Why every company will keep replacing workers with AI — even when they know it's destroying the customers they all depend on.
Imagine a small town with seven restaurants. Every restaurant's customers are… the employees of the other six restaurants. The cooks, the servers, the dishwashers — they all eat out after work. Now someone invents a robot chef that's cheaper than hiring a human cook.
Here's what should happen: each restaurant owner thinks, "If we all replace our cooks, nobody will have money to eat out anymore. Let's be careful." The cliff ahead is visible to everyone.
This isn't a parable. It's an economic theorem. And it's what may be unfolding right now with AI.
In February 2026, Block (the company behind Square and Cash App) cut nearly half of its 10,000 employees, with CEO Jack Dorsey stating AI had made those roles unnecessary. Over 100,000 tech workers were laid off in 2025 alone, with AI cited as the primary driver in more than half the cases. Salesforce replaced 4,000 customer-support agents with AI. Goldman Sachs deployed AI coders that let one senior engineer do the work of five. Researchers estimate roughly 80% of U.S. workers hold jobs with tasks susceptible to automation by large language models.
None of this is hidden. Every CEO can see what's happening. So why can't they stop?
Before we see why the trap is inescapable, we need to understand the pieces on the board.
Think of a company as a collection of tasks — answering customer emails, writing code, managing projects, filing reports. Traditionally, each task is done by a human worker earning a wage. Now AI arrives, and each task could be done by a machine at a lower cost.
Each firm chooses an automation rate — what fraction of tasks to hand over to AI. We call this α (alpha), a number between 0 (keep all humans) and 1 (replace everyone).
Here's the crucial twist. Workers spend their wages buying things. When you fire a worker, you don't just reduce your costs — you remove a consumer from the economy. The paper captures this with a single, powerful number:
Each automated task saves the firm s = w − c (the wage minus the cheaper AI cost). But each automated task also destroys ℓ in demand across the whole sector. The question is: which force wins?
This is the paper's central discovery — and it's more subtle than you might expect.
When a firm replaces workers with AI, it destroys demand worth ℓ across the whole sector. But under competitive pricing, each firm only gets hit with ℓ/N of that loss (where N is the number of firms). The rest lands on competitors.
This is like a group of fishermen sharing a lake. Each one knows overfishing will collapse the stock. But each one also knows that the fish they don't catch will be caught by someone else. So everyone fishes as hard as they can.
The paper proves that competitive firms will always automate more than what's collectively optimal. The gap between what firms actually do and what they should do is:
Adjust the sliders to see how the gap between actual and optimal automation changes.
We normally think competition is good for consumers. But here, more competition dilutes each firm's share of the demand damage, making them care even less about the collective consequence. The most fragmented industries — think customer support, freelance services, small retailers — face the worst version of this trap.
When integration friction vanishes — when AI can do everything equally easily — the game becomes starkly simple.
Imagine AI becomes so capable that every task can be automated at the same low cost (no more "hard tasks"). Now each firm faces a binary choice: automate everything, or automate nothing. The paper shows this reduces to a classic Prisoner's Dilemma.
Example with illustrative parameters: Each cell shows per-firm profit.
| Other Firms | ||
|---|---|---|
| Keep Workers | Replace with AI | |
| Your Firm: Keep Workers |
Π₀ (best collective outcome) |
Π₀ − ℓL(N−1)/N (you lose from others' layoffs) |
| Your Firm: Replace with AI |
Π₀ + L(s − ℓ/N) (you gain while others suffer) |
Π₀ + L(s − ℓ) (everyone worse off — the trap!) |
When s < ℓ: the bottom-right cell gives less profit than the top-left. Yet replacing workers is individually rational no matter what others do. That's the trap.
This isn't a coordination failure that could be solved by "just talking to each other." Even if every CEO in the industry got together and agreed to restrain automation, each one would still be better off breaking the agreement. Automating is a dominant strategy — it's optimal regardless of what anyone else does.
The paper's most striking finding is that over-automation doesn't just hurt workers — it hurts the firms doing the automating.
This chart shows how owner profits and worker income change as the automation rate increases. Both peak before the Nash equilibrium — meaning the competitive outcome leaves everyone worse off.
Parameters: c/w = 0.30, λ = 0.5, η = 0.30, N = 7, k = 1. Both curves normalized to cooperative optimum = 1.0.
The formal result (Proposition 2 in the paper) proves that for any weight a social planner might place on workers versus owners, the competitive outcome involves too much automation. Even a planner who cared only about firm profits — zero weight on workers — would reduce automation below the market level.
The paper evaluates six proposed policy responses. The results are sobering: only one fully works.
Raising η — the fraction of displaced income that gets replaced through new jobs — directly shrinks the demand loss ℓ and narrows the wedge. Every unit increase in η toward 1 helps.
But: Historically, displaced workers suffer large, persistent earnings losses. Reaching η = 1 (full replacement) is extremely difficult, and anything less leaves a positive wedge.
UBI raises everyone's spending floor, but it doesn't change the marginal calculus of automating one more task. It changes payoff levels but not payoff differences.
Worse: Under free entry, higher profits attract new firms, fragmenting the market further — and the wedge grows with N. UBI can paradoxically widen the externality.
A proportional tax on profits (1−t)πi just scales the entire profit function. The positive scalar (1−t) cancels from the first-order condition, leaving the automation rate unchanged.
Key distinction: This is fundamentally different from a per-unit "robot tax." Taxing profits ≠ taxing automation.
Giving workers a share of profits recycles capital income back into demand. This narrows the wedge — but can never close it when workers spend less than 100% of their income in the sector (λ < 1).
And: No firm would voluntarily share profits. The marginal cost (a dollar given away) always exceeds the marginal demand benefit (λ/N of that dollar comes back). This must be mandated.
"Can't firms just negotiate?" No. Because automation is a dominant strategy, no voluntary agreement is self-enforcing. A coalition of M < N firms still externalizes (1 − M/N) of the damage.
Four barriers: Agreements aren't self-enforcing, the externality is multilateral and diffuse, automation rates aren't verifiable by rivals, and AI investments are irreversible.
A per-task tax set to τ* = ℓ(1 − 1/N) makes each firm pay for the demand it destroys on rivals. This aligns private and social incentives perfectly.
Bonus: Revenue can fund retraining (raising η), which shrinks ℓ, which reduces the needed tax over time. The tax may be self-limiting — a bridge, not a permanent fixture.
The idea is beautifully simple. Each firm already absorbs ℓ/N of the demand damage from its own automation. The tax charges it for the remaining ℓ(1 − 1/N) that falls on rivals. The firm now "feels" the full social cost of each layoff.
For each task automated, the firm pays 30% of the worker's wage as a tax.
The paper suggests two channels, each with different properties:
1. Direct transfers (wage insurance, severance) → immediately replaces lost income (raises η), but may weaken incentive to retrain.
2. Retraining programs → slower but builds durable human capital, making gains in η self-sustaining.
A good theory should survive contact with reality. The paper tests five major objections.
No — it makes it worse. This is the paper's Red Queen effect. When AI produces more output per task (not just cheaper), each firm sees a market-share incentive: "If I automate more than rivals, I capture a bigger slice of spending." But at equilibrium, every firm has automated equally, so the market-share gains cancel out. Only the extra displacement remains.
The cooperative optimum doesn't change (total revenue still depends on spending, not output), but the Nash equilibrium moves further away. "Better" AI amplifies the over-automation wedge.
Free entry reshapes the problem but doesn't solve it. In the frictionless case, three regimes emerge based on entry costs:
In the convex-cost case (k > 0), over 94% of tested parameterizations show over-automation persisting under free entry.
Endogenous wages raise the threshold but can't close the wedge. As firms automate, displaced workers flood the labor market, pushing wages down. Lower wages reduce both the cost saving and the demand loss, but the structural problem — each firm bearing only 1/N of the damage — remains untouched.
The strongest version of this argument is that wages fall so far that automation becomes unprofitable. But that's a Pyrrhic resolution: workers who keep their jobs earn barely more than machines. Purchasing power collapses through wage depression instead of displacement. The disease looks different, but the patient is just as sick.
Capital-income recycling narrows the wedge but can't close it. When owners consume a fraction η̂ of their profits in the sector, it replaces some of the lost worker spending. The effective demand loss falls to ℓη̂ = ℓ − η̂s.
Closing the wedge requires η̂ ≥ ℓ/s. When ℓ > s (the most dangerous case, where the planner would prefer zero automation), the required recycling rate exceeds 100% — owners would need to spend more than they earn. Recycling is impotent precisely where the externality is most harmful.
Richer competition (Cournot quantity competition, Bertrand pricing, differentiated products) adds new forces but doesn't eliminate the core mechanism. Whether revenue is split by competitive pricing, Cournot shares, or Bertrand undercutting, each firm still bears only a fraction of the aggregate demand destruction it causes. The wedge persists whenever a firm lacks full monopoly power.
With tasks that are complements rather than perfect substitutes, the production function itself restrains automation (removing human tasks reduces output). But even so, the demand externality remains positive as long as displaced workers lose income.
| Extension | Changes Threshold? | Effect on Wedge | Eliminates Externality? |
|---|---|---|---|
| Better AI (ϕ > 1) | No | Widens ⬆ | No |
| Endogenous entry | No | Persists (can widen) | No |
| Endogenous wages | Yes (raises) | Narrows ⬇ | No |
| Capital-income recycling | Yes (raises) | Narrows ⬇ | Partially |
| Imperfect competition | — | Persists | No |
The model predicts the problem is worst not in dominant tech firms but in fragmented industries deploying capable AI — sectors with many competitors, where each firm's share of the demand damage is tiny:
Standard economic models predict that cost-reducing technology raises profits. If we observe profit erosion coinciding with mass layoffs — firms cutting costs yet still seeing declining returns — that would be the hallmark of the demand externality.
Most policy debate focuses on what to do after displacement happens: retraining, income support, regulation. This paper reframes the question: do competitive incentives drive firms to automate beyond what is collectively optimal? The answer is yes. And no amount of after-the-fact support will change the incentives driving the arms race. Only a tax on automation itself alters the calculus.
What you can confidently explain to someone else after reading this.