A paper posted on arXiv by Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University argues that AI-led layoffs can push the economy toward collective over-automation. The paper says companies keep 100% of the labor-cost savings from replacing workers with AI, while the loss in consumer demand from displaced workers is spread across the broader market.
The paper, titled The AI Layoff Trap, was posted on March 21, 2026, and classified on arXiv under theoretical economics. Its core question is simple: if machines replace everyone, who is left to buy the output?
Firms capture all the savings but only a fraction of the demand damage
The model assumes a market with N identical firms, each deciding how much work to hand over to AI. For every task that gets automated, a company keeps the difference between the former wage bill and the cost of AI. But the demand lost when laid-off workers spend less is not borne fully by the company that made the cuts. That firm absorbs only 1/N of the damage, with the rest spread across competitors.
The article gives a simple illustration. In a market with 20 companies, a firm that cuts jobs bears only 5% of the resulting demand loss, while the other 95% is effectively paid by the rest of the market. The cost savings, however, stay entirely with the company making the decision.
The paper writes the size of over-automation as ℓ(1-1/N)/k, where ℓ represents the demand loss caused by each displaced worker and k is the friction cost of deploying AI. In that setup, a larger N means tougher competition and a smaller share of the demand loss borne by any one firm, which widens the gap. A lower k, meaning AI is cheaper and easier to use, widens it as well.
Automation becomes a dominant strategy in the model
Tsoukalas said, 「No matter what you do, no matter what other companies are doing, your best strategy is to adopt as much AI technology as possible. In economics, that is called a dominant strategy.」
That means the decision is not a game in which firms wait to see what rivals do before acting. Under the model, automating more remains the best move regardless of others’ choices. The paper says negotiation alone does not solve the problem. It goes on to argue that only an all-inclusive business coalition, with no firm left outside it, could reproduce the optimal outcome. If even one company stays out, that company still faces demand losses for which nobody compensates it.
UBI, retraining, profit-sharing, and capital gains taxes do not close the gap
The paper runs through several policy responses often discussed in the AI labor debate and concludes that none of them fully repair the problem.
- Capital gains tax: the authors say firms maximize after-tax profit, but the tax-rate coefficient drops out of the first-order condition, so it does not change the marginal choice of whether to automate one more position.
- UBI: a universal basic income does increase demand by putting cash in people’s hands, but in the model it changes the payoff level rather than the marginal condition where the externality sits.
- Employee profit-sharing: sharing profits with workers can narrow the gap, but fully offsetting it would require a profit-sharing ratio equal to the inverse of the worker consumption rate. The paper says the gap remains positive even if 100% of profits are distributed.
- Retraining: raising the income replacement rate after unemployment directly reduces demand loss, but eliminating the gap entirely would require a replacement rate of 1, meaning layoffs would have no effect on income at all.
According to the paper, those policies deal with the aftermath of layoffs rather than the decision point when firms choose whether to automate.
The paper points to a Pigouvian tax on automation
The only policy tool that changes the equation in the model is a Pigouvian tax on automation itself. The proposed tax rate is equal to the portion of demand loss that firms do not internalize. Tsoukalas said, 「In our model, only taxing automation itself really changes the equation.」
The paper also says the tax revenue could be used for retraining, raising the income replacement rate for displaced workers. As demand loss falls, the tax rate would fall too, meaning the tax could gradually phase itself out.
Authors say the result is theoretical, not a real-time labor market measurement
The article notes that the paper is a theoretical exercise describing an equilibrium path inside the model, not an empirical reading of current labor conditions. Its extreme end state of infinite productivity and zero demand holds only under assumptions of full labor replacement and no income replacement at all for the unemployed. The real economy has not reached that point.
The piece also notes that the authors are not simply two economists in the conventional sense. Falk’s field is cryptography and coding theory. Tsoukalas teaches information systems and is also a senior fellow at the Wharton School.
Other estimates cited in the article
The article also cites the World Economic Forum, which estimates that by 2030, 59 out of every 100 workers worldwide will need retraining, and 11 of them will receive no assistance. Goldman Sachs, by contrast, is cited as saying that fears of an AI employment apocalypse are overstated, while about 15 million people could still be displaced over the next decade.