The AI Layoff Trap
Brett Hemenway Falk, Gerry Tsoukalas
econ.TH
2026-03-21
Each firm pockets the full AI cost saving but bears only 1/N of lost demand. Competitive automation can double the cooperative rate; only a per-task tax restores it.
The Acemoglu-Restrepo task framework asks whether the labor market restocks itself after automation: old tasks vanish, new ones hire people back. Falk (Penn CIS) and Tsoukalas (Boston University) move the camera to the product market. Even if reinstatement arrives later, laid-off workers are also customers. Unreplaced wages cut the spending every firm in the sector needs.
Public debate often treats the slide as a cliff with no brake. Rational firms that can see it should stop. This paper's answer is that seeing it is not enough. Each firm pockets the full cost saving from its own layoffs and absorbs only 1/N of the demand those layoffs destroy. Rivals eat the rest. That is a demand externality, not a forecasting error.
N symmetric firms each hold L task slots and choose an automation rate α in [0,1]. AI performs a task at cost c, humans at wage w, so the per-task saving is s=w−c. Later tasks are harder to integrate, captured by a quadratic friction (k/2)Lα². Workers spend a share λ of income in the sector; owners spend none in the baseline. A fraction η of displaced wages comes back through reemployment or transfers. The rest leaves the sector. Effective demand loss per automated task is therefore ℓ=λ(1−η)w.
Competitive pricing splits total expenditure evenly. Automating one more unit cuts aggregate demand by ℓ and the firm's own revenue by only ℓ/N. The private first-order condition is s−ℓ/N−kα=0; a cooperative planner faces the full ℓ. Nash automation is (s−ℓ/N)/k, the cooperative rate is (s−ℓ)/k, and the wedge is ℓ(1−1/N)/k. The gap grows with N and shrinks with friction k. A monopolist (N=1) internalizes everything; as N goes to infinity the wedge tops out at ℓ/k. The threshold N=ℓ/s: if the market is concentrated enough that ℓ/N still exceeds the saving, nobody automates.
When frictions vanish (k=0), the game is all or nothing. Once N exceeds N, full automation is strictly dominant. If s<ℓ as well, the cooperative optimum is zero automation and the equilibrium is full displacement. A firm that holds back forgoes the saving and still eats rivals' demand destruction. Talk does not change a dominant strategy, so voluntary restraint does not hold.
Under the illustrative parameters c/w=0.30, λ=0.5, η=0.30, the competitive-limit wedge equals the cooperative rate itself: firms automate at twice the profit-maximizing collective level. The same parameters with N=7, k=1 and wage normalized to 1 give Nash 0.65 against cooperative 0.35. The Nash point sits to the right of the profit peak, so both worker income and owner surplus fall short of the cooperative point. This is deadweight loss, not a transfer from workers to owners. Weighting worker welfare only pushes the planner's rate lower and widens the gap.
Policy instruments are scored on this same margin:
| Instrument | Moves the automation rate? | Closes the externality? |
| Retraining (raise η) | Shrinks the wedge | Only at η=1 |
| Universal basic income | No | No |
| Capital-income or profit tax | No | No |
| Worker equity ε | Shrinks the wedge | Not when λ<1 |
| Coasean bargaining (coalition of M<N) | Partially | No |
| Per-task Pigouvian tax τ=ℓ(1−1/N) | Implements the cooperative rate | Yes |
UBI and profit taxes shift profit levels and cancel from the first-order condition, so the layoff incentive does not move. Sharing profits with workers recycles some demand, but closing the wedge requires ε=1/λ, which exceeds 100% of profits whenever λ<1. Voluntary sharing has ε=0 as a dominant strategy. A coalition of M firms internalizes only M/N of the loss; only the grand coalition restores the cooperative rate, and automation is neither observable nor contractible across rivals.
The instrument that hits the margin is a per-task Pigouvian tax equal to the uninternalized demand loss. Revenue spent on retraining can raise η, lower ℓ, and shrink the future tax. Once AI has replaced essentially all labor, the remaining failure is distributional rather than allocative, and a profit-funded UBI takes over.
The externality survives all five extensions. If AI produces ϕ>1 units per task, a market-share motive appears: each firm thinks automating more than rivals steals expenditure. At the symmetric equilibrium those gains cancel, Nash rises, the cooperative benchmark does not, and the wedge widens. Free entry, endogenous wages, capital-income recycling, and second-stage price competition change the size, not the sign. Wages falling all the way to c can erase the wedge, but only by pushing remaining workers toward machine cost. The demand problem is then hidden, not solved.
This is an incentive map: a layoff that is rational on your own books can still undercut the demand wall the industry shares. Better AI does not self-correct; the Red Queen effect widens the gap. Waiting for UBI, retraining, or a verbal pact among rivals does not change a dominant strategy.
For policy, the paper splits how to compensate people after the fact from how to change the layoff incentive beforehand. The first supports living standards. The second is what moves allocation. A unilateral tax could push automation offshore, which is why the authors flag multilateral coordination or a border adjustment, in the spirit of carbon policy.
This is theory. The illustrative parameters are not estimates. Follow it if you work on AI labor policy. If you train models, there is no module to copy.
The model is one sector, one period, symmetric firms, by design. In a frictionless general equilibrium the lost spending may be only a relative-price reshuffle: a falling interest rate turns owners' extra saving back into investment. The demand loss is genuinely destroyed only when rates are already near zero, or when displaced workers cannot borrow through the transition. The appendix is explicit about this, and a fuller GE treatment is still in progress.
There is no empirical estimate of the wedge. U.S. employers announced more than a million cuts in 2025, with AI named in about 55,000 of them, and Block cut nearly half of its 10,000-person workforce in February 2026. Those figures are scene-setting, not an identification of ℓ. The authors say that if reabsorption keeps pace, the externality may be too small to detect. The paper names a structural vulnerability, not a measured crisis.
Automation rates sit inside the firm. A tax authority would approximate them from procurement and payroll. Welfare loss is quadratic in the wedge, so a roughly right tax still has a first-order gain. That defense is fair; the paper does not quantify how noisy the measurement would be. Endogenizing AI research could also tilt investment further toward labor-replacing systems.