The AI Layoff Trap

An interactive companion to Falk & Tsoukalas (2026)1 — why rational firms race to automate past the cliff they can plainly see.

1. The trap, in plain English

You run a company in a competitive industry, say customer support or software development. An AI tool arrives that can do much of what your human employees do. The AI costs about $18,000 per role per year; a human costs $60,000, all-in. Every role you automate saves about $42,000.

You'd have to be sentimental or bad at math to leave that on the table. So you automate. Your competitors do too.

The workers you let go don't stop needing things; they stop buying things. Some of their spending went to your products, some to your competitors'. The income they would have circulated is mostly gone. Only a fraction returns via new jobs or unemployment insurance. Your cost-saving move just shrank the market you sell into.

Monopolies avoid the trap A monopolist would feel the full sting of the shrunk market. Every dollar of demand their automation destroys would come out of their own revenue. They'd automate carefully, weighing savings against demand damage, and stop where the trade-off turned against them. The trap only exists because the market is shared.

But you're not alone. You share customers with N competitors. When you automate one role and destroy some demand, only 1/N of that damage lands on you. The rest spills onto your rivals. From any one firm's point of view, the cost savings look full-sized and the demand cost looks tiny. The same is true for every firm in the market.

Every firm reaches the same conclusion and automates. Collectively, demand falls further than any single firm would have chosen. Profits fall, worker incomes fall, and a chunk of value evaporates outright.

Everyone could see this coming. The paper's model is deliberately transparent; every firm knows exactly what's happening. Rationality and foresight don't help. Automation is a strictly dominant strategy: no matter what anyone else does, each firm is individually better off automating. So everyone does.

Nash equilibrium Game theorists' term for this kind of outcome: a state where no single firm can improve its position by unilaterally changing its mind, given what everyone else is doing. It's "stable" in the narrow sense that nobody wants to move first, even though everyone moving together would leave them all better off.
Cooperative optimum What the firms would collectively choose if they could commit to a single automation rate together: the level that maximizes combined profit, accounting for the full demand damage. The wedge between Nash equilibrium and cooperative optimum is the trap the paper studies.
The Red Queen effect The paper's counter-intuitive result: more competition makes the trap worse. The wedge grows with N. This runs against the usual story where competition disciplines firms into serving customers. Here, competition is exactly what pulls firms past the cooperative rate. Named for the Red Queen in Lewis Carroll's Through the Looking-Glass, running faster and faster just to stay in place.
Show the math

N symmetric firms each choose an automation rate $\alpha_i \in [0,1]$. Saving per role automated is $s = w - c$. Demand lost per role automated is $\ell = \lambda(1-\eta)w$, where $\lambda$ is the share of worker income spent in the sector and $\eta$ is the share of displaced wages replaced elsewhere. A quadratic friction $\tfrac{k}{2}\alpha_i^2$ captures the increasing difficulty of integrating AI task-by-task.

$$\alpha^{\mathrm{NE}} = \frac{s - \ell/N}{k},\quad \alpha^{\mathrm{CO}} = \frac{s - \ell}{k},\quad \text{wedge} = \frac{\ell(1 - 1/N)}{k}.$$

The Nash equilibrium subtracts only the firm's own share (1/N) of the demand damage; the cooperative optimum subtracts all of it. Wedge is positive whenever N > 1 and ℓ > 0.

2. Watch the mechanism

Four cupcake bakeries in a large city compete for the same customers. Each employs 50 workers (bakers, decorators, retail staff), who are themselves among the city's cupcake buyers. A new wave of automation (order kiosks, mixing robots, AI customer service) can handle about 40% of those workers' tasks at roughly a third the cost. Every firm will decide to automate; the private math is obvious. Click Next to advance one quarter at a time, and watch the lower chart: it tracks the spending these 200 workers contribute to the sector. By the end, every firm is worse off than if they'd coordinated.

Still employed Laid off even under coordination Laid off only because of the race
Four consumer-retail chains. Each employs 50 frontline workers at $60,000 a year, all-in. Those workers are also customers; they spend part of their income at retailers in this sector. Everyone's making money.
Step 0 of 5

Click Next to advance. Each step waits for you; read the caption, look at the firms and the demand bar, then continue.

10%
Starting citywide cupcake revenue: $36.00M
If they'd coordinated, ends at: $35.24M
Under the race, ends at: $34.99M
Trap's cost: $252k/yr (0.70% of citywide)

The workers' own spending drop is fixed by the model (about $1M/yr, 28% of their baseline). The slider only scales citywide context: a smaller share means a bigger city, so the same damage looks smaller on the chart.

3. Play with the model — for a single industry

Drag the sliders to set up a single industry like the consumer-retail example in §2, then watch how the Nash equilibrium pulls every firm past what would be collectively optimal. Both workers and owners lose. Section 4 widens the lens to the whole US economy.

Parameters

Market structure
4

More firms means each feels less of its own demand damage, which widens the trap.

Cost structure
$60,000

All-in yearly cost of employing one human: salary, benefits, overhead.

30% ($18,000)

AI replacement cost relative to a human doing the same work. At 30%, AI is about a third as expensive.

savings per role automated: $42,000
$60,000

How costly the last tasks are to automate. Higher values make it harder to push automation all the way.

Demand & spending
20%

Of every dollar a worker earns, how much comes back as sales in this sector. Realistic ranges: a narrow industry ~5%, a mid-sized sector (retail, consumer software) ~15-25%, broad consumer services ~30%+. Economists call this the sectoral "marginal propensity to consume."

30%

How much of a laid-off worker's wages is recovered via new jobs or transfers. The rest is gone from the economy.

demand lost per role automated: $8,400
Policy

A tax equal to the external cost an activity imposes. Carbon tax is the classic example; this is its equivalent for demand destruction. Each firm pays for the demand damage its automation causes to the other firms.

τ = $0 per role (currently OFF)

Equilibrium

Nash equilibrium αNE
—
Cooperative αCO
—
Wedge (over-automation)
—
Activation threshold N*
—

Outcomes: independent vs. coordinated decisions

Independent AI automation (each firm deciding alone) Coordinated AI automation (firms committing together)

4. The Bigger Picture

We've seen how the mechanism works inside a single industry, with four consumer-retail firms sharing a sector's demand. What happens when we widen the scope to the US economy as a whole?

A deliberately simple model The model applied at economy scale overstates cross-sector demand linkage but ignores several amplifying channels — benefits decoupling, precautionary savings among employed workers, absorption capacity constraints — that the paper treats as fixed parameters but that are endogenous to displacement rate. We treat the simple projection as a rough proxy for the combined effect.
US civilian workforce
~160M
BLS, 2024
Mean annual wage
~$65,000
BLS OEWS, 2024
Wage-driven consumer demand
~65% of wages
NIPA personal consumption

Scenario inputs

40%

Massenkoff & McCrory2 find roughly 30% of US workers have zero observed exposure (cooks, mechanics, lifeguards, dishwashers). The remaining 70% have some, with the top quartile averaging 39% task coverage. The 40% default is a midpoint for "workers in roles where the trap could bite."

20%

Of an exposed worker's role, what fraction automation actually replaces when every firm decides independently. For reference, Massenkoff & McCrory2 put top-quartile observed coverage at 38.8% today; the most exposed individual occupations, like computer programmers, sit around 74%.

15%

With foresight and coordination, how deep would automation go? Some is still productive; the cooperative level is lower than the race-to-the-bottom pace. The wedge between this and the independent path is the trap's cost.

30%

How much of a laid-off worker's wages returns via new jobs, unemployment insurance, or other transfers.

Aggregate outcomes, per year

Independent path
Coordinated path
The trap's cost
Workers displaced
—
—
—
Worker income destroyed
—
—
—
Consumer demand lost
—
—
—

In historical context

Each historical bar is that downturn's peak real PCE drop (%) applied to today's ~$19T PCE base — a uniform sizing exercise, not a direct dollar comparison.