What happens to work after superintelligence?

In 1930, in the depths of the Great Depression, John Maynard Keynes published a short essay with an unexpectedly cheerful title: Economic Possibilities for our Grandchildren. Within a century, he predicted, living standards in rich countries would be four to eight times higher, and people would work perhaps fifteen hours a week. The real problem of the future, he wrote, would be how to use our leisure.

He was right about the first part. He was spectacularly wrong about the second. We became far richer and kept working almost as hard.

Keynes’s miss is a good place to start thinking about work after superintelligence, because it shows how poorly even brilliant economists predict what people will do with new abundance. It also raises the question that matters most: this time, will there be any work left to do?

The reassuring history

Fear of machines taking jobs is old. Between 1811 and 1816, English textile workers known as the Luddites smashed the new mechanised looms that threatened their livelihoods. Their fear was reasonable. Many of them did lose their trades.

And yet, over the following two centuries, automation did not produce mass unemployment. Machines took over spinning, farming, calculating and switchboards, and new jobs appeared that nobody had imagined. The economist David Autor and his colleagues estimated that around 60 percent of American workers in 2018 were employed in job titles that did not exist in 1940.

The pattern has a simple logic. Machines do some tasks more cheaply, which makes goods cheaper, which leaves people with more to spend on other things, which creates demand for new kinds of work. As long as there are tasks humans do better than machines, there are jobs.

The part of the history we skip

The reassuring story leaves something out: the transition. The economic historian Robert Allen described a period in early nineteenth century Britain, which he called Engels’ pause, when output per worker rose steadily for decades while working-class wages barely moved. The gains from industrialisation eventually reached ordinary people. It took roughly two generations.

“Eventually” is cold comfort to the people who live through the middle. The long-run statistics of the Industrial Revolution look wonderful. The lives of the handloom weavers did not.

What the evidence shows so far

With current AI, the early evidence is mixed and interesting. In 2023, researchers at OpenAI and the University of Pennsylvania estimated that around 80 percent of US workers held jobs where at least a tenth of their tasks could be affected by language models, and about 19 percent where at least half could.

Affected is not the same as replaced. A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed customer support agents given an AI assistant. On average, they resolved 14 percent more issues per hour. The least experienced agents improved by around a third, while the most experienced barely changed. The tool had spread the know-how of the best workers to everyone else.

So far, then, AI mostly behaves like earlier technologies: it changes tasks within jobs faster than it eliminates jobs. The question is whether that can last.

Why superintelligence might be different

Every past wave of automation had a limit. Machines took the physical tasks, then the routine mental ones, and humans moved to the work machines could not do: judgment, creativity, persuasion, care. The escape route always existed because there was always something left.

A superintelligence is defined by closing that route. If a system outperforms humans at virtually every cognitive task, and robotics eventually handles the physical ones, then the old logic of “new jobs will appear” meets a problem. New jobs will appear, but the machine may be better at those too.

In 1983, the Nobel economist Wassily Leontief made an uncomfortable comparison. Horses were not made unemployed by bad policy. They simply stopped being the cheapest way to do the work, and their numbers fell. He suggested that the role of human labour as the most important factor of production could shrink in a similar way.

There is a serious counterargument. Economists point to comparative advantage: even if a machine is better at everything, it still makes sense to trade with humans for the tasks where our disadvantage is smallest. People may also keep wanting other people, for care, for teaching, for performance, for the simple fact that a human made it. Some of this will surely hold. Whether it holds at a decent wage is another matter.

The real danger is a world where work no longer pays enough to live on, while the machines that replaced it produce more wealth than ever.

The real question is ownership

If machines do most of the productive work, the economy could become astonishingly rich. The question is who receives the income. Today, most people get their share of the economy through wages. In a world where wages matter less, the share goes to whoever owns the machines.

That is why economists like Daron Acemoglu and Simon Johnson, in their 2023 book Power and Progress, argue that the benefits of technology are never automatic. They depend on choices about who controls it and how its gains are shared. Proposals already on the table range from a universal basic income to public ownership stakes in AI systems to new ways of taxing automated production. None of them is simple, and all of them are political.

Keynes imagined that abundance would free us from work. He may yet be proved right. But freedom from work only feels like freedom if it comes with a share of what the machines produce. That turns the economic question into a question of power, which is where the next article in this series goes: who will control superintelligence?