The AI Layoff Trap: Game Theory, Layoffs & What to Do

Last updated: April 2026

Key Takeaways

• Researchers at UPenn and Boston University proved mathematically that AI-driven layoffs are a Prisoner's Dilemma: every company automates because the alternative is being undercut by competitors, but collective automation destroys the consumer demand all companies depend on. Neither UBI, upskilling, capital taxes, nor voluntary corporate agreements can fix it. The only solution in their model is a Pigouvian automation tax.
• The numbers are already alarming: approximately 80,000 to 99,000 tech workers have been laid off through April 2026, with nearly half of those cuts explicitly attributed to AI. Goldman Sachs warns displaced workers face longer job searches, 3%+ pay cuts, and earnings gaps that widen for a decade.
• Digital sovereignty becomes economic resilience: when income is uncertain, every recurring subscription and every dependency on rented infrastructure is a liability. Eliminating ISP equipment rental fees, building demonstrable AI skills on local hardware, and hardening your network security are concrete steps that reduce your exposure regardless of what policymakers do next.

What "The AI Layoff Trap" Actually Says

On March 2, 2026, Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University published a paper titled "The AI Layoff Trap" (arXiv:2603.20617). It has since gone viral, generating intense discussion across social media, economics blogs, and financial media. Most of that discussion has been breathless summaries of the abstract. The actual paper deserves a closer look.

The model is deliberately simple, which is a strength. Multiple symmetric firms in a sector each decide what fraction of their workforce to replace with AI. Automated tasks cost less to perform, so each firm has a direct incentive to automate. But displaced workers are also consumers. When they lose income, they stop spending. Each firm that automates captures the full cost savings from cheaper labor, but the resulting demand destruction is spread across all firms in the sector. Your competitor's layoffs reduce your revenue, and your layoffs reduce theirs.

This creates what economists call a demand externality. Each firm bears only a fraction of the damage its own automation causes. The rest falls on competitors. Because automating is always individually profitable regardless of what other firms do, it is what game theorists call a strictly dominant strategy. No amount of foresight changes the calculus. Every CEO can see the cliff ahead. None of them can afford to brake, because braking while competitors accelerate means losing market share and dying alone.

In the paper's frictionless limit, where every task is equally easy to automate, the game collapses into a textbook Prisoner's Dilemma. Every firm displaces its entire human workforce. Productivity goes to infinity. Demand goes to zero. Everyone loses.

A critical finding that most viral coverage has missed: this is not a transfer of wealth from workers to owners. It is a deadweight loss that harms both. Firm profits decline alongside worker income. The companies that fire everyone do not get richer. They get poorer, because the people they fired were also their customers and the customers of every other firm in the economy.

The paper also identifies what it calls the "Red Queen effect," named after the character in Through the Looking Glass who has to run faster and faster just to stay in place. Better AI does not mitigate the problem. It amplifies it. Each firm sees a market-share advantage in automating faster than rivals. But at equilibrium, those advantages cancel out, and only the additional demand destruction remains. More capable AI models widen the gap between what is individually rational and what is collectively optimal.

Why the Proposed Fixes Don't Work (According to the Model)

The paper systematically evaluates six policy instruments against the externality margin. The results are uncomfortable for anyone hoping market forces or popular policy proposals will self-correct the problem.

Universal Basic Income (UBI) raises the floor on living standards, which matters for human welfare. But it does not change the per-task automation incentive. Firms still capture the full cost saving from each automated task and bear only a fraction of the demand loss. The automation rate stays the same. UBI treats the symptom without touching the mechanism.

Upskilling and retraining narrow the gap between the equilibrium automation rate and the cooperative optimum, but cannot close it. Even when workers are highly skilled and quickly reabsorbed into new roles, the externality persists as long as there is any lag between displacement and reemployment. Upskilling helps workers. It does not fix the structural incentive driving firms to over-automate.

Capital income taxes operate on profit levels, not on the per-task margin where the externality actually lives. Taxing profits does not change the relative cost-benefit calculation of automating one more task. The equilibrium automation rate is unchanged.

Worker equity participation (giving workers ownership stakes) narrows the wedge slightly, because workers who hold equity benefit when the firm's stock rises from automation-driven cost cuts. But the demand externality is cross-firm. Your equity in Company A does not compensate you for the demand destruction caused by Company B's automation. The structural problem remains.

Coasian bargaining (firms voluntarily agreeing to restrain automation) fails because automation is a dominant strategy. Any agreement to slow down is not self-enforcing. The moment one firm defects, every other firm must follow or lose market share. This is the same reason cartels are inherently unstable, except worse: in a cartel, defection is profitable but risky. Here, defection is the only rational move regardless of what anyone else does.

Pigouvian automation tax: the only instrument that works in the model. A tax set equal to the uninternalized demand loss per automated task forces each firm to bear the full social cost of its automation decisions. The tax revenue funds retraining programs that accelerate worker reabsorption, which shrinks the externality over time. In theory, the tax is self-limiting: as reabsorption improves, the optimal tax rate falls.

The Numbers Behind the Theory

The paper is theoretical. The layoff data is not. Here is what the first four months of 2026 look like across major companies that have explicitly attributed workforce reductions to AI or AI-adjacent restructuring.

Sources: Programs.com AI layoffs tracker, Tom's Hardware/Nikkei Asia, InformationWeek 2026 tracker, Fortune CFO survey, individual company announcements.

The aggregate numbers are striking. According to Nikkei Asia reporting cited by Tom's Hardware, approximately 78,557 tech workers were laid off globally from January through early April 2026, with more than 76% of those cuts in the United States. Of those, 37,638 — roughly 48% — were explicitly attributed to AI and automation. A separate real-time tracker from SkillSyncer counts 146 layoff events impacting 99,283 workers as of mid-April 2026.

Goldman Sachs published an analysis warning that AI-displaced tech workers face longer job searches, pay cuts of 3% or more, and earnings gaps that widen over a decade. The labor market is currently rewarding specificity over breadth: a cybersecurity engineer who can apply AI to threat detection is in a fundamentally different market than someone with a generic AI certification competing for the same entry-level positions as thousands of other recent completers.

The Honesty Check: AI Washing Is Real

Not every layoff attributed to AI is genuinely caused by AI. Cognizant's Chief AI Officer Babak Hodjat told Nikkei Asia that AI sometimes becomes "the scapegoat from a financial perspective, like when a company hired too many, or they want to resize, and it gets blamed on AI." OpenAI's Sam Altman made a similar observation, acknowledging that "some AI washing" is occurring alongside genuine displacement.

A working paper from the National Bureau of Economic Research surveyed 750 U.S. CFOs and found that fewer than half plan AI-related job cuts. When extrapolated, that amounts to roughly 502,000 roles out of 125 million — significant, but a fraction of the doomsday projections. Fortune reported that if the study's numbers are correct, it would still represent a 9x increase over the approximately 55,000 AI-attributed layoffs reported in 2025.

Perhaps most uncomfortable for the "AI is transforming productivity" narrative: Goldman Sachs senior economist Ronnie Walker noted that "we still do not find a meaningful relationship between productivity and AI adoption at the economy-wide level." Some workers have reported that AI tools are making them less productive, with time spent on certain tasks increasing by as much as 346%. Economists have a name for this gap: Solow's paradox, coined by Nobel Laureate Robert Solow in 1987, describing the observation that transformative technology can appear ubiquitous while remaining invisible in productivity statistics.

The honest assessment: the truth is probably both. Some displacement is genuine. Some is convenient narrative. And some companies are making permanent, irreversible headcount decisions based on a technology whose productivity gains have not yet materialized at scale. The paper's model assumes AI delivers real cost savings. If companies are cutting workers based on AI's potential rather than its demonstrated performance, the economic consequences could be even worse than the theory predicts.

What This Means for You: The Sovereignty Response

ModemGuides cannot fix macroeconomic game theory. No blog can. But the paper's core insight — that structural forces are pushing toward displacement regardless of individual decisions — has a practical corollary for anyone who depends on a paycheck: reduce your exposure to recurring costs, build skills that differentiate you, and harden the infrastructure that your household depends on.

This is the same principle we apply to every topic on this site. Whoever controls the infrastructure controls the experience. When you rent your modem, your ISP controls your network. When you run AI through cloud APIs exclusively, a platform provider controls your tools and your data. When income becomes uncertain, every dependency on rented infrastructure becomes a liability that grows heavier each month.

Reduce Your Recurring Digital Cost Footprint

Start with what you are paying every month for digital infrastructure you could own outright.

ISP equipment rental: The average ISP modem or gateway rental fee is $13 to $18 per month — $156 to $216 per year — for hardware you never own. A quality DOCSIS 3.1 modem costs $80 to $150 and pays for itself in six to nine months. After that, every month is savings. If you are on Xfinity, Spectrum, Cox, or most other cable providers, you can replace your rented equipment with your own modem and eliminate that line from your bill permanently. The financial case is straightforward. The security case is even stronger: ISP-provided gateways typically include remote management protocols (TR-069), internal network scanning capabilities, and in some cases public Wi-Fi hotspots broadcasting from inside your home. When you are job searching from your home network, you want to control what your front door looks like.

Cloud storage subscriptions: If you are paying $10 to $15 per month for cloud storage, evaluate whether a one-time purchase of a local NAS or even an external drive could serve the same function for files that do not need to be constantly synced across devices. The goal is not to eliminate every cloud service — some are genuinely worth it — but to audit which recurring charges are delivering ongoing value versus which ones persist through inertia.

VPN as a sovereignty tool: A VPN prevents your ISP from building a behavioral profile of your job searching, interview preparation, and online activity. Proton VPN offers annual plans that reduce the per-month cost significantly. Mullvad VPN costs a flat 5 euros per month with no commitment, no email required, and no data logging. Neither will break a budget during a transition period, and both are investments in privacy rather than ongoing fees for access to infrastructure someone else controls.

The principle: every subscription is a recurring obligation. Infrastructure you own has a one-time cost and continues delivering value regardless of your employment status. During stable income, the difference between renting and owning is convenience. During income disruption, it is the difference between a fixed cost you have already absorbed and a monthly drain you cannot stop.

Build AI Skills on Infrastructure You Control

The Goldman Sachs displacement analysis points to a clear pattern: the labor market in 2026 is rewarding domain expertise combined with AI fluency, not AI fluency alone. Prompt engineering job postings grew 777% over 18 months. AI governance postings grew 1,257%. But these numbers describe demand for people who can apply AI within a specific field — not for people who completed a generic certification.

Local AI tools give you the ability to build real, demonstrable skills without paying for cloud API access or surrendering your work product to platform providers. Here is what a practical local AI stack looks like:

Ollama runs open-source language models on your own hardware. One command to install, one command to download a model. It provides an API endpoint on your local machine that is compatible with most AI tools, making it a drop-in replacement for cloud services in development and experimentation. If you have a machine with 16 GB of RAM, you can run capable 7B to 8B parameter models that handle everyday coding, writing, and research tasks. Our guide to the best hardware for running local AI models covers what you need at every budget level, from a Raspberry Pi to a dedicated GPU build.

Obsidian + Ollama as a local knowledge base. Andrej Karpathy's "LLM Wiki" pattern uses a local model to compile, interlink, and maintain a structured knowledge base from your own documents. You can set this up entirely on hardware you already own with zero cloud dependencies. Our step-by-step guide to building a local LLM knowledge base with Obsidian walks through the full setup. Your notes, research, and documentation stay on your machine in plain markdown files you own forever, with no vendor lock-in.

The hiring signal this creates: "I run open-source AI models on my own hardware and have built a local knowledge management system" is a demonstrably different credential than "I have used ChatGPT." It shows systems thinking, infrastructure awareness, and self-directed technical depth — exactly the domain-plus-AI combination that Goldman's data says the market is rewarding.

If you already have a local AI setup, make sure it is properly secured. Exposed API ports and misconfigured services are common vulnerabilities in home AI deployments. Our guide to securing local AI infrastructure covers isolation, hardening, and supply chain risk mitigation for Ollama, MCP servers, and automation tools.

Harden Your Digital Security (Laid-Off Workers Are Targets)

This is the section nobody else covering the AI Layoff Trap paper is writing, and it matters.

Credential theft campaigns spike during mass layoffs. Attackers know that newly unemployed workers are more likely to click on fake recruiter emails, phishing messages disguised as job offers, and fraudulent unemployment benefit portals. The emotional and financial stress of displacement makes people less cautious at exactly the moment they need to be more cautious.

At the same time, the software supply chain is under active attack. The TeamPCP campaign that compromised LiteLLM, led to the Cisco data breach through Trivy, and culminated in the axios npm supply chain attack demonstrated that trusted open-source packages can be weaponized to reach thousands of developers and organizations through a single compromised dependency. If you are a developer or engineer who has been laid off and is now doing freelance work or contributing to open-source projects from your home network, your security posture is directly relevant to your professional viability.

Concrete steps that cost little or nothing:

Pi-hole for DNS-level visibility. A Pi-hole running on a Raspberry Pi blocks known malicious domains, advertising trackers, and telemetry at the DNS level. More importantly, it gives you a query log showing every domain every device on your network is contacting. If a phishing email installs something that starts calling home, you will see it. The hardware costs under $80. The software is free.

Network segmentation. If your job search, your personal devices, and your IoT equipment all sit on the same flat network, compromising any one device gives an attacker a path to everything else. Even basic segmentation using your router's guest network feature limits the blast radius. For more control, pfSense or OpenWrt on dedicated hardware provides enterprise-grade isolation at no software cost.

Password manager and hardware security keys. AI-generated phishing is now indistinguishable from legitimate email. A unique, random password for every account (via 1Password or Bitwarden) and a hardware security key (YubiKey) as your second factor eliminate entire categories of account takeover attacks. Our digital security checklist for 2026 covers the full personal and network security stack.

The connection between security and sovereignty is direct. Every piece of infrastructure you control is a piece of infrastructure that cannot be compromised by someone else's operational failure. Running your own DNS means no third-party DNS provider sees your queries. Owning your modem means no ISP remote-management backdoor. Controlling your router firmware means you decide when patches are applied. These are not abstract principles. They are practical defensive layers that matter more, not less, during periods of economic stress.

What the Paper Gets Right, What It Misses, and What We Don't Know

What it gets right: The demand externality is real, intuitive, and underexplored in the economics literature. Most AI displacement research focuses on the labor market — what happens to workers. Falk and Tsoukalas focus on the product market — what happens to revenue when workers stop spending. The Prisoner's Dilemma framing is clean, the systematic evaluation of why popular policy proposals fail is valuable, and the finding that both workers and owners lose challenges the simplistic "capital vs. labor" framing that dominates public discussion.

What it misses: The model is deliberately simplified, and the authors acknowledge its limitations. It operates within a single sector with no new task creation — the "reinstatement effect" that has historically offset displacement since the Industrial Revolution. Acemoglu and Restrepo's work shows that new tasks and occupations have historically emerged to replace automated ones, though Autor et al. (2024) find that this replacement has slowed over the past four decades and is disproportionately failing entry-level workers. The paper also does not model international trade dynamics, sectoral differences, or the possibility that some forms of automation genuinely expand markets rather than merely cutting costs.

Most critically, the paper assumes AI automation delivers real productivity and cost advantages. The Solow's paradox evidence — that economy-wide productivity gains from AI remain statistically invisible despite massive investment — suggests that some companies may be automating based on narrative rather than measurement. If firms are cutting workers in anticipation of productivity gains that never materialize, the demand destruction happens without the offsetting cost savings that the model assumes partially cushion the blow. This is a scenario the paper does not model, and it could be worse than what the theory predicts.

What we do not know: The paper is a theoretical proof, not an empirical measurement. It proves that the externality exists and that market forces cannot self-correct it. It does not measure how large the wedge is in the real economy, because that depends on variables no one can observe in real time: how fast displaced workers find new employment, how much of their spending shifts to other sectors, and whether AI creates new categories of work fast enough to offset the displacement it causes.

The honest summary for readers: this paper does not prove that economic collapse is inevitable. It proves that the incentive structure driving AI layoffs is structurally broken in a way that no individual company and no voluntary market mechanism can fix. Policy matters. The specific policy the authors recommend — a Pigouvian automation tax — faces enormous political obstacles and has no serious legislative momentum in any major economy. Which means that for the foreseeable future, individual preparation is not a substitute for structural solutions, but it is the only lever you actually control.

Frequently Asked Questions

What is "The AI Layoff Trap" paper?

It is a game theory paper by Brett Hemenway Falk (University of Pennsylvania) and Gerry Tsoukalas (Boston University), published on March 2, 2026 and available on arXiv (paper ID 2603.20617). The paper uses a competitive task-based model to prove that AI-driven automation creates a demand externality — a Prisoner's Dilemma — where every firm rationally over-automates even though collective restraint would benefit everyone. It systematically evaluates six proposed policy solutions and finds that only a Pigouvian automation tax can correct the distortion.

Who are the authors and where was it published?

Brett Hemenway Falk is at the University of Pennsylvania's Department of Computer and Information Science. Gerry Tsoukalas is at Boston University's Questrom School of Business. The paper was posted to arXiv on March 21, 2026, and is also available via SSRN (abstract ID 6448898). It has not yet been through formal peer review, but the mathematical framework builds on well-established task-based automation models by Acemoglu and Restrepo.

Why can't companies just agree to slow down AI automation?

Because automation is a strictly dominant strategy in the paper's model. A dominant strategy means it is the best choice for each firm regardless of what other firms do. Any voluntary agreement to restrain automation is not self-enforcing — the moment one firm defects and automates faster, every other firm must follow or lose market share. This is why the authors find that Coasian bargaining (voluntary negotiation) cannot solve the problem. It is the same structural reason that cartels are inherently unstable, except here the incentive to defect is even stronger.

Does the paper say UBI would fix AI displacement?

No. The paper explicitly finds that UBI does not change the equilibrium automation rate. UBI raises the floor on living standards, which is valuable for human welfare, but it does not alter the per-task cost-benefit calculation that drives each firm's automation decisions. Firms still capture the full cost saving from each automated task and bear only a fraction of the demand loss. The externality persists with or without UBI.

What is a Pigouvian automation tax?

A Pigouvian tax is an economics concept named after economist Arthur Pigou. It is a tax levied on activities that generate negative externalities — costs imposed on third parties who are not part of the transaction. A carbon tax is the most familiar example: it forces polluters to bear the cost of the pollution they create. In this context, a Pigouvian automation tax would charge firms for the demand destruction caused by each automated task, forcing them to internalize the full social cost of displacement. The paper finds this is the only instrument that implements the cooperative optimum. The tax revenue would fund retraining, which accelerates worker reabsorption and makes the tax potentially self-limiting over time.

Am I at risk of AI-driven job loss?

The paper cites research by Eloundou et al. (2024) estimating that roughly 80% of U.S. workers hold jobs with tasks susceptible to automation by large language models. It is important to understand what this means: task exposure is not the same as job replacement. A job with 30% of its tasks exposed to AI automation is not a job that will be eliminated — but it is a job that will change, potentially significantly, in ways that affect compensation and bargaining power. The roles most concentrated in early 2026 layoffs have been customer support, operations, compliance processing, and middle management. Domain expertise combined with AI fluency appears to provide the strongest insulation, according to Goldman Sachs analysis.

What can I do right now to prepare for AI-driven economic changes?

Three concrete actions. First, audit and reduce your recurring digital costs: eliminate ISP equipment rental fees by owning your own modem, review cloud subscriptions for services you could host locally, and consolidate where possible. Second, build demonstrable AI skills on local infrastructure you control — running open-source models, building a local knowledge base, and understanding AI tooling from the infrastructure layer up. Third, harden your digital security — laid-off workers are prime phishing targets, and AI-generated attacks are now indistinguishable from legitimate communication. A password manager, hardware security key, and DNS-level monitoring via Pi-hole are low-cost, high-impact steps.