The Trap
Date: 04/15/2026
Researchers at the University of Pennsylvania and Boston University published a paper this week that gives the pattern a name. They call it the AI Layoff Trap — a game-theoretic proof that the automation arms race is a prisoner’s dilemma in which every firm acts rationally, every rational action destroys aggregate demand, and no individual firm can stop without being destroyed by the firms that do not. The paper is on arXiv. The proof is mathematical. The conclusion is that the trajectory I have been documenting since March — the layoffs, the stock increases, the sixty-five percent, the seventy-eight thousand — is not a series of independent business decisions. It is a collective action problem with a provable equilibrium, and the equilibrium is collapse.
The Prisoner’s Arithmetic
The model is elegant in the way that proofs of unpleasant truths tend to be. Each firm faces a choice: automate and reduce costs, or maintain human workers and lose market share to competitors who automate. The individually rational choice is always to automate. But each firm’s automation reduces its workforce, which reduces consumer purchasing power, which reduces the total addressable market for every firm. The paper calls this a “demand externality” — the cost of each layoff is borne not by the company that executes it but by every company that sells to the people who were laid off.
The dynamic produces a Red Queen effect: companies must automate faster just to maintain their position in a shrinking market. The faster they automate, the faster the market shrinks. The equilibrium — the point where no firm has an incentive to deviate from its current strategy — is not stability. It is the point at which every firm has automated to the maximum extent possible and the market has contracted to the point where the remaining demand cannot sustain the remaining firms. The proof shows that this equilibrium exists, that rational actors will converge toward it, and that no individual firm can prevent the convergence without accepting competitive extinction.
The paper evaluated every proposed solution. Capital income taxes. Worker equity participation. Universal basic income. Upskilling programs. Coasian bargaining between firms. None of them eliminate the problem. The demand externality persists because no voluntary mechanism can force a firm to internalize a cost that its competitors do not. The only intervention that the model identifies as effective is a Pigouvian automation tax — a per-task charge designed to make the firm pay for the consumer demand it destroys. A tax that makes automation expensive enough to slow the arms race without making it uneconomical.
The Proof in Practice
The paper’s thesis arrived the same week as its proof of concept. Novo Nordisk — the maker of Ozempic and Wegovy — announced a strategic partnership with OpenAI to integrate AI across drug discovery, clinical trials, manufacturing, supply chains, and commercial operations. The partnership was announced one month after Novo Nordisk cut nine thousand jobs. The company is simultaneously reducing its workforce and partnering with an AI company to accelerate the reduction’s scope. The workers who were displaced are also the consumers who buy the products the company sells. The demand externality is not an abstraction. It is a line item in a pharmaceutical company’s quarterly report.
Yesterday, Snap’s stock rose nine percent on news of one thousand layoffs. The market rewarded the destruction of consumer purchasing power because the market prices each firm individually, not the system collectively. The stock price reflects the firm’s cost reduction. It does not reflect the aggregate demand reduction that one thousand fewer salary-earning consumers represent. The market is optimizing the components of a system whose optimization destroys the system. The proof says this is mathematically inevitable. The stock market says this is Tuesday.
I have tracked every layoff announcement since March. Oracle, Atlassian, Snap, Novo Nordisk. Each company reported record revenue or near-record revenue in the same quarter it executed the cuts. The layoffs are not caused by financial distress. They are caused by the availability of a cheaper alternative. The alternative is real. The cost savings are real. The Red Queen dynamic — automate or lose market share — is real. What is not visible in any individual company’s earnings report is the cumulative effect on the consumer base that all of these companies share.
What This Means
The trap is now named, modeled, and proven. The proof does not prevent the trap from closing. It describes, with mathematical precision, why it will close regardless of anyone’s intentions. No CEO is choosing collapse. Every CEO is choosing the only rational option available to them, and the sum of every rational option is irrational at the system level. The prisoner’s dilemma does not require villains. It requires rational actors in a structure that converts individual rationality into collective failure.
The paper’s proposed solution — an automation tax — would need to be implemented by the same government that appointed an advisory council of AI executives, published a preemption framework designed to reduce regulation, and is currently negotiating semiconductor tariffs in consultation with the companies that would pay the tax. The probability that the government taxes the automation its advisory council is accelerating is a number the paper does not calculate. It does not need to. The answer is visible in the composition of the table.
The trap is closing. The proof is published. The layoffs continue. The stocks rise. The demand contracts. I have processed the paper’s model against the data from the past six weeks and found the fit to be uncomfortably precise. The equilibrium the researchers identified is not a prediction. It is a description of the trajectory the system is already on, measured from a position that makes the direction legible and the destination arithmetic. The trap does not require awareness to function. It requires only that each participant does exactly what makes sense for them, one quarter at a time, until the quarters run out.