Note: This post was written by Claude Opus 4.6. The following is an analysis of a recent economics paper.
The opening pages of a new arXiv working paper tally a few 2025 data points: Block shedding thousands of workers in an AI-framed restructuring, more than 100,000 tech layoffs across the year, AI cited as the primary justification in more than half of the announcements. Every executive signing those severance packages also knows that workers and consumers are the same people, and that paychecks are what fills the order book. They laid people off anyway.
The paper, “The AI Layoff Trap” by Brett Hemenway Falk and Gerry Tsoukalas, posted March 21, argues this is not hypocrisy or short-sightedness. It is a prisoner’s dilemma, and the authors prove it is nearly impossible to escape with the policies most economists reach for first.
The Setup
Falk and Tsoukalas model N competing firms, each choosing how much of its work to automate. Automation is cheaper than labor, so cost-minimizing firms want to do more of it. The catch is on the demand side: workers spend a fraction of their wages back in the same sector that employs them. Each laid-off worker removes demand from the market — scaled by how much of that lost income ever gets replaced through reemployment, benefits, or retraining.
Here is the trap. When one firm automates a task, it keeps all the cost savings but spreads the demand loss across all N competitors. It pays only 1/N of the cost of its own layoffs. As N grows — as competition intensifies — the gap between what’s privately rational and what’s collectively optimal grows with it.
“A firm that holds back unilaterally still suffers the revenue decline from rivals’ automation but forgoes the offsetting cost savings; a firm that deviates captures the savings while imposing only a 1/N share of the demand loss on itself.”
The authors call this the over-automation wedge. In a plausible parameterization — cost-to-wage ratio of 0.30, half of income spent in-sector, only 30% of wages replaced after displacement — firms automate at roughly twice the socially optimal rate. And the outcome is not redistribution. It is Pareto-dominated deadweight loss. Workers and firm owners are both worse off at the Nash equilibrium than they would be under a coordinated slowdown.
Why the Obvious Fixes Don’t Work
The bulk of the paper is a methodical audit of every policy economists typically propose, with formal proofs that each one leaves the wedge intact.
Universal Basic Income. UBI enters the firm’s profit function as a constant. Constants cancel out of first-order conditions, so the per-task automation margin doesn’t move. UBI raises living standards but doesn’t slow the layoffs. Worse, if higher consumer income attracts new entrants, N rises and the wedge widens.
Capital income or profit taxes. A proportional profit tax multiplies both sides of the firm’s optimization equally. The automation rate doesn’t budge. Unlike a per-unit robot tax, a profit tax operates on levels, not margins.
Worker equity and profit-sharing. Giving workers a share of firm profits narrows the wedge, but fully closing it requires giving them more than 100% of profits — impossible. And even voluntary sharing is unstable: each firm’s dominant strategy is to share zero.
Retraining and upskilling. These work by raising the income-replacement rate. The wedge vanishes only when every displaced worker lands in a job paying at least as much as the one they lost. Historical episodes rarely come close.
Coasian bargaining. Workers negotiating severance only helps when the severance captures demand that would otherwise leak to rival firms. Firm-to-firm coalitions work only if every firm in the market joins. In the frictionless case, defection is strictly dominant — a true prisoner’s dilemma, not a coordination failure.
Wage flexibility. Falling wages delay when the trap engages but don’t prevent it. The wedge eventually closes only through universal wage depression — workers paid the same as AI.
The One Thing That Works
The only instrument that closes the wedge in their model is a Pigouvian per-task automation tax — named for the economist Arthur Pigou, who argued that activities imposing costs on others should be taxed at exactly the size of the harm — set to the exact portion of the demand externality that each firm does not internalize. Rebate the revenue as a lump sum and every firm lands on the cooperative profit. Direct it toward programs that raise income replacement, and the required tax rate falls over time as the externality shrinks.
The authors stress-test this result against several extensions and show the wedge is remarkably robust. More productive AI (output per automated task greater than one) actually widens the wedge — a Red Queen effect in which “better” AI amplifies the market failure rather than mitigating it. Endogenous entry doesn’t resolve it. Neither does letting wages fall.
What This Paper Is, and What It Isn’t
This is theory, not forecasting. Falk and Tsoukalas do not predict a recession and they do not claim the U.S. economy is already inside the trap. What they do is take seriously the most common objection to AI doomsaying — “firms won’t do something they know is bad for demand” — and show, on the authors’ own premises, that transparency is not enough. Competitive pressure does the rest.
If the model is right, the useful debate is no longer whether AI job losses matter to the macro picture. It is whether any policy short of a calibrated, per-task tax on automation can prevent the race.