What the Steam Engine Teaches Us About AI Inequality

In 1760, England was a mostly agricultural country where the majority of people worked the land using methods that would have been recognizable to their medieval ancestors. By 1850, it was the workshop of the world, producing half of all the iron and cotton cloth on Earth. The transformation was staggering in its scale and speed. It was also, for the people living through it, frequently catastrophic.

The standard economic history of the Industrial Revolution tends to focus on the aggregate: GDP rose, life expectancy eventually improved, standards of living ultimately increased for nearly everyone. This is true. It is also almost perfectly misleading as a guide to what the experience of industrialization felt like for the people who lived through its first several decades. Because for those people — the handloom weavers whose skills were made worthless overnight by power looms, the agricultural laborers displaced by enclosure acts, the urban factory workers laboring fourteen-hour days in conditions that shortened lives — the Industrial Revolution was not primarily an experience of rising prosperity. It was an experience of disruption, displacement, and deepening inequality.

This tension — between the long-run aggregate benefit and the short-run distributional disaster — is the most important fact about the Industrial Revolution for understanding what AI is doing to us now.

The economic historians have a name for the painful period between when a transformative technology arrives and when its benefits spread broadly enough to improve most people’s lives. They call it the Engels Pause, after Friedrich Engels, who documented the degraded living conditions of Manchester’s working class in the 1840s — at the height of the industrial boom, when England was becoming spectacularly wealthy by any aggregate measure. The pause lasted roughly sixty to eighty years. Productivity surged. Profits for factory owners and capital holders soared. Real wages for workers stagnated or declined. Then, gradually, policies shifted, labor organized, and the gains began to spread.

The analogy to AI is not precise — no historical analogy ever is — but the structural similarities are striking enough to be instructive. We are watching a technology that dramatically increases the productivity of capital-holders and highly skilled workers while disrupting the livelihoods of people whose comparative advantages have been in information processing, content creation, customer interaction, and routine cognitive work. These are not narrow sectors. They describe the majority of white-collar employment in advanced economies.

There is a particular parallel worth dwelling on: the enclosure movements. Between roughly 1750 and 1850, English common lands — shared resources that rural communities had depended on for centuries — were enclosed by acts of Parliament, converted to private property, and incorporated into large agricultural estates. The enclosure movements didn’t just displace agricultural workers. They transferred the value of a collectively held resource to private owners, effectively privatizing wealth that had previously been distributed, however unevenly, across communities.

Data is the commons of the AI economy, and what has happened to data over the past two decades rhymes uncomfortably with enclosure. The conversations, transactions, creative outputs, and behavioral traces of billions of people have been accumulated by a handful of platforms, used to train AI systems, and are now being deployed to compete with those same people for economic opportunities. The person whose customer service conversations trained a language model is not compensated when that language model replaces them in a customer service role. The writer whose work was scraped from the internet is not compensated when an AI trained on that work generates content that competes with theirs.

This is not a conspiracy; it is a structural dynamic. The Industrial Revolution’s factory owners were not primarily motivated by malice toward the handloom weavers they displaced. They were responding to market incentives in predictable ways. So are AI companies. The question is whether, as in the industrial case, the political economy will eventually shift to redirect more of the gains to people who bear the costs.

The history of what actually reduced industrial inequality is more complicated than optimists prefer and more hopeful than pessimists acknowledge. The transition from Engels Pause to broadly shared prosperity was not automatic or inevitable. It required specific interventions that were, at the time, politically contentious: mandatory public education that gave workers the skills to participate in the industrial economy, labor organizing that gave workers bargaining power over wages and conditions, health and safety regulation that imposed costs on factory owners that they would not have voluntarily incurred, and eventually social insurance systems that distributed some of the economy’s risks across the population rather than concentrating them among those least able to absorb them.

None of these interventions was framed at the time as a response to inequality per se. Education was framed as national competitiveness and moral improvement. Labor law was framed as preventing exploitation. Social insurance was framed as reducing social unrest. The effect was to broaden participation in industrial prosperity, not because anyone particularly planned it that way, but because the political pressure of industrially displaced people eventually translated into institutional change.

The uncomfortable lesson is the timeline. Sixty to eighty years is a long time to wait if you are a handloom weaver in 1790 or a call center worker in 2026. The aggregate eventually improves, but “eventually” can mean multiple lost generations in the interim. The historical optimist says: the technology creates more jobs than it destroys, and the living standards of everyone eventually improve. The historical pessimist says: tell that to the workers who died young in industrial Manchester.

What policies worked, and which didn’t? The most durable interventions were those that addressed supply-side human capital — education systems that gave workers the ability to do the new jobs that the technology created, rather than just protecting the old ones. Protectionist interventions that tried to preserve existing industries — there were serious proposals to ban power looms — bought time but couldn’t ultimately prevent the transition, and they delayed the adaptation that workers needed to make. Redistributive interventions that shared productivity gains through wages and social insurance were the most effective at shortening the painful transition period without blocking the technology’s adoption.

The analogy suggests that AI retraining programs, on-the-fly as they are, address a real need but in insufficiently ambitious ways. The Industrial Revolution required building entirely new educational systems, not just retraining individual workers for individual jobs. The scale of the challenge was institutional, not individual. If the AI transition is comparably disruptive — and there are reasons to think it may be more so, given its pace and breadth — the response needs to be comparably institutional in ambition.

What the steam engine analogy cannot tell us is how long the current pause will last. The industrial transition involved physically relocating workers from agricultural to manufacturing settings, which took generations. The AI transition may be faster because it involves cognitive work rather than physical, and cognitive adaptation can happen more quickly than geographical and occupational relocation. Or it may be slower because the skills being displaced are harder to retrain toward than operating a loom. We don’t know yet.

What we do know is the pattern. A general-purpose technology arrives. Its benefits initially concentrate in the hands of those who own or deploy it. The majority of the population experiences disruption rather than gain during the transition. Political pressure eventually builds. Institutions adapt. Benefits broaden. The technology is judged, in retrospect, as unambiguously good — and the suffering of the transition period is systematically underweighted in the retrospective.

That systematic underweighting is itself worth interrogating. It shapes how we respond to current disruption — treating the suffering of the transition as a necessary cost of long-run progress rather than as a policy failure to be addressed. The handloom weavers of England did not need to suffer as badly as they did. The political choices of the early industrial period could have been different. The same is true now.

The steam engine teaches two things simultaneously, and you have to hold both: that transformative technologies do eventually benefit humanity broadly, and that the transition is genuinely painful in ways that policy can either mitigate or worsen. The question is not whether AI will, in aggregate and in the long run, improve human welfare. It probably will. The question is what we choose to do about the sixty years in between.

There is one dimension of the industrial analogy that may be more alarming than encouraging. The Industrial Revolution’s inequality was primarily national: it happened within countries, between capitalists and workers in the same economy. The political pressure that eventually drove institutional change — labor organizing, progressive legislation, social insurance — worked through national political systems. Workers had political representation, however imperfect. They could vote, organize, and eventually extract concessions through democratic and quasi-democratic processes.

The AI transition’s inequality has a significant global dimension that the industrial analogy doesn’t fully capture. The gains from AI are concentrating not just within countries but between countries: in the United States, China, and a handful of wealthy nations with strong AI ecosystems, at the expense of developing economies whose comparative advantage in lower-cost service work is being eroded. Workers in the Philippines whose call center jobs are being automated by AI systems built in California have no political representation in California’s legislature. They have no ability to vote for the policies that would distribute AI’s gains more broadly. The institutional mechanisms that eventually corrected industrial inequality were largely national institutions, and we don’t have well-developed global equivalents.

This is the dimension of the analogy that should focus minds. The domestic inequality of the Industrial Revolution was eventually addressed, imperfectly, through institutional innovation: unions, progressive taxation, public education, welfare states. The global inequality of the AI revolution will require global institutional innovation of comparable ambition — and the current state of global governance suggests that developing such institutions will take much longer than the transitions that happened within national borders. The steam engine teaches us that these transitions can be navigated. It also teaches us that “can be” and “will be” are not the same sentence.