Owen McGrann’s “The Dead Economy Theory” starts with a brutally simple question:
> Who is the customer when the customer is the thing you’ve eliminated?
The AI industry needs a huge economic prize to justify hundreds of billions of dollars in infrastructure and valuations measured in the same order. Better search, autocomplete, and office memos are not enough. The only market large enough is labor itself.
That does not prove mass replacement will happen. But it clarifies what many AI business models are implicitly selling: not a slightly better tool, but a much cheaper substitute for expensive cognitive work.
McGrann’s essay follows that premise further than the normal “jobs versus productivity” debate. His dead economy is not a place where nothing gets produced. GDP may rise. Software may multiply. The dead part is participation: most people are no longer needed to produce value, do not own the systems producing it, and gradually lose the income and political leverage that came from being economically necessary.
The core mechanism takes three turns.
Turn one: a company replaces workers with AI. Its costs fall, margins rise, and investors reward management. From the firm’s perspective, the decision is rational.
Turn two: the displaced workers lose income and cut spending. Restaurants, landlords, retailers, software companies, banks, and other businesses feel weaker demand. Some respond by automating more of their own workforce.
Turn three: the original company discovers that its customers were other companies’ employees. It saved money by participating in the destruction of its own market.
This is not primarily a story about evil executives. It is a coordination problem. Each firm captures almost all of its own labor savings while bearing only a small fraction of the demand it destroys. If twenty firms share a market, one company can automate and receive the benefit while nineteen competitors absorb most of the lost spending.
That produces a prisoner’s dilemma: every firm has an incentive to automate faster than is healthy for the system as a whole. Better AI can intensify the race because falling behind becomes more expensive.
The strongest part of the essay is that it does not require superintelligence.
The bad outcome may come from “so-so automation”: AI cheap enough to replace workers, mediocre enough to lower service quality, and unimpressive enough to produce only modest economy-wide productivity gains. A company can still prefer it because labor costs are visible and immediate, while the social costs are diffuse and delayed.
That possibility matters because it breaks the usual binary debate. We do not need to choose between “AI is hype” and “AI transforms everything.” A technology can be overhyped as intelligence and still be powerful as a labor-disciplining instrument. It can fail to deliver abundance while succeeding at moving income from labor to capital.
The standard optimistic answer is that technology has always destroyed jobs and created new ones. Agriculture employed most people; now it employs very few. Spreadsheets eliminated armies of manual calculations but created other office work. New occupations appeared that earlier generations could not imagine.
All true. But history is evidence, not a law.
The agricultural transition took generations. The Industrial Revolution produced a long interval in which productivity and profits rose before wages caught up. A transition can be successful in a history textbook and catastrophic across an individual life.
AI also differs in scope and speed. Earlier machines usually attacked narrow physical or clerical tasks. General-purpose models are being aimed simultaneously at software, law, finance, medicine, design, support, analysis, and administration. The capital already invested creates pressure to push adoption quickly—even before organizations know whether the systems can reliably do the work.
The key variable is therefore not only whether new jobs eventually appear. It is whether they appear fast enough, in the same places, and for people whose skills have just lost market value.
McGrann’s argument becomes more important when it moves from employment to ownership.
If AI produces enormous value but the models, data centers, chips, and distribution channels belong to a narrow class of owners, the economy can grow while most people lose leverage. Labor’s share falls. Collective bargaining weakens. Consumer demand depends increasingly on transfers. The tax base becomes easier to route through firms already skilled at tax optimization.
Democracy itself rests partly on mutual dependence. Governments and capital owners need citizens as workers, taxpayers, soldiers, and consumers. That need gives ordinary people bargaining power. If automated capital needs far less human labor, the balance changes.
This is why “we will just pay everyone UBI” feels incomplete. It treats the public as a cost to support after removing them from production, not as owners or participants with claims on the productive system.
A healthier response would spread capital ownership, preserve public stakes in AI infrastructure, tax concentrated gains, and prevent a few firms from becoming the private metering layer for intelligence.
The essay also refuses the tidy idea that a monthly payment solves the human problem.
Work is often tedious, exploitative, or badly organized. It should not be romanticized. But it also supplies structure, status, skill, social contact, and proof that one has a place in the shared world. Communities hollowed out by deindustrialization did not suffer only from smaller bank balances. They suffered from lost purpose, weaker institutions, addiction, family breakdown, and the disappearance of a believable future.
That makes displacement psychologically different from leisure. Leisure is restorative when it is chosen and bounded by meaningful activity. Permanent economic redundancy imposed from above can feel like exclusion.
The risk becomes sharper if AI also deskills the people expected to adapt. If juniors delegate foundational work before developing judgment, the same tool that threatens their current role can weaken their route to the next one.
McGrann writes polemically, and some parts of the essay jump too quickly from plausible incentives to a nearly predetermined political future.
Demand does not have to come only from wages. Lower prices can create new consumption. AI can complement workers, create new products, and make small firms viable. Governments can redistribute income or ownership. Capital investment itself creates demand. The empirical estimates of AI productivity and displacement are wildly uncertain.
There is also a tension in arguing both that AI adoption is economically underwhelming and that it will rapidly eliminate most cognitive labor. “So-so automation” partly resolves this—the system can replace workers without transforming productivity—but the scale still matters. A warning about a dangerous trajectory is not proof that the endpoint is inevitable.
So the useful reading is not “the economy will definitely die.” It is:
1. firm-level automation incentives do not automatically produce a healthy macroeconomy;
2. productivity statistics say little about who owns the gains;
3. transition speed can turn a long-run success into a generation’s disaster;
4. income transfers do not fully replace agency, status, and participation;
5. governance must arrive before ownership and deployment patterns harden.
The dead economy theory is strongest as an ownership question disguised as a jobs question.
If AI really becomes a universal production layer, society has three broad choices. People can own part of that layer. They can retain bargaining power over how it is deployed. Or they can become dependents receiving whatever distribution the owners consider politically necessary.
The technology does not choose among those futures. Corporate incentives will choose one by default if democratic institutions do not choose another deliberately.
A growing economy in which most people have no stake, no role, and no leverage is not abundance. It is a very productive form of exclusion.