There is a 39-page paper sitting on arXiv that has been quietly making the rounds since late March, and on Wednesday April 29 BusinessToday’s morning desk put it on the front page of every HR-Slack thread in the Fortune 500. The paper is called The AI Layoff Trap. It is by Brett Hemenway Falk at Penn and Gerry Tsoukalas at Boston University. The conclusion, restated in plain English, is the most uncomfortable thing a corporate boardroom has read in a decade.
Every CEO laying off workers in the name of AI productivity is also firing one of their company’s customers. Every one of them privately captures 100% of the cost savings. None of them privately bears more than a sliver of the demand loss, because the demand loss gets spread across every other firm’s revenue line. The math says this is a Nash equilibrium, the math says rational firms cannot stop themselves even when they understand it, and the math says only one policy lever closes the gap.
That lever is a tax.
The model, in three sentences
Falk and Tsoukalas build a task-based competitive equilibrium model. Firms choose how many tasks to automate; AI replaces a fraction of human labor at lower cost; displaced workers see a partial drop in income because the wage-replacement rate of unemployment insurance plus eventual re-employment is less than 1.0. The displaced workers are also customers — their reduced spending feeds back into aggregate demand, which is every firm’s top-line revenue.
The first-order condition for any individual CEO is that they should automate as long as the cost savings on their P&L exceed the demand loss they personally bear. But the demand loss they personally bear is their share of the macro hit — roughly the firm’s revenue share of total consumer spending. For a typical Fortune 500 company that’s a fraction of a percent. For a Big Tech company at peak, maybe a couple of percent. The other 98–99.9% of the demand loss they cause lands on competitors’ P&Ls.
Multiply this by the fact that competitors do the same calculation, and the equilibrium is what the paper bluntly calls “firms automate their way to boundless productivity and zero demand.”
Why this gets called a “trap” instead of a “race”
A race you can drop out of. A trap you cannot.
Falk and Tsoukalas show that the externality survives every standard market mechanism economists usually invoke to clean up coordination failures. They walk through them one by one and the table is brutal:
- Wage adjustments: Wages do drop as labor demand collapses, but they don’t drop enough to make humans cheaper than the AI that just got cheaper. The ratchet only goes one way.
- Free entry of new firms: New entrants automate at the same rate, so they amplify the externality rather than absorb it.
- Capital income taxes: Tax the AI investment, fine — but the externality is on the labor displacement decision, not on capital deployment. Wrong target.
- Worker equity participation: ESOPs and broad-based stock plans give displaced workers a slice of the firm-level upside, but the demand-side feedback loop is at the macro level, not the firm level. Doesn’t help.
- Universal Basic Income: Plugs the income gap of displaced workers and recovers some demand. But the firm choosing whether to automate doesn’t internalize UBI’s cost; it shows up as someone else’s tax bill. The race continues.
- Coasian bargaining: Workers can’t negotiate with the millions of firms whose individual automation decisions collectively drained the labor market.
- Reskilling / upskilling: Helps the marginal worker but does nothing to the per-task cost differential that drove the layoff in the first place.
The only mechanism that survives the model is a Pigouvian automation tax — a tax that makes each firm pay, at the moment of substitution, the actual demand loss its automation decision creates. Set correctly, it makes the private cost of automating equal to the social cost. The race ends because the prize shrinks.
The paper notes — and this is the elegant part — that the tax is self-limiting. Use the revenue to fund retraining and income replacement; the wage-replacement rate goes up; the externality shrinks; the optimal tax rate falls toward zero. The mechanism funds its own decommissioning.
The data the paper points at
The empirical anchors Falk and Tsoukalas invoke read like the LostJobs front page:
- Over 100,000 tech layoffs in 2025, AI cited in more than half.
- Over 92,000 employees laid off across 98 companies in 2026 to date.
- The 47.9% Q1 AI-attribution share Tom’s Hardware logged on April 28 is the single number that makes their model load-bearing rather than theoretical. The first-order condition the paper writes down was, three weeks ago, an academic abstraction. After Microsoft’s first-ever buyout, after Oracle’s 6 AM email, after Meta’s May-20 cut date — it is now describing what is actually happening.
The paper’s punchline, written down in section 5: “More competition and ‘better’ AI amplify the excess.” Translated for the non-economists: as AI gets better, the trap gets tighter, not looser. The release of GR00T N1.7 yesterday — open, free, scaling-law-published — is exactly the “better AI” their model says will accelerate the race.
What this contradicts
Two highly visible counter-narratives, both from this week, both from people who have to be right for a living:
- Box CEO Aaron Levie’s “tech is a contained phenomenon” argument from April 28 Fortune. Levie’s claim is that AI displacement is bounded to tech because the rest of corporate America has slow loops, regulators, and 1998 mainframes. Falk and Tsoukalas would respond: the demand externality doesn’t care which industry was first to automate. It cares whether any industry is automating faster than the wage-replacement rate, and 47.9% of Q1 cuts at AI-attribution by Q1 of any single industry is enough to start the externality cascade across the economy.
- Apollo’s Torsten Slok Jevons-paradox piece from the same Fortune Wednesday. Slok’s argument is that cheaper AI per task expands total task demand, which absorbs the displaced labor. The Jevons reading and the Falk-Tsoukalas reading are not actually in conflict at the task level — both can be true. The conflict is at the income level: even if total tasks increase, the new tasks have to also be performed by humans at high enough wage-replacement rates to keep aggregate demand intact, and the open question is whether that’s true when the marginal new task is “supervise the AI agent” at a fraction of the displaced wage.
The most charitable read is: Levie is right that the speed of cross-industry contagion will be slower than tech’s, Slok is right that some of the displaced labor finds new tasks, and Falk-Tsoukalas are right that without an explicit Pigouvian intervention the equilibrium still ends in the trap. All three positions can be simultaneously correct.
What LostJobs is watching
- Whether any G7 finance ministry runs a trial balloon on an “automation tax” within the next four quarters. France floated something similar in 2017 (rejected). South Korea has a reduced robot tax credit since 2018, which the Brookings working group calls the closest existing instrument. The Falk-Tsoukalas paper is the first time the case has been made on demand-externality grounds rather than fairness grounds. That argument is harder for a Treasury committee to dismiss.
- Whether the Hacker News thread — which, as of Wednesday morning, was the top story on HN — converts into a Tier-1 newspaper editorial within the week. Once the FT and the WSJ have written it up, the Davos crowd has cover.
- Whether the Q2 BLS unemployment-to-vacancy ratio crosses 1.0. Falk and Tsoukalas treat the wage-replacement rate as exogenous; if labor markets tighten back up, the externality shrinks naturally. If they slacken, the model says the trap intensifies. May 2 BLS prints first.
The dry coda: the second-most-shared screenshot on Wednesday was the bottom of page 12 of the paper, where Falk and Tsoukalas write — about the firms making the layoff decisions — “we show that knowing this is not enough for firms to stop it.” The paper has been on arXiv since March 2. The first SSRN post went up the same week. Eight weeks of every CFO who reads economics blogs being able to read this paper, knowing what their automation decision actually does to the macro picture, and continuing to make the same decision anyway. That is the empirical confirmation of the paper’s central claim, before you even get to the equations.