symmetric market with N competing firms, the paper's baseline mechanism gives an automating firm the full cost saving while it bears only a fraction of the demand loss generated by the displaced worker.
02Why competition can make the problem worse
The uncomfortable logic is straightforward. If one company refuses to automate while its competitors do, it may be left with a higher cost base. Its competitors can cut prices, increase margins or gain market share. So even if every CEO understands that excessive automation could weaken aggregate demand, each individual CEO still has an incentive to automate.
The result is a prisoner's-dilemma-like outcome: restraint may be collectively better, but unilateral restraint can be commercially dangerous. The model therefore predicts an automation level above the level that would maximise collective surplus.
The Automation Arms Race
03The paper's most provocative finding: better AI can worsen the trap
In the model, more capable and cheaper automation increases the incentive to substitute machines for labour. More competition also strengthens the incentive. That produces a counterintuitive result: technological progress can increase the size of the gap between private incentives and the socially preferred amount of automation.
This does not mean better AI is economically bad. It means that productivity gains can create a policy problem if the transition destroys purchasing power faster than new work and income streams replace it.
04What about UBI, retraining and worker ownership?
The authors test several familiar responses — and one that stands apart.
Four policy responses, tested against the model
Universal basic income can raise income or demand, but it does not eliminate the firm's marginal incentive to automate. The decision to substitute machines for labour is untouched — only its downstream effect on the worker is cushioned.
Upskilling can make re-employment faster and reduce the size of the demand loss, but it does not remove the underlying externality — the firm still doesn't bear the cost of the demand it destroys.
Worker equity can redistribute some of the gains from automation, but it does not necessarily close the incentive gap between what's privately rational and what's collectively optimal.
A Pigouvian automation tax is designed to make firms internalise the demand cost associated with displacement. Within the model, this is the one lever that can align the private automation decision with the collective optimum — rather than just softening the landing after the decision is made.
05But this is not a mathematical proof that AI will destroy the economy
This distinction matters. The paper is a theoretical working paper, not empirical proof that the real-world economy is heading toward zero demand. Its results depend on the assumptions of its task-based competitive model, including how labour displacement affects purchasing power and how quickly the economy can reabsorb displaced workers.
Economic history also provides an important counterweight. Previous waves of automation have displaced particular tasks while creating new tasks, occupations and sources of demand. The central uncertainty with generative AI is whether this reinstatement process will be fast and broad enough to offset displacement.
The real risk isn't mass unemployment tomorrow. It's the erosion of the entry-level rung.
Many young workers enter professional employment through junior roles — analysts, support staff, developers, accountants, customer-service agents, research assistants and back-office professionals. If AI removes a large share of these roles, the economy could face a paradox: more demand for high-skill, AI-enabled workers, but fewer opportunities for inexperienced workers to acquire the experience needed to become those workers.
The question isn't simply how many jobs exist. It's whether the economy still provides a credible progression from education to first job to experience to higher productivity.
06The real AI debate: productivity versus purchasing power
The strongest lesson from the paper is not that AI is inherently destructive. It is that the distribution of AI's productivity gains matters.
| If productivity gains are broadly shared | If productivity gains mainly accrue to capital |
|---|---|
| Lower prices, new products, higher wages, new occupations emerge | Displaced workers lose income and new employment arrives too slowly |
| The demand problem can be mitigated | The transition becomes much harder |
That leaves policymakers with a difficult question: should policy focus only on helping workers after displacement — or should it also change the economic incentives that determine how quickly firms automate?
07Conclusion
The AI Layoff Trap is best understood as a warning about incentives, not a prophecy of economic collapse. Its provocative contribution is the idea that competitive firms can rationally automate beyond the level that would be best for the economy as a whole.
The next phase of the AI debate should therefore move beyond the simplistic question, 'Will AI take jobs?' The harder questions are: