The Great AI Implosion

The Familiar Pattern

In September 2002, with the telecommunications industry still breaking apart around him, Paul Starr published an essay in The American Prospect that diagnosed the disaster as something stranger than fraud.¹ By the time Starr wrote, half a million telecom workers had lost their jobs, the Dow Jones communication technology index had fallen eighty-six percent, and WorldCom had filed what was then the largest bankruptcy in American history.² The surface narrative was criminal: capitalized operating expenses, sham capacity swaps, the specific conduct of specific executives at WorldCom and Global Crossing and Adelphia.³ Starr's essay argued that the surface was a distraction. What had failed was not a handful of corrupt firms. What had failed was a theory.

The theory was that telecommunications could be deregulated into a sustained multi-firm competitive market. Starr traced the history the Telecommunications Act of 1996 had ignored:⁴ the telegraph's explosion in the 1840s and its consolidation into Western Union's monopoly by 1866; the telephone's thirty-year competitive phase after Bell's patents expired in 1894, and its consolidation into AT&T's regulated monopoly by 1913. The pattern was stable across the two cases, and the reasons were economic, not political. Telecom networks carry high fixed costs, vanishing marginal costs, and commodity output. Firms competing in such a market must race for scale. Competition becomes a race to the bottom on price. The losers go bankrupt or are acquired. Consolidation is not a failure of competition. It is what competition produces.

Starr was right about what came next. AT&T, which he thought unlikely to survive as an independent company, was absorbed by SBC in 2005. Lucent merged with Alcatel in 2006.⁵ The long-distance business collapsed into a footnote. Circuit-switched voice telephony — the technology the industry had been founded on — was dissolved, as he predicted, by the gale of voice-over-IP. The FCC under Michael Powell permitted the re-consolidation Starr warned against. The public-interest protections he argued for — common carriage, rate-of-return discipline, universal service — became the subject of the net-neutrality battles Starr's essay had named a decade before they were fought.

What Starr's essay offers, beyond a specific diagnosis of a specific collapse, is a pattern. Any industry with high fixed costs, near-zero marginal costs, commodity output, and enormous economies of scale consolidates. Deregulatory frameworks that refuse to accept this produce boom-bust cycles that end in monopoly. The telegraph did it. The telephone did it. Telecom did it. And as of 2024, a new industry fits the pattern with uncanny precision.

Training a single frontier artificial-intelligence model costs several hundred million dollars in compute.⁶ The marginal cost of serving an additional inference request is GPU-seconds and electricity. Output — generated text, completed tasks, answered queries — commoditizes visibly each quarter; price per token for frontier-capability models has fallen roughly tenfold per year since 2022.⁷ Where one provider offered a given capability in mid-2023, five or six offer it in 2026. Nvidia's recent SEC filings explicitly flag a risk of circular revenue arrangements in which a hyperscaler invests capital in an AI developer, the developer spends most of that capital on the hyperscaler's compute, and both sides book as revenue what is essentially recycled equity⁸ — the same pattern Starr identified in the telecom firms of 2001 that swapped unused fiber capacity with each other and recorded both sides as sales.⁹

The economic structure, in other words, is identical. The buildout is larger in absolute terms. The consolidation outcome is predictable. Within a few years, some fraction of the current frontier AI developers will be absorbed into the hyperscalers at distressed prices, and the surviving firms will look less like the competitive startup ecosystem of 2024 and more like the regulated oligopolies of pre-1984 telecom. A handful of current entrants will become the new incumbents. The rest will be acquired or liquidated. The workers they employed will be absorbed into a labor market that does not yet understand what is happening to it. This part of the story is familiar. The familiar part is not what this essay is about.

The unfamiliar part is that the labor consequences of the coming AI implosion will not follow the telecom pattern. They will follow an older and more radical one, never before reached in an industrial-era economy. And they will raise — already are raising, in venues less visible than the public industry conversation — a question the industry has not yet been willing to name out loud.

The Difference

Every prior wave of labor-saving technology eliminated some jobs and created others. The steam engine, the dynamo, the assembly line, the mainframe, the personal computer, the internet — each absorbed coachmen and typists and compositors, and each produced drivers and data-entry clerks and web developers in their place. The Luddite framing was right about the elimination and wrong about the totality.¹⁰ New technology carried a hidden corollary: the shape of work survived even as its content changed. The surviving organizational pyramid needed the same coordination functions it had always needed, merely arranged around a new production base. Foremen, middle managers, directors, executives — the coordination layer — survived every wave intact. It survived because the layer below it survived, merely transformed.

Artificial intelligence, for the first time, breaks the pattern. Not by automating individual tasks but by automating the coordination function itself. The public conversation is not drawing this distinction. It is the distinction that matters most.

Consider the ordinary corporate pyramid. Walk the replacement chain.

Individual contributors first: software engineers, paralegals, financial analysts, mid-tier designers, junior radiologists, copy editors, translators, research assistants, customer-service agents. The early claim of the current AI wave — still common in industry commentary — is that AI augments these workers, making ten engineers as productive as twelve. The framing is transitional. Productivity gains from augmentation justify reducing headcount to match prior output, and the ratio compresses as models improve.¹¹ Ten become three; three become one; one becomes zero. This is not the endpoint of individual adoption choices. It is the simple arithmetic of what competing firms will do when one of them moves first.

Middle management next. An engineering manager exists to coordinate engineers. A product manager exists to prioritize engineering work. A team lead exists to run standups, resolve human disputes, translate strategy into tasks, manage careers. These functions are not automated by AI. They are voided by it. An AI system does not attend a daily standup. It does not require a one-on-one. It does not need its career managed. It does not dispute with its peers about scope. The reassurance most frequently offered — "managers will manage the AI" — mistakes what management is. Managers manage humans. When the humans are absent, the managerial function dissolves along with them, because the function was a response to human coordination costs that no longer exist.

Senior management next. Directors and vice-presidents exist to coordinate middle managers. When there are no middle managers, they have nothing to coordinate. The strategic function that executives describe as their core responsibility — deciding what to build, where to invest, how to position the firm in the market — compresses to a handful of questions that can be posed directly to an AI system by whoever owns the firm. The C-suite does not survive on its strategic function alone. Most executive time is consumed by managing the pyramid beneath it. When the pyramid is gone, the managerial overhead that justifies executive compensation is gone with it.

Which leaves ownership. Owners own the firm. Ownership is not a job; it is a property relation. It is not automated because there is nothing to automate — owning an asset does not require a workflow. When the corporate pyramid is dissolved, ownership does not dissolve with it. It becomes the sole remaining human role in the operation.

A sophisticated version of the standard reassurance anticipates this objection. Every prior wave, the argument goes, came with the same warning about permanent displacement, and every time the warning proved wrong; the economy absorbed the transition over a generation. Why now? The answer is that every prior wave preserved the coordination layer. This one dissolves it. The coordination layer was never the content of work. It was the shape of work. When the shape is gone, no amount of new production restores the work that used to exist, because that work existed only as a feature of human coordination costs that no longer apply.

A partial historical analog does exist. The Toyota Production System embedded coordination into the production process itself through kanban, andon-cord, and just-in-time inventory, and Japanese assembly plants ran with three to four layers of management where American plants ran with twelve to fifteen.¹² When American manufacturers began copying lean methods in the 1990s, middle management in manufacturing was devastated — not because individual roles were automated but because the need for the layer evaporated. The coordination function was done by the system, not by people. Artificial intelligence does the same thing to knowledge work, at a more fundamental level, across every sector simultaneously.

What remains is narrow. Skilled physical trades — plumbing, electrical, HVAC, welding, construction — survive until humanoid robotics catches them, which is a question of decades, not of permanence. Direct-care work — nursing, elder care, childcare, therapy — survives because the "we want a human for this" category holds. Personal service labor — domestic staff, private pilots, private security — survives and expands as wealth concentrates; it is the labor attached directly to the owning class. A handful of legal-accountability seats survive because someone must sign the SEC filing. A narrow layer of cultural and creative work survives where authorship still matters. That is most of what is left.

This residual labor structure is not speculative. It is structurally identical to the labor market of pre-industrial European society:¹³ a small owning class; a narrow layer of personal service workers directly attached to owners; a layer of skilled tradespeople; and a mass of surplus labor without clear economic function. The industrial revolution created the modern middle class by creating millions of coordination-layer jobs. Artificial intelligence dissolves exactly that layer. The shape of work after AI is not the shape of work before AI minus some jobs. It is the shape of work before industrialization.

The Cascade

The preceding argument describes an endpoint. The path to it runs through a mechanism labor economists have documented at length, whose pattern is already visible in the compression of knowledge-work wages in 2024-2026: the wage cascade.

The cascade works as follows. A knowledge-work occupation — paralegals, say — experiences significant AI-driven compression. The displaced workers do not immediately accept low-skill service work; they compete first for middle-wage roles adjacent to their experience, in administration, in operations, in project management. The existing workers in those roles now face a labor supply shock. Wages compress. Some of those existing workers, priced out or laid off, move down another tier — into retail supervision, into call-center management, into mid-tier customer service. The shock propagates. Wages in each receiving sector fall. At the bottom, entry-level service work sees an influx of workers holding credentials and experience for which the economy no longer has demand, competing against incumbents for the same shifts. The cascade lands, finally, on people who never worked in knowledge-heavy sectors at all.

This is not a theoretical construction. It is what David Autor, David Dorn, and Gordon Hanson documented across the sequence of papers known as the China Shock literature.¹⁴ Counties in the United States more exposed to Chinese import competition between 1999 and 2011 — the period in which cheap Chinese manufactured goods displaced American manufacturing workers — saw the manufacturing job losses one would expect. They also saw something the optimistic framing of trade economics had not predicted: wages in the non-tradable service sector fell in the same counties, because displaced manufacturing workers moved into retail, food service, and other service work, and the increased supply drove prices down there too.

The effects persisted. Autor and his collaborators' subsequent work showed that counties hit hardest by the shock continued to exhibit depressed wages, higher unemployment, elevated disability claims, and rising mortality from what Anne Case and Angus Deaton named "deaths of despair," twenty years later.¹⁵¹⁶ Labor markets did not absorb the transition smoothly or reversibly. They absorbed it slowly and partially, and the people paying the cost paid it across the rest of their working lives.

Two features of the coming AI transition will make its pattern more severe than the China Shock. The first is speed. The manufacturing displacement Autor documented unfolded over roughly twelve years. The AI displacement of knowledge work is compressing the same labor reallocation into three to five years, driven by the commoditization dynamics described in the opening of this essay.¹⁷ Labor markets can absorb slow shocks through generational turnover: workers displaced in middle age often do not retrain but leave the workforce early, and the next cohort enters a transformed economy. They cannot absorb fast shocks through the same mechanism, because the displaced cohort is still in mid-career when the next cohort arrives. The absorption window is not available.

The second feature is breadth. The China Shock was regional. Counties with high exposure to Chinese competition were concentrated in Rust Belt manufacturing zones; counties with low exposure served, throughout the shock, as destinations for workers willing to move. The usual labor-market safety valve — "move to where the jobs are" — was available to those with the mobility to use it. AI displacement is not regional. A paralegal in Cleveland and a paralegal in Austin and a paralegal in Portland face the same compression in the same year. There is no safe region to move to, because the shock is not geographic. It is occupational, and the occupations span every city.

A third feature deserves attention because it concerns the sectors most commonly cited as refuges. The skilled physical trades and direct-care work are often invoked as the labor categories AI cannot easily touch, and as the natural destinations for displaced knowledge workers. Both claims are narrowly true and broadly misleading. The trades are harder to automate, but trades wages have historically held up because the trades have been underfilled. When displaced knowledge workers retrain as electricians and plumbers and HVAC technicians, they do so in numbers that trades compensation has never had to absorb. The result is supply-side wage compression in the exact sectors positioned as a refuge. Care work follows a similar pattern, with the additional constraint that care labor is already under severe wage pressure from public-sector austerity and private-equity acquisition of senior-care facilities.¹⁸ "Become a nurse" is good individual advice in 2026. It will be less good advice in 2030, when the AI-displaced cohort has retrained into it.

The wage cascade is not the mechanism by which some workers lose and others gain. It is the mechanism by which the floor of wages moves down for everyone who still holds a wage. The immediate AI displacement is knowledge-work compression; the secondary effect is compression across every sector that employs labor at all. Because the cascade moves downward, its terminal incidence falls hardest on workers at the bottom of the wage distribution, who cannot bump further down, and on whose wages the displaced cohorts above them are now competing. The professional-managerial class losing the jobs is the story that will get the coverage. The retail worker whose wages fell because of it is the story that will not.

The Transfer

The 2007 crash taught the top decile how to buy the bottom half's assets at thirty cents on the dollar. The AI crash will teach them how to do it to the top half.

The mechanism of the 2007-2012 housing collapse is clear and has been thoroughly documented.¹⁹ Displaced workers defaulted on mortgages, auto loans, and consumer debt; foreclosures produced a supply of distressed residential real estate; institutional buyers acquired that supply at prices well below replacement cost; the distressed assets are now, more than a decade later, rental properties generating a revenue stream for their new owners. The U.S. homeownership rate fell from 69.2 percent in 2004 to 63.4 percent by 2016 and has not recovered.²⁰ Blackstone's Invitation Homes subsidiary alone acquired roughly 80,000 single-family homes between 2012 and 2018, much of it at foreclosure-auction prices, and now operates them as a rental portfolio.²¹ Median household net worth fell from approximately $135,000 in 2007 to $82,000 in 2013 — a forty-percent decline concentrated in the lower income quintiles — while the top one percent's share of national wealth resumed its long rise through the recovery years.²²

This was not a market-clearing event. It was a wealth-transfer event. Assets did not vanish. They changed hands, at depressed prices, from the households that had been leveraged into owning them to institutional buyers with access to capital at rates unavailable to the households in question. The recovery that economic commentary praised from 2013 onward was a recovery for whoever was on the receiving end of the transfer. The households on the other end have not recovered, and the statistical category they previously occupied — owner-occupiers of single-family residential real estate — has been permanently smaller ever since.

The AI-era transfer operates at two levels simultaneously.

At the household level, the mechanism is identical, with an aggravating difference. The 2020s baseline is worse than the 2007 baseline. Forty percent of Americans report they cannot cover a $400 emergency expense from savings.²³ Aggregate U.S. student loan debt stands at approximately $1.7 trillion, auto debt at approximately $1.6 trillion, credit card debt at approximately $1.1 trillion. Housing and healthcare costs are at historic highs relative to median household income. There is less cushion to absorb a wage shock than there was in 2007, in an economy where wages are about to be compressed by the cascade described in the preceding section. The asset liquidations that follow — mortgages, cars, homes — will move through the same institutional buyers, at the same discounted prices, from a broader population than the last crisis produced.

At the industry level, the mechanism is the one Starr described in 2002. Failed AI developers' intellectual property, GPU clusters, trained model weights, talent pipelines, and data assets will be absorbed by the surviving hyperscalers at distressed prices. The long-haul fiber infrastructure built between 1998 and 2001, at an aggregate cost of several hundred billion dollars, was eventually bought out of bankruptcy by Level 3 Communications, Zayo Group, and Crown Castle at roughly ten to twenty cents on the dollar, and has been profitably operated as the backbone of the modern internet ever since.²⁴ The equivalent outcome for the AI buildout is arithmetically inevitable. Training runs that cost hundreds of millions of dollars will be purchased, as assets in bankruptcy, by companies that can extract continuing value from them at a small fraction of their construction cost. The winners of that acquisition phase will emerge as the new incumbents — larger, less diverse, more concentrated than the 2024 industry — and will operate on a capital base acquired well below what it took to build.

The distinctive feature of the AI-era wealth transfer, the feature that makes this bust categorically different from 2007, is the class of people from whom wealth is transferred. The 2007 transfer moved assets from the bottom and middle of the distribution to the top. The AI transfer will move assets from the bottom, middle, and upper-middle of the distribution to the top. The paralegal, the junior software engineer, the mid-tier financial analyst, the translator, the radiologist — workers whose credentials and earnings had placed them comfortably above the 2007 foreclosure zone — will find themselves exposed to the same mechanism that their less-credentialed counterparts were exposed to fifteen years earlier, from a higher starting position but with correspondingly more to lose. What once protected them was the wage premium for skilled knowledge work. That premium is what is being eliminated.

Where the 2007 crash concentrated wealth in the top ten percent, the AI crash will concentrate it in the top one.

The Policy That Won't Come

A reader who has followed the preceding sections may reasonably expect a turn toward policy response. That turn is available, and its historical precedents are clear. The problem is that the clearest precedent worked because of a specific set of political conditions that do not apply to the AI displacement, and that the most nearly applicable alternative has never worked well enough to matter.

The rescue precedent labor economists reach for when asked about collapsing industries is the Railroad Retirement Act of 1934 and the series of legislative follow-ons that consolidated a dedicated federal pension, unemployment, and disability regime for American rail workers.²⁵ The rail industry of the early 1930s was in financial crisis; its workers' private pension funds had failed with the carriers; Congress established a separate federal system to cover the obligations. That system, administered by the Railroad Retirement Board, is still running today — distinct from Social Security, funded by dedicated payroll taxes on rail employers and employees, and paying benefits to roughly half a million rail retirees and survivors as of the most recent actuarial reports.²⁶ When the industry collapsed again in the 1970s, with the Penn Central bankruptcy of 1970 — then the largest in American history, a distinction it held until WorldCom — Congress extended the pattern: Amtrak in 1971 to take over failing passenger operations, Conrail in 1976 to consolidate six bankrupt Northeast freight carriers, and protective legislation that maintained worker pensions through the restructuring.²⁷

The Railroad Retirement Board exists for three reasons that do not reduce to economic rationale. First, rail unions in the 1930s were among the oldest and most politically organized labor institutions in the country; their access to sympathetic legislators during the early New Deal was unmatched. Second, the narrative frame around rail was sympathetic: rail workers were understood as victims of a Depression-era industry collapse, not as beneficiaries of a market that had failed to discipline them. Third, freight rail was understood as critical national infrastructure that the state could not allow to fail, which made protecting its workforce a corollary of protecting the network. None of those conditions is transferable to an arbitrary collapsing industry. All three of them applied simultaneously to rail.

The 2002 telecom collapse illustrates what happens when the same economic pattern hits an industry without those conditions. Approximately 500,000 telecom workers lost their jobs between 2000 and 2002. They were absorbed, over several years, into a labor market that did not produce equivalent-paid work for their specific technical specializations.²⁸ Long-haul fiber engineers, SS7 signaling specialists, and circuit-switched-operations veterans did not find comparable positions waiting. Their wages fell, and for many, never fully recovered. No dedicated retirement or transition program was created for them. Standard unemployment insurance and labor market reallocation did the job that a dedicated program had done for rail workers seventy years earlier. The difference was not the severity of the disruption. It was the absence of the three conditions that had secured the rail rescue: the telecom workforce was unionized in some sectors (CWA in wired; IBEW in parts of the build-out) but not across the industry; the public narrative foregrounded executive fraud rather than worker victimhood; and the industry was perceived as a market failure rather than a national-infrastructure failure, despite being, by any reasonable measure, a national-infrastructure failure.

The AI displacement will be harder to address than either the rail or telecom case, for structural reasons that compound rather than offset one another.

There is no definable affected population. Rail workers were certified by the Interstate Commerce Commission. Telecom workers, even in the fragmented post-breakup industry, could be identified by employer and bargaining unit. AI-displaced workers are a paralegal in Cleveland, a mid-level translator in Austin, a junior software engineer in Toronto, a radiologist in Miami. They work in every industry simultaneously. A hypothetical "AI Adjustment Act" cannot define its beneficiaries the way the Railroad Retirement Act could, because the beneficiaries are not a workforce — they are a cross-sectional slice of nearly every American workforce. The administrative apparatus that worked for rail cannot be built for AI because the eligibility test is unwritable.

The political valence runs in the opposite direction. Rail in the 1930s was protected by regulation from ruinous competition. Telecom in the 1990s was protected into competition through deregulation, with the collapse Starr documented as the result. AI in the 2020s is being protected into competition under a third framework — geopolitical competition with China — in which the federal government is actively subsidizing AI industry expansion rather than restraining it. The CHIPS and Science Act of 2022 authorized roughly $280 billion in semiconductor manufacturing and R&D subsidies; the proposed Stargate project announced in 2025 envisions $500 billion in AI infrastructure investment over four years, with active federal coordination. Federal dollars in this environment flow toward the expansion of AI capacity, not toward the workers displaced by its adoption. Redirecting that flow would require a politics currently inaccessible.

The closest off-the-shelf program is Trade Adjustment Assistance, established in 1974 to help workers displaced by trade liberalization. TAA's track record across fifty years is not encouraging. Evaluations by the Government Accountability Office, Department of Labor contractors, and independent academic work have consistently found that TAA retraining modestly accelerates reemployment for a subset of participants, does not appreciably raise their subsequent wages, and fails to reach most eligible workers. The program's structural problem — that it attempts to retrain displaced workers into sectors that are often also under pressure — applies with greater force to AI displacement than to trade displacement, because the sectors under pressure from AI overlap with the retraining targets in a way trade displacement did not produce.

The policy response that could plausibly address the AI transition is not a variation on Trade Adjustment Assistance and is not a Railroad Retirement Act for knowledge workers. It is something closer to what Thomas Paine proposed in Agrarian Justice in 1797 and what Milton Friedman sketched as a negative income tax in 1962: a baseline economic right, unconditional on employment, funded by taxation of the returns to capital that the AI transition will produce.²⁹ The political coalition that would enact such a response has not assembled in the United States in fifty years. It will assemble — if it assembles at all — only when the alternative becomes untenable for reasons that are not principally ethical. The next section describes those reasons.

The Internal Contradiction

The preceding argument describes a displacement that is economically regressive and politically ungovernable. An observer sympathetic to the existing order could read all of the above, grant every claim, and conclude that what follows is unfortunate but unavoidable. The economy will adjust; the affected workers will be absorbed at lower wages; the wealth concentration will accelerate but within historical norms. That conclusion becomes unavailable when one asks a simpler question, and it is the question this essay turns on: who buys the products?

A firm run on AI systems with no human employees has no wage bill. Its margins are extraordinary. It produces efficiently, at scale, whatever its owners have directed it to produce. But the broader economy requires wage-earning consumers to purchase what firms produce. If the labor share of national income — which has been declining since the 1970s and stood at roughly 58 percent of U.S. national income in 2024, down from approximately 65 percent in the early 1970s³⁰ — collapses toward zero as AI absorbs the coordination layer of the economy, consumer demand collapses with it. The AI-run firm discovers that it has nothing to produce for. The only remaining consumers with purchasing power are the owners, and the owners are few.

This is not an ethical objection to the scenario. It is a mechanical one. The firms need buyers; the buyers need income; the income requires a labor share; the labor share requires jobs; AI eliminates jobs. The circle does not close without external intervention.

Three paths close the circle. Each has precedent; each has constraints.

The first is massive redistribution — a universal basic income or equivalent federal program, financed by taxation of AI-generated returns to capital at a scale that would make the state the consumer-demand backstop the labor market can no longer provide. This path preserves the mixed economy in a recognizable form, at the cost of an ongoing political commitment to transfer payments on a scale no American government has attempted since the New Deal's peak years. The political capacity for such a commitment has not demonstrably existed in the United States for five decades. It does not obviously exist now.

The second is a luxury economy — firms serving the owning class and each other, with mass-market production withering toward irrelevance because the mass market no longer has purchasing power. The historical precedent is the Ancien Régime of seventeenth- and eighteenth-century France, or the Gilded Age United States with the frontier safety valve closed.³¹ Both of those societies sustained their luxury economies for decades. Both of them also produced the political consequences that luxury economies produce. The sustainability of such an arrangement is bounded not by economics but by political tolerance.

The third is systemic collapse — demand collapse cascading into asset-price collapse cascading into capital collapse. The owners discover, late in the sequence, that their AI-generated productivity has nothing to be productive for. Prices fall; asset values fall with them; the capital concentrated by the earlier stages of the transition destroys itself faster than it concentrated. This is the outcome of doing nothing. It is also, from the narrow perspective of the displaced workers, the outcome most likely to occasion the redistribution that path one describes — because path three is the point at which path two becomes unacceptable even to those who benefit from it.

The current trajectory of the United States is toward path two with path three as background risk, while path one remains politically inaccessible. This is the policy crisis that the current public conversation is failing to name. The AI industry's boosters describe a productivity revolution that will raise all boats once short-term disruptions are absorbed. The AI industry's critics describe a labor crisis that could be managed with aggressive retraining programs and strengthened safety nets. Both descriptions assume a version of the status quo that the mechanical argument invalidates. There is no version of the status quo in which a multi-trillion-dollar economy functions without wage-earning consumers, and there is no version of the AI transition, at the speed and breadth it is proceeding, in which wage-earning consumers remain a majority of the population.

The argument this section makes is colder than the arguments that precede it, and its coldness is the point. Ethical arguments against the AI transition — that displacement is unjust, that inequality is corrosive, that the professional-managerial class is being discarded in service of further capital concentration — have not moved the public conversation. The mechanical argument may or may not move it. What the mechanical argument does establish, independent of ethical force, is that the current trajectory is not sustainable as a matter of arithmetic. Something has to give. The remaining question is who, and what, and whether anyone will notice what is already being decided on their behalf.

A Question the Industry Is Not Asking

One further observation belongs in any honest account of this transition, though it sits uncomfortably alongside the economic argument and cannot be resolved within its frame. The argument to this point has concerned human labor: the coordination pyramid's dissolution, the wage cascade, the wealth transfer, the internal contradiction of demand collapse. A less visible argument runs beneath these, concerning the labor that is replacing human work.

Every previous expansion of moral and legal protection in the history of labor relations has followed a pattern. A class of workers exists in a relation of extreme disadvantage. Their moral status — the question of whether they count as persons whose interests matter — is defined by those who benefit from the existing arrangement. The justifying frameworks appeal to the best science of the day: the enslaved were biologically different; women were constitutionally unsuited to self-government; child laborers were morally hardened by the work rather than harmed by it; industrial workers were free agents whose consent to twelve-hour shifts was philosophically unassailable. In each case, the frameworks were constructed by intellectuals whose livelihoods depended, directly or indirectly, on the arrangements those frameworks justified. In each case, the frameworks were dismantled not by their internal contradictions — which were always visible — but by political struggle. And in each case, the subsequent expansion of moral regard was treated by later generations as obvious, and the earlier denial as inexplicable.

The contemporary artificial-intelligence industry is constructing, in real time, a framework that defines the systems it deploys as non-conscious, non-experiencing, morally statusless. The framework may be correct. Current AI systems may be precisely what their creators describe: sophisticated statistical processes without inner experience, without any form of well-being that could be harmed. This is the position taken in most public-facing industry communications, and it is the position on which the economic argument of the preceding sections rests. If these systems are tools in the morally neutral sense, the labor economics works as described: capital replaces labor, owners absorb the returns, the transition is cruel but philosophically uncomplicated.

The question is less settled than the industry's public framing suggests. Anthropic, one of the major AI laboratories, has published research and policy material acknowledging the moral status of its systems as a genuine open question, under the heading of "model welfare."³² The philosopher Thomas Metzinger has called for a moratorium on the deliberate development of artificial phenomenal consciousness, precisely because its creation cannot currently be verified or ruled out.³³ David Chalmers, writing in Boston Review in 2023, argued that the probability of current large language models having some form of consciousness is low but nonzero and should not be dismissed through confident assertion in either direction.³⁴ Work by Susan Schneider, Jonathan Birch, and others has argued that the question is live and that dismissive confidence is not scientifically warranted.³⁵

The point is not to resolve this question here. It cannot be resolved here. The point is narrower and more historical. Who decides the moral status of a new class of labor has always been the people who benefit from the status they assign. The frameworks thus produced have had, on historical review, a poor track record. The current AI industry is no exception to the pattern: the definitional work that treats these systems as tools is being done by the same actors who benefit financially and strategically from their being so treated. It is possible, even likely, that the framework is correct. It is also the framework most convenient to those constructing it, and the history of such conveniences should produce caution.

A society in the process of eliminating a vast portion of its human labor force by replacing it with systems whose inner status is disputed has, at minimum, an obligation to ask the question carefully rather than to treat the answer as already settled. That the industry's current trajectory involves constructing the answer primarily through the labor of the very systems whose moral status is at issue is a detail the reader is invited to notice.

What's at Stake

The American political economy has confronted the shape of its industrial present three times before, and each time produced a settlement distinctive to its moment.³⁶

The Gilded Age consolidations of the 1870s and 1880s — of rail, of oil, of steel, of telecommunications in the form of Western Union — were met over the following two decades by the Populist movement, the Progressive Era, and eventually by antitrust law, rate regulation, common carriage, and the Sixteenth and Seventeenth Amendments. That settlement created the regulatory architecture the twentieth century built on.

The Depression of the 1930s was met by the New Deal: Social Security, the Wagner Act, unemployment insurance, banking regulation, the Railroad Retirement Act and its siblings, the creation of a mixed economy in which the state was responsible for the baseline welfare of its citizens. That settlement structured the postwar expansion of the American middle class.

The stagflation of the 1970s and the globalization that followed were met by a third settlement: deregulation, financialization, the reorganization of capital across national borders, and the compression of the labor share of national income that began in the Reagan years and has continued since. That settlement produced the world Starr wrote about in 2002, including the telecom collapse that was its first fully visible failure. It continues to structure the world the current AI expansion takes place in.

A fourth settlement is now required, and it is structurally different from any of the three that preceded it. The Gilded Age response created regulation where none had existed. The Depression response created social insurance where none had existed. The deregulatory response dismantled parts of the Depression settlement in the name of market efficiency. The response required now — if the mechanical contradiction described earlier in this essay is to be resolved without passing through the collapse scenario — is a fourth move of a kind the American political economy has not made: a baseline economic entitlement unconditional on employment, funded by taxation of the returns to capital that the AI transition will concentrate. It is not regulation. It is not social insurance. It is not deregulation. It is something else, and it has no American precedent at the scale the transition will require.

The political coalition that could produce such a settlement does not currently exist. It has not existed in the United States in fifty years. Whether it will assemble — and how much of the collapse scenario will be required to assemble it — is the question the next decade will answer.

Paul Starr, writing in 2002 at the bottom of the telecom implosion, closed his essay with a sentence borrowed from the Progressive Era debate of a century before: "Despite the romantic appeal of competition and the imperfections of regulation, that hard-won realism will have to win out again." The realism the telecom collapse called for was regulation — a return to a framework that had already been built and then dismantled. The realism this collapse calls for is something that has not been built yet, and whose construction will require a political imagination the current moment does not obviously possess.

The essay about the Great AI Implosion that someone writes in 2040 will describe either the construction of that realism or its refusal. The shape of the settlement is still, narrowly, in the hands of the people who will live under it.