The Vanishing Farmer: What 200 Years of Agricultural Labor Can Teach Us About AI

One of the great fears surrounding artificial intelligence is straightforward: what happens to people when machines can do the work people currently do? We don't know. But we have conducted something resembling this experiment before.

In 1800, approximately three-quarters of the American labor force worked in agriculture. Historical reconstructions put the agricultural share at about 74% in 1800, around 60% in 1850, 40% by 1900, and progressively lower thereafter. Today, only around 1 to 2% of American workers are employed in agriculture, depending on exactly how agricultural employment is defined.

In other words, over two centuries, almost the entire agricultural workforce disappeared as a proportion of American employment. And almost nobody in 1800 could have predicted what those people would eventually do instead.

The Obvious Prediction Would Have Been Wrong

Imagine explaining the future to an American farmer in 1800. You tell him that eventually machines will plow fields, plant crops, harvest grain, milk cows, and perform work that then requires enormous amounts of human and animal labor. Perhaps he asks the obvious question:

What will everyone do? It would have been an excellent question. At the time, agriculture was not a marginal industry. It was the dominant occupation of the country. An estimated 74% of the labor force worked in agriculture.

If someone in 1800 had been told that eventually roughly 98 out of every 100 agricultural jobs, measured as a share of employment, would disappear, "mass unemployment" would have been an entirely reasonable prediction. That isn't what happened.

Instead, entirely new categories of economic activity emerged. Railroads, automobiles, aviation, telecommunications, pharmaceuticals, movies, television, advertising, software, professional sports, financial services, computer programming, biotechnology, and digital entertainment all eventually created economic activity that did not exist in anything resembling its modern form in 1800.

Millions of jobs eventually existed that an intelligent person standing in a field in 1800 could not merely have failed to predict. He would have lacked the conceptual vocabulary necessary to describe them. That distinction matters enormously when thinking about AI.

The Astonishing Part Isn't Merely That the Farmers Disappeared

The agricultural transformation becomes much more impressive when we look at what happened to agricultural output. In 1800, America had approximately 5.3 million people. Today it has well over 300 million. Yet the fraction of people required to produce agricultural output has collapsed.

The agricultural system therefore accomplished several things simultaneously. The population exploded, food availability per person increased, agricultural labor as a percentage of employment nearly disappeared, and the United States became one of the world's largest agricultural exporters. This isn't a story of Americans deciding that they no longer needed much food and therefore needing fewer farmers. It is a story of an extraordinary increase in output per worker.

One Worker Instead of Dozens

It is difficult to calculate agricultural caloric productivity precisely across two centuries because the underlying data aren't comparable. There was no USDA collecting comprehensive food-supply statistics in 1800. Definitions of agricultural employment changed. Modern agriculture produces exports, animal feed, industrial crops, and biofuel inputs as well as food. Imports further complicate the calculation.

But a rough calculation illustrates the magnitude. Suppose the average American around 1800 ultimately consumed or was supplied something like 2,000 to 2,700 calories per day. The precise figure is uncertain and should not be treated as established fact.

With approximately 5.3 million Americans and roughly 1.4 million agricultural workers, each agricultural worker was effectively supporting only a few people.

Today, one agricultural worker effectively supports well over 100 Americans in terms of domestic food availability, while American agriculture simultaneously produces enormous quantities for export and non-human uses.

Depending on definitions and assumptions, the improvement in food calories supported per agricultural worker is plausibly on the order of 50 to 100 times. The exact multiplier isn't the important finding. The order of magnitude is.

Where Did Everybody Go?

This is where the agricultural analogy becomes particularly relevant to AI. The workers didn't simply become unemployed farmers. Labor migrated. Agricultural productivity released enormous quantities of human time and effort that could then be used elsewhere in the economy. The National Academies describes the release of workers from farming as helping fuel the growth of the rest of the American economy.

People moved into manufacturing, then increasingly into services, and eventually into occupations that couldn't have existed before the technologies supporting them existed. The economy did not contain a fixed quantity of useful work that gradually got consumed by machines.

Human wants expanded as productive capacity expanded. Once a society could feed itself with a tiny fraction of its population, it didn't conclude that everyone else was unnecessary. It started doing other things.

Something Even Stranger Happened to Work

There is another part of the story that is easy to miss. As agricultural employment collapsed, participation in the formal market economy eventually expanded dramatically among women. The conventional historical statistics show an enormous increase in female labor-force participation during the twentieth century. Female participation in rich countries, including the United States, rose dramatically before leveling off around the turn of the twenty-first century.

But this needs an important qualification. Women in agricultural America weren't sitting around doing nothing. They were raising children, preparing food, making clothing, preserving food, caring for animals, working gardens, helping on farms, and participating in family businesses. Much of that productive activity simply wasn't counted as employment.

Recent historical work demonstrates how misleading the conventional statistics can be. One reconstruction finds that counting unpaid family workers raises estimated labor-force participation among free American women in 1860 from 16% to 57%. Another reconstruction of American working hours since 1870 finds substantial female participation in unpaid agricultural production and shows that women's total work follows a much more complicated historical pattern than conventional employment statistics imply.

So the remarkable change isn't simply: People used to work less and now they work more. It is: Work moved from farms and households into an increasingly specialized market economy. And that may be an even more important lesson for AI.

Technology Doesn't Merely Eliminate Work. It Changes What Counts as Work.

Consider a farm family in 1850. One person might grow food. Another might preserve it. Someone might repair equipment. Someone might make clothing. Someone might educate children. Someone might care for elderly relatives. Much of this occurred within a household or farm and therefore never appeared as a market transaction. Modern economic development progressively unbundled these functions.

Food processing became an industry. Clothing production became an industry. Education became an enormous profession. Elder care became an industry. Equipment repair became a specialized occupation. Transportation became an industry. Accounting became a profession.

A tremendous amount of activity moved from generalized household production into specialized market production. Agricultural productivity was therefore not simply a labor-destruction mechanism. It was part of a massive reorganization of human activity. AI may do something similar.

The Jobs AI Destroys Are Easier to Imagine Than the Jobs It Creates

This creates a systematic forecasting problem. We can see the occupations AI threatens because those occupations already exist. We cannot easily see occupations enabled by AI because many of them do not exist yet. The asymmetry is enormous.

A person in 1800 could look at a mechanical harvester and understand which farm workers it might replace. He could not look at that machine and infer that someday someone would earn a living designing smartphone applications. The causal chain was far too long.

The same problem affects predictions about AI. We can examine accounting, programming, graphic design, customer support, law, medicine, or administration and identify tasks that AI might perform. Those are visible.

What isn't visible is the economic activity that becomes possible when intelligence, analysis, organization, translation, programming, design, and information processing become radically cheaper. That second-order world may prove much larger than the first-order displacement.

This Doesn't Mean the Transition Will Be Painless

The agricultural precedent shouldn't be romanticized. People lost livelihoods. Communities declined. Skills became obsolete. Families moved. Economic power shifted. The transition occurred over generations rather than overnight.

An individual farmer whose livelihood disappeared didn't necessarily become a software engineer. Historical economies adapt at the population level in ways that can be extremely difficult for particular individuals. AI could also operate much faster than agricultural mechanization did.

That is a legitimate reason for concern. But it is different from assuming that eliminating existing tasks eliminates the human capacity to find valuable things to do. Two centuries of agricultural history provide powerful evidence against that assumption.

The Lesson From the Field

In 1800, roughly three out of four American workers were involved in agriculture. Today, agriculture employs only a tiny fraction of the workforce. If employment were fundamentally a process of dividing a fixed quantity of necessary work among human beings, this should have been disastrous. Instead, the opposite happened.

Hundreds of millions more Americans eventually lived at vastly higher material standards while consuming more food and producing an extraordinary variety of goods and services that didn't exist in 1800. The disappearing farmer did not produce the disappearing worker.

He produced the liberated worker. That doesn't tell us exactly what happens with AI. But it gives us a powerful warning about predictions. When someone asks, "What will people do when AI can do these jobs?", demanding that we name the replacement jobs may impose an impossible standard. Ask an American in 1800 what the displaced agricultural workers would eventually do.

He couldn't possibly have answered. Neither could the economist. Neither could the farmer. Neither could the government. The future occupations didn't exist. And yet the people eventually found things to do. The deepest lesson from agricultural mechanization may therefore be neither optimism nor pessimism about AI.

It is epistemic humility. We are very good at seeing the work technology is about to eliminate. History suggests we are much worse at imagining the work that becomes possible afterward.