From Luddites to Humanoids: What the Steam-Engine Era Tells Us About AI Robotics Replacing Human Labor

The thesis I want to defend in this piece is simple, and possibly wrong: the rivalry between human labor and machine capital that started in March 1811, when Nottinghamshire framework-knitters smashed wide stocking frames in the name of “Ned Ludd,” is the same rivalry we are about to live through again — except this time the machine has perception, dexterity, and language, and the diffusion curve has been compressed from fifty years to fifty months.

That is not a moral claim. It is a structural one. Once you read the actual economic-history literature on the Luddites, Captain Swing, the Silesian weavers, and the so-called Engels’ Pause, you find that the 19th-century debate had almost nothing to do with whether machines would take jobs and almost everything to do with distribution, speed, and political response. Those are exactly the three axes on which the AI-and-robotics debate is being lost in 2026.

I’ll walk through what the protesters were actually protesting, why real wages stagnated for half a century while output exploded, what the modern task-based models of automation say about reinstatement of labor, and where perception-aware and language-capable robotics breaks the historical analogy. I’ll cite the work where I’m leaning on someone else’s data — including Caprettini & Voth (2020) in AER: Insights, Allen (2009) in Explorations in Economic History, Acemoglu & Restrepo (2019) in JEP, Acemoglu (2024) NBER WP 32487, Autor (2024) NBER WP 32140, Brynjolfsson, Li & Raymond (2025) in QJE, Eloundou et al. (2024) in Science, Mokyr, Vickers & Ziebarth (2015) in JEP, and most heavily on Juhasz, Squicciarini & Voigtländer’s 2024 Annual Review of Economics essay, which is the cleanest bridge in the literature from the framework-knitters to the BMW Spartanburg humanoid pilots.

Table of contents

I. The Luddites were not anti-technology

The folk version of the Luddite story — credulous artisans terrified of progress, smashing the future with hammers — is wrong on every empirical count. The framework-knitters of Nottinghamshire, the croppers of Yorkshire, and the handloom weavers of Lancashire were among the most technically sophisticated workers in Britain. They had spent careers operating, maintaining, and adapting the machines they later destroyed. What they were objecting to was not machinery as such. It was a bundle.

The bundle had four ingredients:

1. Wage collapse. Lancashire cotton spinners’ wages fell from 24 shillings/week in 1815 to 18s/week in 1818. Handloom weavers — about 200,000 strong at peak in 1812 — saw piece rates collapse and their numbers grind down to roughly 7,000 by 1861, in what Leunig and Voth (2018) in European Review of Economic History document as the canonical case of an entire skilled occupation hollowed out by power-loom diffusion.
2. Skill devaluation. The wide stocking frame, the gig mill, and the shearing frame did not eliminate the work; they let employers replace apprenticed craftsmen with unapprenticed “colts” at a fraction of the wage. The grievance was that customary apprenticeship rules — which had functioned as a labor-market institution for centuries — were being unilaterally voided.
3. Output adulteration. Many of the machines produced “cut-up” goods of inferior quality that undercut the market for properly finished work. The Luddite manifestos of 1811–12 explicitly distinguished between machines producing “fraudulent” output and those that did not.
4. Loss of autonomy. The factory regime — bells, gates, supervisors, fines for late arrival — replaced piece-work in cottages with industrial discipline. This is the pre-Marxian version of “alienation,” and it was felt before Marx had a word for it.

The protest was a bundle protest. Mechanization was the trigger but rarely the constitutive grievance. Hobsbawm and Rudé’s Captain Swing (1969) made this case for the 1830 agricultural unrest, and modern empirical work has tightened it considerably.

The most rigorous quantitative treatment is Caprettini and Voth (2020), “Rage Against the Machines: Labor-Saving Technology and Unrest in Industrializing England,” AER: Insights. They study the ~3,000 incidents of the Captain Swing riots, when southern English laborers smashed horse-powered threshing machines that had displaced something like a quarter of agricultural employment in winter. Instrumenting machine adoption by the share of heavy soils (where threshing machines were more profitable), they find threshing-machine intensity roughly doubles the probability of riot — from 13% to 26.1% in affected parishes. But the result that matters most for our question is the second one: areas with industrial outside options had less unrest. Where displaced agricultural workers could walk to a textile mill or a railway works, they did. Where they could not, they rioted.

That is the single most important historical finding for thinking about the AI/robotics era, and we will come back to it.

II. Engels’ Pause: the wage debate the AI literature has forgotten

Here is the part of the historical record that gets routinely waved away in techno-optimist commentary: the Industrial Revolution did not deliver rising wages to workers for at least the first fifty years of its existence, and possibly longer. The literature calls this the Engels’ Pause, and it is one of the few episodes in modern economic history where there is genuine, ongoing scholarly disagreement among first-rate economists about something this fundamental.

The pessimist position is Allen (2009), “Engels’ Pause: Technical Change, Capital Accumulation, and Inequality in the British Industrial Revolution,” in Explorations in Economic History 46(4). Allen argues real wages were essentially flat from 1790 to 1840, while output per worker rose ~46% and the profit share of national income roughly doubled. He models this as a Lewis-style surplus-labor regime in which capital had to be deepened — that is, accumulated faster than labor — before workers could share in the gains. The mechanism is brutal in its honesty: when the elasticity of labor supply is high (rural-to-urban migration, women and children pulled into the labor force, Irish immigration), profits are reinvested rather than competed away into wages, and the wage bill is anchored to subsistence until the underlying labor surplus is exhausted.

Feinstein (1998) in JEH had earlier constructed the canonical real-wage index showing only ~12% gain across 1780–1840 — a half-century of working-class stagnation in the country that was, by every other measure, becoming the wealthiest in human history.

The optimist counter is Clark (2005) in JPE, who argues that with different price weights, real wages rose meaningfully earlier. The most recent reassessment is Crafts (2021) in Oxford Economic Papers, who synthesizes the debate: yes, the Engels’ Pause is real, but it is smaller and shorter than Allen claims, and it is largely explained by what Crafts calls “pro-rich growth” — a rising profit share, terms-of-trade shifts, and measurement issues. Most importantly, Humphries and Schneider (2019) in EcHR add a distributional layer that almost no AI commentator engages with: women and child spinners, who are invisible in the standard male-wage series, were low-wage throughout the entire period. Even when male artisan wages held up, the household-level picture was much worse.

You can pick your side of this debate. What you cannot do — if you want to be honest — is claim that the historical record is one of immediate and broadly-shared gains from mechanization. Output and wages decoupled for a long time. The gains accrued to the owners of capital first, and to labor only after a long, contested, and sometimes violent political process.

This is the pattern that should be on every page of every AI labor-displacement memo. It is not.

III. Why wages eventually rose, and how long it took

Mechanization did, eventually, raise wages enormously. Real wages in Britain rose roughly fivefold between 1840 and 1914. The mechanism is well understood, and it is not “the market sorted it out.” It was four mechanisms layered together over roughly seventy-five years:

1. Capital deepening. The 19th-century capital-labor ratio rose by something like an order of magnitude. Machines complemented labor only after enough of them had been built that the marginal product of labor itself rose. This took decades of forced saving and reinvestment by industrialists.
2. New task creation. Railways did not exist in 1810. Steamships, telegraphy, gas lighting, retail distribution, civil engineering, mass schooling, white-collar clerking — none of these existed at scale before the 1840s. Mokyr’s The Lever of Riches (1990) is the canonical account: the “useful knowledge” base widened the menu of profitable tasks faster than mechanization closed the existing menu.
3. Political response. This is the part that gets deleted from most accounts. The Combination Acts (1799/1800) were partially repealed in 1824, legalizing trade unions. The Factory Acts (1833, 1844, 1847, 1850, 1867) limited child labor and working hours and — critically — created an enforcement inspectorate funded by the state. The Reform Acts of 1832, 1867, and 1884 progressively extended the franchise to working-class men, generating the political coalition that made redistribution possible. The Corn Laws fell in 1846. The 1870 Forster Act put the state in the business of mass primary education. None of these were technological. All of them were necessary.
4. Education’s race against technology. Goldin and Katz’s The Race between Education and Technology (2008) frames the 20th-century version of the same argument: as long as the supply of skilled labor grew faster than technology’s demand for it, the skill premium fell and labor’s share rose. As soon as that race tipped — which they date around 1980 — inequality reasserted itself.

The shortest honest summary: mechanization raised wages, after about fifty to seventy-five years, and only because of political institutions that the Luddite generation had to fight to put in place. The “machines lift all boats” story is true on a long enough timescale. It is also true that the people who lived through 1811–1840 did not get to use that timescale.

IV. The modern framework: displacement and reinstatement

The cleanest contemporary lens for thinking about this is the task-based automation model of Daron Acemoglu and Pascual Restrepo. Read in order:

• Acemoglu & Restrepo (2018), “The Race between Man and Machine,” AER 108(6)
• Acemoglu & Restrepo (2019), “Automation and New Tasks,” JEP 33(2)
• Acemoglu & Restrepo (2020), “Robots and Jobs: Evidence from US Labor Markets,” JPE 128(6)

The model decomposes any automation episode into two opposing forces:

• The displacement effect: machines take over tasks previously done by labor, mechanically reducing labor’s share of output.
• The reinstatement effect: technological change creates new tasks where labor has comparative advantage, raising labor’s share back up.

Whether labor’s share rises or falls is a horse race between the two. The empirical claim Acemoglu and Restrepo defend is that since around 1987, the displacement channel has accelerated in the United States while the reinstatement channel has weakened. Their robot study estimates that one industrial robot per thousand workers reduces the local employment-to-population ratio by ~0.2 percentage points and wages by ~0.42% — a meaningful effect, especially because industrial robots in the period studied were dumb, blind, and bolted to the floor.

A complementary line of work is the polarization literature. Autor, Levy and Murnane (2003) in QJE introduced the “routine vs. non-routine” task framework, and Autor and Dorn (2013) in AER showed that computerization hollows out middle-skill routine jobs while expanding both low-skill services and high-skill professional work — the “polarization” pattern that has dominated rich-country labor markets for thirty years. Goos, Manning and Salomons (2014) in AER confirmed this across European economies.

Two forecasting traditions sit on top of this framework. The widely-cited Frey and Osborne (2017) estimate — “47% of US jobs at high risk of automation” — was occupation-level rather than task-level, used an expert-judgment training set, and contained no productivity or price feedbacks. Arntz, Gregory and Zierahn (2016) at the OECD redid the same exercise at task level and got 9% rather than 47%. The ex-post correlation between Frey-Osborne risk scores and actual job loss is, by ITIF’s 2022 retrospective, about −0.26 — meaning the high-risk occupations grew slightly more than the low-risk ones, the opposite of the prediction.

This is the part of the literature where I would gently caution any reader against putting weight on a specific number. The frameworks are good. The point estimates are point-in-time guesses about a deeply non-stationary process.

V. Why perception-aware and lingual robotics breaks the analogy

Here is where the historical analogy gets harder, and more interesting. Every prior wave of automation — steam, electricity, internal combustion, the computer, the industrial robot — automated routine tasks. The mathematical statement of this is Polanyi’s paradox: we cannot automate what we cannot articulate. Tacit knowledge, dexterity, real-time perception, social judgment, language — these were the moats that kept human labor inside the production function long after the machines had taken over the routine middle.

Multimodal foundation models and humanoid robots attack those moats directly.

This is the substantive reason “this time might be different.” It is also the reason the most thoughtful economists in the field disagree about magnitudes:

• Acemoglu (2024), “The Simple Macroeconomics of AI,” NBER WP 32487 uses Hulten’s theorem on task-level exposure and cost savings to bound AI’s TFP gains modestly: at most 0.66% over ten years, or 0.53% if you adjust for hard-to-learn tasks. He projects that AI widens the capital-labor gap but is, unusually, distributed roughly equally across demographic groups — eroding the standard “skill up to be safe” answer.
• Korinek and Trammell (2023), “Economic Growth under Transformative AI,” NBER WP 31815 sketch scenarios where output expands tenfold while wages collapse, if the share of tasks where humans hold comparative advantage shrinks faster than new tasks emerge.
• Autor (2024), “Applying AI to Rebuild Middle Class Jobs,” NBER WP 32140 is the most genuinely optimistic recent paper. Autor argues that if AI is deployed as expert-judgment augmentation for sub-elite workers, it could actually re-fatten the middle class by extending licensed-expertise tasks (medicine, law, code, engineering) to a broader workforce. The empirical anchor is Brynjolfsson, Li and Raymond (2025) in QJE: a customer-support deployment showed a 14% productivity lift overall, 34% for novices, and almost zero for top performers. Compression toward expert quality.
• Eloundou, Manning, Mishkin and Rock (2024) in Science estimate that 80% of US workers have at least 10% of tasks affected by LLMs, and 19% have at least half their tasks affected. Crucially, higher-income jobs show greater exposure — an inversion of the skill-biased technical change pattern that defined 1980–2020.
• Brynjolfsson, Rock and Syverson (2021), “The Productivity J-Curve,” in AEJ: Macro revives Paul David’s “dynamo paradox” — electrification took ~30 years to show in factory productivity statistics — and predicts a measured-productivity dip during organizational reorganization, then a steep climb. The political-unrest window is the dip.

Two structural facts about humanoid robotics deserve to be foregrounded, because both are categorically new and both are missing from analogies that frame this as “just another wave.”

The first is simultaneity. Steam, electrification, the assembly line, the computer, the industrial robot — each automated a slice of the task space and left adjacent slices intact. The factory dismantled craft production but created the clerk, the foreman, the machinist, the engineer. Computers dismantled clerical work but expanded the analyst, the consultant, the customer-service representative. Each wave shifted labor out of one domain and into another, and the domain it shifted into was protected by some combination of perception, dexterity, judgment, language, or social knowledge that the machine of that era could not replicate. A multimodal foundation-model agent attached to a humanoid robot platform attacks all of those moats simultaneously. There is no adjacent domain in the way there was in 1830 or 1985. This is the part of the analogy that I think is genuinely without precedent.

The second is manufacturability. Human labor is supplied at a rate constrained by gestation, childhood, education, and migration — a function with a 22-year lag and a ceiling set by demography. Humanoid robots are supplied at a rate constrained by industrial-fabrication throughput, which has historically scaled like consumer electronics — orders of magnitude faster than population growth. Once the platform problem is solved, the substitution ratio is bounded only by capital expenditure and energy, not by labor supply. The 19th-century capital-deepening story took fifty years partly because building a steam engine was hard. Building the millionth Optimus is, mechanically, about as hard as building the millionth iPhone. The historical reinstatement effect was always given decades to find a domain. The robotics reinstatement effect may be given quarters.

A third structural fact, related but distinct: speed kills institutions before it kills jobs. The closest historical precedent for institutional repurposing on a multi-year rather than multi-decade timeline is wartime mobilization — the United States went from a peacetime economy to roughly 40% of GDP devoted to war production in eighteen months between 1940 and mid-1942, with rationing, price controls, the War Production Board, and women entering the labor force at unprecedented scale. The lesson from that episode is not encouraging for our purposes. It worked because there was existential consensus, centralized coordination authority, and a clearly defined endpoint. None of those conditions are present in the AI/robotics transition. We are asking institutions to repurpose themselves at wartime speed without wartime authority, consensus, or endpoint clarity. Organizations that are stressed eventually adapt. Organizations that are rushed often fracture instead.

These three facts together — a simultaneous attack on every prior labor moat, capital-good-rate supply scaling, and an adjustment timeline that compresses institutional reform into the unrest window — are why I am unsure the historical reinstatement story applies. They don’t make me certain it doesn’t. But they should make anyone uncertain, and most of the popular discourse is not.

The single best paper bridging all of this back to the historical record is Juhasz, Squicciarini and Voigtländer (2024), “Technology, Skills and Geography in the Spatial Diffusion of the Industrial Revolution”, with a related MIT Annual Review of Economics essay drawing the line directly from the early factory system to humanoid robotics. If you only read one paper from this post, read that one.

VI. What changes if physical labor is fully replaced

The reason the reinstatement question matters as much as it does is that almost every load-bearing institution of modern political economy is built on the assumption that human labor is the primary input to production. If that assumption fails — even partially, even regionally — the cascade of consequences runs much further than “unemployment goes up.” Let me sketch what I think actually changes. None of the items below are speculative tail risks; each one is a direct mechanical consequence of physical labor losing its central role in the production function.

1. The wage stops being the primary distribution mechanism

For two and a half centuries, the wage has been the dominant mechanism by which national output gets distributed to citizens. Tax bases, social insurance, retirement security, household formation, the political identity of “the working class,” the design of the entire welfare state — all of it sits on the foundation that human labor produces output, gets paid for it, and the distribution of those payments approximately tracks the distribution of citizens. If physical labor is replaced, that distribution mechanism stops working at the margin.

Universal Basic Income is the most-discussed alternative, but UBI as commonly described is a wage-tax-funded transfer from the residual workers to everyone else; it doesn’t survive a world where the wage tax base has collapsed. The serious alternatives are sovereign-wealth-fund-style citizenship dividends, worker ownership of robot capital (the cooperative answer), or some new asset-based distribution mechanism that doesn’t currently exist at meaningful scale.

The two largest live experiments are instructive. The Alaska Permanent Fund, founded in 1976, holds about $80 billion in oil-royalty-derived assets and pays a roughly $1,500–$3,000 annual dividend to each Alaskan resident — popular, durable, but small in absolute terms and entirely dependent on a single commodity. Norway’s Government Pension Fund Global, founded in 1990, holds roughly $1.7 trillion (about $300,000 per Norwegian) and funds the welfare state through annual draws constrained by a 3%-of-fund spending rule, but does not pay individual dividends. Singapore’s Temasek and GIC operate similarly without direct distribution. What these have in common is small, homogeneous polities; a clearly defined commodity or asset rent base; constitutional or political-cultural commitment to intergenerational savings; and a sustained track record of disciplined investment management. Federalizing that model at the scale of a US or EU economy — and funding it from a tax base composed of frontier-AI and humanoid-robot capital, not oil — is something no major economy has attempted. The design questions are open, the political coalitions to support it do not yet exist, and the time horizon to construct it is shorter than the time horizon over which the wage base may erode.

2. The service sector loses its embodiment shield

Roughly 80% of OECD employment is in services. The implicit reason these jobs were considered safe — through three previous waves of automation — was Polanyi’s paradox: a haircut, a heart-rate check, an ER intake, a plumbing repair, a restaurant meal all required physical presence, real-time perception, dexterity, and social judgment. These were the moats. Autor (2015), “Why Are There Still So Many Jobs?” JEP 29(3) was, in retrospect, largely an essay about why Polanyi’s paradox was holding. Humanoid robots invalidate that essay’s central premise on a sector-by-sector basis as the platforms mature.

The implication is not that all service jobs vanish next year. It is that the structural argument for why they were categorically protected is gone. Whatever timeline you assign to the displacement, the conceptual basis for “service work is safe because it requires a body” no longer exists. That mattered for fifty years of policy thinking. It matters less now.

3. The reinstatement effect requires a domain that may not exist

Acemoglu and Restrepo’s reinstatement channel requires some domain where labor has comparative advantage and where new tasks can be created. Historically that domain has always existed: cognitive work when machines did physical, non-routine cognitive when computers did routine, perception/social/dexterity when ICT did non-routine cognitive. Each wave handed labor a new continent to migrate into.

If perception, dexterity, language, and social interaction all fall to the same technology platform, the remaining domain shrinks to: status goods (“made by humans” as a luxury attribute), authentic interpersonal relationships (therapy, religion, friendship-as-service), and roles where human accountability is legally required (judges, certain medical decisions, legitimacy-bearing officials). These are real categories. They may not be large enough to absorb a working-age population. The honest answer in the literature is that we do not know whether they are or aren’t — Korinek and Trammell’s scenarios in which output expands tenfold while wages collapse are a formalization of exactly this uncertainty.

4. Capital ownership concentration vs. dispersion

The 19th-century industrial capital stock eventually became broadly distributed through homeownership, pensions, mutual funds, and 401(k)s. That broadening took roughly a century and required deliberate political construction — the GI Bill, mortgage subsidies, ERISA, the entire architecture of mass-market financial intermediation. The current AI/robotics capital stock is concentrated in a small number of model providers and a slightly larger number of humanoid-robot manufacturers.

If physical labor is replaced and capital ownership remains as concentrated as it currently is, the income distribution becomes Piketty’s r > g on steroids — capital gets a near-total share of output, and that capital is owned by a small fraction of the population. The question is whether ownership will broaden the way 19th-century capital ownership eventually did, or whether the political-economy dynamics of frontier-AI deployment prevent that broadening. This is the central distributive question of the next two decades. It is not resolved by markets alone. It is resolved by policy.

5. The fiscal architecture of the modern state needs new foundations

US federal revenue is roughly 50% individual income tax, 35% payroll tax, 10% corporate, with the remainder excise, customs, and estate. About 85% of federal revenue is labor-derived. Every European welfare state has a similar fiscal architecture, with VAT in place of some of the income/payroll mix but the same underlying assumption that human work is the primary taxable base.

If labor income shrinks meaningfully, the fiscal architecture has to reorient — toward capital income (at much higher rates than current), consumption (VAT-style), automation taxes (the Bill Gates “robot tax” idea), sovereign-wealth equity stakes in AI/robotics firms, or land-value taxes. Each of these has a deep theoretical literature and small-scale precedent. None has been the primary tax base of a modern state. The fiscal transition is at least as politically difficult as the labor-market transition, and it has to happen in parallel — the welfare state cannot wait while a new revenue base is invented.

6. Energy becomes the binding factor of production

If robot labor scales as a manufactured capital good rather than as gestated population, the binding constraint on aggregate output stops being “how many workers do we have.” It becomes “how many kilowatt-hours can we deliver, and at what price.” The geopolitics of compute (semiconductors, data centers — which I wrote about in the Cold War 2.0 piece and the GPU compute market piece) extends naturally to the geopolitics of energy.

Cheap and abundant electricity becomes the new wage rate. The countries and regions that win the transition are those with structural electricity-price advantages: hydroelectric basins, abundant solar, eventually fusion. Manufacturing relocates not toward cheap labor but toward cheap energy. The terms of trade between energy-rich and energy-poor countries become as decisive as labor-cost differentials were in the 1990s offshoring wave — and probably more decisive, because energy is harder to substitute for than labor was.

7. Demography inverts from liability to asset

Aging societies — Japan, Korea, China, Italy, Germany, much of the OECD — currently face fiscal crises driven by too few workers supporting too many retirees. The entire dependency-ratio framework assumes labor is the productive input that gets transferred via taxation to the non-working population.

If robot labor is unbounded, the relevant statistic stops being workers-per-retiree and becomes capital-and-energy per capita. Aging societies with high accumulated capital, built infrastructure, and energy supply may outperform young societies with abundant labor. The economic logic of immigration shifts from “we need workers” to “we need consumers and citizens.” Migration policy becomes a debate about national identity, consumption rights, and asset ownership rather than labor markets. This is one of the most counterintuitive consequences and probably the one most underweighted in current discussion.

The clearest live test will be Japan. Japan combines the OECD’s most aged population, world-leading industrial robotics manufacturing (Fanuc, Yaskawa, Mitsubishi Electric), high accumulated capital, dense infrastructure, and a culture relatively comfortable with humanoid form factors. Under the conventional dependency-ratio framework, Japan is the canonical fiscal-crisis case. Under the robot-capital framework, Japan is the canonical advantage case — possibly the first AI superpower of the embodied era. South Korea sits adjacent. Germany has the manufacturing base but a weaker software ecosystem. The mirror-image risk lands on emerging economies: India’s demographic dividend is supposed to pay off through the same export-services and offshoring pathways that lifted China — but those pathways depend on cost-of-labor arbitrage that humanoid robots eliminate. African industrialization-via-manufacturing, the textbook development pathway, may close before it opens. Whether emerging economies find new development models or get locked out is the most important development-economics question of the next two decades, and it is barely being asked.

8. Geography reverses 200 years of urbanization

Cities formed because labor and capital had to co-locate to produce industrial output. Manchester, Pittsburgh, Shenzhen, São Paulo are all answers to the same question: where do you put the workers next to the machines? Without labor in the loop, manufacturing optimizes for energy cost, materials access, and logistics — not worker proximity.

The 19th-century urbanization that built modern cities runs in reverse for production purposes. Cities survive only as consumption, cultural, and governance hubs — the trajectory already visible in San Francisco, Manhattan, and central London, where the productive base has thinned and the cultural-consumption layer is what remains. Whether this is good or bad is a separate question. That it is a 200-year reversal is not.

The reverse pattern is also possible: a small number of cities may re-concentrate even as production disperses. Cultural production, governance, in-person services that retain symbolic premium (high-end restaurants, live performance, religious life), and the consumption of urbanity itself may concentrate harder in fewer places, while production migrates to energy hubs in West Texas, Quebec, the UAE, the Australian outback, the Norwegian fjords. The global map of the late 21st century may have a small number of dense consumption-cities and a sparser network of energy-and-fab production zones, with most of the existing 19th-century industrial cities — the rust belts of every continent — emptying further. New industrial deserts in places that thought themselves permanently urbanized.

9. Comparative advantage between nations dissolves

Ricardo’s framework — the foundation of every theory of international trade since 1817 — rests on differing relative labor productivities across countries. If robot capital can produce any good anywhere, given energy and materials, comparative advantage collapses to: energy cost, regulatory regime, access to specific natural resources, and the location of irreplaceable infrastructure (semiconductor fabs, port logistics, certain mineral deposits).

International trade theory developed for a 19th-century world of mobile goods and immobile workers needs a new foundation for a world of mobile capital, mobile compute, and economically irrelevant labor. That is a lot of theoretical work for a literature that has not yet started.

10. The intra-capital fight: who becomes the railroad barons of this era

The discussion above frames the conflict as labor versus capital. The larger and probably more consequential fight is within capital. The 19th-century equivalent is illuminating: the railroad era did not just destroy farm and craft labor; it also created concentrated capital fortunes — Vanderbilt, Gould, Stanford, Hill, Harriman — that triggered the antitrust era (Sherman 1890, Clayton and the FTC 1914), trust-busting under Roosevelt, and a generation of progressive-era reform that constrained capital itself.

The AI/robotics era is producing the same dynamic, faster. The candidate concentrations are visible already: frontier-model providers (OpenAI, Anthropic, Google DeepMind, xAI, Meta), humanoid-robot OEMs (Tesla, Figure, 1X, Apptronik, the Chinese entrants), data-center hyperscalers (AWS, Azure, GCP, CoreWeave, Oracle), and specialized hardware (Nvidia, AMD, Broadcom, TSMC). The Cold War 2.0 layer adds national champions (Huawei, ByteDance, the Chinese SOE robotics ecosystem) that could either out-compete or be excluded from Western markets through export controls.

Within capital, the most interesting fight is not “incumbents vs. challengers” but commons vs. enclosure. Open-weights models (Meta’s Llama, DeepSeek, Mistral) commoditize the model layer and threaten the rents of frontier providers. Open-source robot platforms could do the same to humanoid OEMs. Whether the AI/robotics economy ends up looking like the late-19th-century railroad oligopoly (concentrated, regulated, eventually broken up) or like the post-1970s software industry (commoditized at the platform layer, with rents migrating to applications) is a live question, and the answer will determine how much intra-capital conflict spills into the broader political economy. Antitrust policy on AI infrastructure is the new chapter of the Sherman Act — and we have not written it yet.

11. Identity, dignity, and the politics of embodiment

There is a non-economic layer to this transition that the equations don’t capture. Work is not just a wage. It is a primary source of identity, dignity, social belonging, and meaning — a fact the dignity-and-meaning literature (Sandel’s Tyranny of Merit (2020), Case and Deaton’s Deaths of Despair (2020), Goldin’s Career and Family (2021)) has spent the past decade documenting in detail. The 1810 framework-knitter had a craft identity tied to a generations-long tradition of work. The 2030 knowledge worker has a professional identity that is, for many, the dominant component of selfhood. Replacing the wage with a transfer payment does not replace the identity.

Two specific dynamics deserve flagging. First, the humanoid form factor invites psychological response in a way disembodied software does not. The same model that produces equanimity as a chat interface produces visceral reaction when embodied as a robot in a factory or a hospital ward. Anthropomorphism cuts both ways: it makes robots more useful and also more politically salient. The historical analogue is the difference in protest energy between abstract financial reorganization (low public reaction) and visible industrial change like steam engines, threshing machines, factory floors (high public reaction). Humanoid robots are visible. They will be politicized in ways that pure software automation has not been.

Second, the politics of who is replaced first will interact with identity in ways likely to surprise even careful analysts. If displaced workers are predominantly male (industrial, construction, agricultural humanoid deployment) the political backlash will look like the deaths-of-despair coalition that produced Brexit, Trump, and Meloni. If displaced workers are predominantly female (caregiving, retail, administrative) the backlash will organize through different institutions and look different. If both are displaced simultaneously, the historically unusual coalition that emerges has no obvious template. None of this is captured in a labor-share regression.

The honest summary

I count at least eleven load-bearing institutions of modern political economy that are partially or wholly built on the assumption that human labor is the primary input to production — and on the further assumption that the political response to industrial-scale capital concentration arrives slowly enough that institutions can absorb it. If those assumptions fail, the institutions don’t fail simultaneously — they fail at different rates, with different transition costs, and through different political channels. But each one eventually has to be rebuilt or replaced.

The 19th-century answer was to build new institutions inside the unrest window: the factory inspectorate, the trade union, mass schooling, the welfare state. Each of those took decades and was contested at every step. The 21st-century question is whether we can do that on a faster timeline, with a more concentrated capital base, across institutions that took two centuries to build the first time, with a population that retains a craftsman’s psychological need for work in a world that may no longer need a craftsman’s labor.

That is what I mean by “dramatic.” It is not “robots take jobs.” It is that the load-bearing institutions of the modern state are mostly built on a foundation that may stop load-bearing. There is no precedent in industrial history for replacing eleven institutions in a single generation.

VII. What 19th-century unrest predicts about the AI/robotics era

Three things from the historical record translate directly to a forecast.

1. Compressed adjustment will overload the political safety valves

Power-loom diffusion in cotton took roughly fifty years from Cartwright’s 1785 patent to industry dominance in the 1830s. Threshing-machine diffusion in southern England took about thirty. ChatGPT reached 100 million users in two months. Figure 02 logged ninety thousand part-handlings on a real BMW production line in Spartanburg within roughly twelve months of demos. Optimus iterations are landing every quarter.

This matters because of the Caprettini-Voth result: outside options dampen unrest. The handloom weaver who could walk to a Lancashire mill or a Liverpool railway yard rioted less, even when his own occupation was being destroyed. When the new technology is general-purpose rather than sector-specific — and a multimodal robot policy that can drive a forklift, fold laundry, and answer customer email is precisely general-purpose — the outside options shrink at the same rate as the displacement.

The historical safety valve — “go work somewhere the new machine isn’t yet” — depended on the diffusion being slow enough and sector-specific enough that somewhere remained unautomated long enough for retraining and relocation. It is not obvious that condition holds anymore.

2. Inverted incidence will produce a politically novel coalition

In 1811 the displaced were skilled artisans (croppers, framework-knitters); in 1830 the displaced were unskilled rural laborers; in 1844 the displaced were Silesian linen weavers competing against mechanized British cotton. The 19th-century adjustment proceeded skill-tier by skill-tier, slowly enough that political coalitions could form one occupation at a time.

The Eloundou et al. and Acemoglu (2024) results suggest the AI/robotics episode inverts this. Higher-income, college-educated knowledge workers are more exposed than the past three waves of automation predicted. Service workers were protected by the embodiment gap that 2025–26 humanoid pilots are now closing. The political coalition that emerges from this is historically rare: simultaneously professional and service-class, simultaneously college-educated and not, with no obvious organizing structure inherited from previous labor movements.

If you want to forecast the political economy of this period, that is the key variable to watch. Coalitions that span education tiers are powerful when they form and difficult to form.

3. The Luddites lost the technology fight and won the political one — slowly

Did the Luddites “lose”? On the narrow question of whether mechanization proceeded: yes. The leaders were hanged or transported at the 1813 York trials; the Combination Acts suppressed organizing; Peterloo (1819) and the Six Acts (1819) made labor protest a capital concern. By 1850 the framework-knitters and croppers were gone as occupations.

But on the broader question of whether the working class extracted gains from industrialization: they won, eventually, through political mechanisms invented largely because of the unrest. The Combination Acts were repealed in 1824/25. The Reform Act of 1832 began the franchise extension that made working-class electoral power possible by 1884. The Factory Acts of 1833 onward built an enforcement state. The 1870 Forster Act put primary education on a national footing. By 1914 a British factory worker’s real wage was roughly five times what his great-grandfather earned in 1810.

The lesson — and this is the central thesis of Acemoglu and Johnson’s Power and Progress (2023) — is that the distributive consequences of a general-purpose technology are politically determined, not technologically inevitable. The same steam engine could have produced either Manchester in 1842 or Manchester in 1914. Which one it produced depended on the franchise, the unions, the inspectorate, and the schools.

VIII. The policy levers, then and now

If the 19th-century lesson is that politics decides distribution, the obvious follow-up question is: which 19th-century levers have working analogues today?

The comparison is humbling. The 19th-century levers worked on a fifty-to-seventy-five year timescale and required violent unrest, mass political mobilization, and a war or two to produce. The 21st-century analogues are mostly absent, partial, or contested.

Two policy questions stand out as genuinely new. First: the Combination Acts inverted. In 1811, capital was free to combine while labor was illegal. In 2026, the asymmetry is more subtle but real — model providers concentrate at the frontier, while exposed workers have no analogous coordination structure. Competition policy on AI infrastructure is the modern version of the Combination Acts question. Second: algorithmic auditing as inspectorate. The 1833 Factory Act was important not because it set new rules but because it created paid inspectors with authority to enforce them. AI safety regimes that lack an inspectorate-equivalent are the 1819 Cotton Mills Act, not the 1833 Factory Act, and the historical record on rules-without-enforcement is unkind.

There is a deeper problem the table glosses over. The 19th-century institutions were built into a relatively trusting polity — a society that had not yet figured out what national institutions could do, and was therefore willing to let new ones try. By every measure (Gallup, Pew, the General Social Survey, the World Values Survey) public trust in government, in major institutions, and in fellow citizens is now at or near multi-decade lows in most rich democracies. Polarization makes consensus harder. Regulatory capacity has been hollowed out by decades of underinvestment. Building algorithmic-auditing infrastructure, AI competition policy, citizenship-dividend mechanisms, and reskilling systems all simultaneously — at unprecedented speed and inside low-trust polities — is the structural challenge that gets mentioned least often and matters most. The 19th-century reformers were building from scratch. The 21st-century reformers are trying to repair, repurpose, and replace institutions inside societies that have already lost faith in them. That is a harder problem, and the historical record on institution-building under conditions of low public trust is unkind.

IX. What could go unusually well

I have spent most of this essay on the dark version of the transition because the dark version is underweighted in popular discourse. But there is a serious upside case that deserves attention, and intellectual honesty requires me to make it as carefully as I made the downside one. There are at least six places where AI and robotics could produce outcomes meaningfully better than anything the historical record offers.

1. The expert middle, restored. Autor’s argument in Applying AI to Rebuild Middle Class Jobs is that AI as expert-judgment augmentation can let sub-elite workers perform tasks previously locked behind credentials and tacit knowledge. A community-college-trained nurse with an AI clinical-reasoning tool can carry meaningful diagnostic load. A paralegal with a competent legal-research agent can do the work that previously required a junior attorney. This is the genuinely middle-class-rebuilding scenario, and the early Brynjolfsson-Li-Raymond results are consistent with it. If most of AI’s distributive effect runs through this channel rather than through outright displacement, the rich-country middle class re-fattens for the first time in fifty years.

2. Cost-disease sectors finally bend. Healthcare, education, government services, eldercare — the Baumol-cost-disease sectors — have absorbed an ever-growing share of national income for fifty years because productivity grows slowly and wages must keep pace. If robotics breaks that, the relative price of these sectors collapses, and the consumer surplus released is enormous. American healthcare spending alone is over $4 trillion annually and rising. Even modest cost reductions there represent the largest welfare gain in modern economic history.

3. Care abundance. The scarcest goods in late-modern societies are not material — they are time, attention, and human care. If robotics delivers eldercare, childcare, mental-health support, and patient companionship at quality and scale impossible with human labor, the felt-quality-of-life effect on aging societies could be larger than any GDP statistic captures. The most important consequence may be invisible to economists.

4. Scientific acceleration. Drug discovery, materials science, climate-relevant research, energy R&D — these all run faster with AI tooling. The case for AI-accelerated science is the case for compounding civilizational gains: a 2x or 5x speedup in the rate of useful scientific progress is, on long horizons, the most important thing that could happen to the species.

5. Emerging-economy leapfrog. The pessimistic version of the AI-and-development question is that robot labor closes the export-manufacturing pathway. The optimistic version is that countries which never built industrial bases may not need to — they can import productivity as a service. India delivering high-quality public services to a billion people at a fraction of US per-capita cost is technologically conceivable in a way it wasn’t five years ago. The leapfrog scenario has a real foothold.

6. New asset-ownership paradigms. The most underweighted upside is that AI/robotics may create new mechanisms for distributing capital ownership: tokenized fractional ownership of compute infrastructure, citizenship-equity stakes in sovereign AI/robot funds, data dividends paid for training-data contribution, cooperative robot-fleet structures. The 19th century invented the public corporation, the mortgage, the mutual fund, and the pension. The 21st century may need to invent its equivalents — and unlike institutional reform of the welfare state, financial-instrument design is something rich societies are reasonably good at.

These are real and they are not vaporware. They are also not automatic. The downside scenarios in Section VI happen by default if institutions don’t act. The upside scenarios in this section happen only with deliberate institutional construction. The fact that they are technologically possible is necessary but not sufficient. Whether we get the dark version of Section VI or the light version of this section is a political-economic question, not a technological one.

X. Open empirical questions worth being honest about

I want to end on uncertainty rather than overclaim, because the literature itself does not converge:

• Magnitude of reinstatement under near-AGI capability. Acemoglu (2024) bounds it modestly; Korinek-Trammell sketch scenarios where it explodes. The data does not yet discriminate, and anyone telling you confidently which way it goes is selling something.
• J-curve compression. The Brynjolfsson-Li-Raymond customer-support results suggest in some occupations the productivity dip may be skipped entirely. In others — physical tasks requiring humanoid-robot capital deepening — the J-curve may stretch closer to the 30-year electrification benchmark.
• Distributional incidence. Eloundou et al.’s SBTC inversion is striking. Whether it persists once humanoid robots fully close the embodiment gap is genuinely unknown. The political-economy implications differ wildly depending on whether knowledge workers and service workers are exposed simultaneously or sequentially.
• Outside-option scaling. The Caprettini-Voth dampening result is an empirical regularity from 1830, not a theorem. Whether it holds when the technology is general-purpose is the single most important question, because it determines whether the historical safety valve still works.
• The high-wage-economy reversal. Allen’s argument for why Britain mechanized first was that high British wages made labor-saving capital profitable. If the same mechanism operates in 2026, the high-wage knowledge economies — which is to say, exactly the rich-country professional class that thought itself safe — become the prime targets for AI deployment for the same reason.
• Speed itself as a policy variable. Mokyr, Vickers and Ziebarth (2015) in JEP catalog past technological anxieties — mechanization, automation, computerization — that failed to materialize as predicted. They are right that “this time is different” is a claim that must do work. The work it has to do is showing that the transition can be politically and socially absorbed at compressed speed, not that new tasks eventually emerge.

XI. What I actually think

I will stake out a position, possibly wrong:

The reinstatement effect will, eventually, arrive. Foundation-model agents and humanoid robots will create new tasks I cannot currently imagine, just as the 1810 framework-knitter could not have imagined a railway dispatcher or a chemical engineer. The five-times-real-wage outcome of 1810→1914 is, on a long enough timescale, the modal scenario.

But the transition is not the destination. The fifty-to-seventy-five year window of the Engels’ Pause was real, it was painful, and it produced political institutions — unions, factory inspectorates, mass schooling, the franchise — that the AI/robotics era has not yet produced equivalents of. The compression of the diffusion timeline from decades to years means we are trying to build those institutions inside the unrest window rather than in advance of it. The Caprettini-Voth result tells me the safety valve of “go work somewhere the new machine isn’t” weakens to the degree the technology is general-purpose, and a robot with perception and language is the most general-purpose technology since electrification.

And the dramatic version of the question — what changes if physical labor is fully replaced — is not a far-future thought experiment.

It is the question of whether the wage stops being the load-bearing distribution mechanism, whether the fiscal base of the welfare state survives, whether comparative advantage between nations still means anything, whether 200 years of urbanization runs in reverse, whether demography flips from liability to asset, whether the railroad-baron concentration of AI/robotics capital triggers a new antitrust era, and whether the dignity and identity that work has supplied for two centuries can be replaced by a transfer payment. Those are eleven institutions I count, possibly more. They are not all going to fail. But they are all going to be contested, in parallel, on a timeline shorter than any of them took to build.

If you are an entrepreneur, an investor, or a policymaker, that is the operating environment to plan for, not the comforting one in which jobs simply rotate from old industries to new ones.

The Luddites were not wrong about the bundle. They were wrong about the timeframe over which the political response would arrive. We are not wrong about the bundle either. The open question is whether we will be wrong about the timeframe in the same direction — believing the political response will be slow when it needs to be fast — or in the opposite direction, building safeguards against a transition that turns out to be far less disruptive than the model would suggest.

I genuinely do not know which it is. What I do know is that the version of this debate that pretends the 19th century was a story of immediate broadly-shared gains is not engaging with the historical record. The record is harder than that. We should expect ours to be too.

Appendix: Essential papers, with one-sentence whys

If you want to follow the literature without reading everything, this is the spine. Read in this order, the picture builds.

• Juhasz, Squicciarini & Voigtländer (2024), MIT/Annual Review of Economics — the cleanest scholarly bridge from the early factory system to humanoid robotics; if you read only one paper from this post, read this one.
• Allen (2009), “Engels’ Pause,” Explorations in Economic History — the canonical pessimist case on Industrial Revolution wages; sets the frame for thinking about long pauses between productivity and wage growth.
• Crafts (2021), “Slow Real Wage Growth during the Industrial Revolution,” Oxford Economic Papers — the synthesis that resolves the Allen-vs.-Clark pause debate without dismissing either side.
• Caprettini & Voth (2020), “Rage Against the Machines,” AER: Insights — the empirical anchor on what predicts unrest, with the all-important “outside options dampen unrest” finding.
• Humphries & Schneider (2019), “Spinning the Industrial Revolution,” Economic History Review — the distributional layer most accounts skip: women and child spinners were low-wage throughout, even when male artisan wages held up.
• Acemoglu & Restrepo (2019), “Automation and New Tasks,” JEP — the cleanest framework paper on displacement vs. reinstatement; required vocabulary for the rest of the field.
• Acemoglu (2024), “The Simple Macroeconomics of AI,” NBER WP 32487 — the modest-bound view on AI productivity gains, with the key insight that AI’s distributive effect cuts across demographic groups rather than along skill lines.
• Korinek & Trammell (2023), “Economic Growth under Transformative AI,” NBER WP 31815 — the formalization of the radical-disruption scenarios where output expands tenfold while wages collapse.
• Autor (2024), “Applying AI to Rebuild Middle Class Jobs,” NBER WP 32140 — the most genuinely optimistic recent paper, with a serious mechanism (expert-judgment augmentation) rather than vague techno-optimism.
• Brynjolfsson, Li & Raymond (2025), “Generative AI at Work,” QJE — the empirical anchor for the optimist case; 14% productivity lift overall, 34% for novices, ~0% for top performers.
• Eloundou, Manning, Mishkin & Rock (2024), “GPTs are GPTs,” Science — the SBTC inversion: higher-income jobs are more exposed than past automation waves predicted.
• Brynjolfsson, Rock & Syverson (2021), “The Productivity J-Curve,” AEJ: Macro — why AI productivity gains may dip before they climb, and why the political-unrest window is precisely the dip.
• Mokyr, Vickers & Ziebarth (2015), “The History of Technological Anxiety,” JEP — the cool historical perspective: many past technological anxieties failed to materialize as predicted, and “this time is different” is a claim that must do work.
• Acemoglu & Johnson, Power and Progress (2023) — the political-economy frame: distributive outcomes are politically determined, not technologically inevitable; the central thesis informs how to read every other paper on this list.

If you have time for only three: Juhasz et al. for the bridge, Acemoglu (2024) for the framework, Acemoglu & Johnson for the politics. The rest fills in the texture.