The Fifteen-Hour Week
In 1930 Keynes predicted that within a century the rich countries would be four to eight times richer and would work fifteen hours a week. The first half came true with seven years to spare. The second did not come close, and the machines he feared would cause mass unemployment never did. Ninety-four years of data on output, hours, pay and jobs show where the productivity went: mostly into income rather than time, none of it into idleness, and since 1980, in America above all, a large part of it to the people at the top. Nobody chose that split; it was made for them. This first part follows the hours; the second, Where the Hours Went, follows the income.
In the autumn of 1930, with the world sliding into the worst depression in its industrial history, John Maynard Keynes published a short essay that ignored the depression almost completely. “Economic Possibilities for our Grandchildren” had started as a talk to schoolboys at Winchester two years earlier, was polished for an audience in Madrid in the summer of 1930, and appeared in two instalments in The Nation and Athenaeum that October.[1] Its argument was that the gloom was a misreading. What looked like the end of prosperity was, in Keynes’s phrase, the growing pains of over-rapid change. Underneath it, technical progress and the accumulation of capital were compounding at a rate that would, within a hundred years, do something no generation before had come near: solve what he called the economic problem, the struggle for subsistence that had occupied the species since it existed.
He made two predictions with numbers attached. The first was about income: “the standard of life in progressive countries one hundred years hence will be between four and eight times as high as it is to-day.” The second was about time. With needs that modest to meet and productivity that high, people would work far less; he suggested “three-hour shifts or a fifteen-hour week”, and spent the rest of the essay worrying about what a species bred for toil would do with the leisure. On the way he gave a name to the fear that runs under every argument about machines and work: “technological unemployment”, meaning unemployment caused by our finding ways to economise on labour faster than we find new uses for it. He thought it a temporary maladjustment. Plenty of people since have thought it the whole story.
The century is nearly up. It is 2026; Keynes’s grandchildren, had he had any, would be old. This essay asks three questions of the data. Has productivity really grown exponentially, or only fast? Where did the gains go: into more things, into more time, or into fewer jobs? And what does the answer say about the second machine age now arriving, which is aimed at exactly the kind of work that was supposed to be safe from the first?
Is it really exponential?
Mostly, yes, with an asterisk. “Exponential” means growth at a roughly constant percentage rate, so that each year adds not a fixed amount but a fixed fraction of what is already there. On an ordinary axis that shows up as a curve that looks flat for a long time and then shoots upward; on a logarithmic axis, where equal distances mean equal ratios, it is a straight line. That is the test. Figure 1 plots output per hour worked, the cleanest measure of what an hour of human work produces, for the United States, Britain, France, Germany and Portugal from 1890 to 2024, on a log scale.[2]
The American line is close to straight for a hundred and thirty-four years. A single constant rate, 2.2 per cent a year, fits it with an R² of 0.987: nearly all of the variation in the logarithm of American productivity since 1890 is explained by the passage of time alone. At 2.2 per cent a quantity doubles every thirty-two years, and that is what has happened, nearly four times over. An American hour of work in 2024 produced about seven times what it did in 1930; a British hour about six times; a French or German hour about ten times, and a Portuguese hour twelve.
The asterisk is that the rate is not constant in detail, and the recent detail is not encouraging. For the United States the rate was 1.9 per cent a year from 1890 to 1929, then about 2.6 to 2.7 per cent through the Depression, the war and the post-war boom until 1973, then 1.3 per cent until 1995, a burst of 2.4 per cent in the decade when computers finally showed up in the statistics, and 1.3 per cent again since 2005. Europe and Japan are more dramatic: France and Germany grew at five to six per cent a year from 1950 to 1973 as they caught up with American methods, Portugal at seven, and since 2005 all of them have grown at under one per cent, Britain at about half a per cent. The European lines in Figure 1 bend upward in the middle of the century and flatten at the end. What looks like one exponential is several, at different rates, joined end to end.
Over the really long run the picture changes again. Before about 1800 output per person barely moved for millennia, and the modern two per cent is itself a break from a near-zero past, which is why economic historians describe the whole of human history as faster than exponential: the growth rate itself grew, from nothing to two per cent, over two centuries.[3] Inside the last hundred and thirty years, though, and inside the countries that led, “exponential at around two per cent, slowing lately” is the honest summary.
As for what drives it, technology is only part of the answer, and not in the way the word usually suggests. Robert Solow’s famous 1957 calculation found that only about an eighth of the growth in American output per hour between 1909 and 1949 could be attributed to more capital per worker; the rest was a residual he called technical change, which is to say everything the accounting could not name.[4] Later work named some of it. Much of the residual is education, the rise in the skill of the average worker, which Goldin and Katz trace through the American high-school movement and the expansion of universities.[5] Some is the movement of people out of farming, where output per hour was low, into factories and offices, where it was high: in 1900 about four in ten American workers were on farms, and by 2000 fewer than two in a hundred.[6] Some is cheap energy, some is the scale of markets, and some is institutions, which decide whether any of the rest can be used. Robert Gordon’s argument, the most influential pessimistic account, is that the great burst between 1920 and 1970 came from a handful of one-off inventions, electricity, the internal combustion engine, running water, chemistry, telecommunications, whose effects could only be felt once, and that nothing since has matched them.[7] The flattening at the right of Figure 1 is the evidence he points to.
Four to eight: right, early
Keynes’s income forecast was a statement about GDP per head, and it can be checked directly. Taking 1930 as the base, real GDP per head by 2023 had risen by a factor of about 7.7 in the United States, 6.7 in Germany, 5.9 in France and 5.1 in Britain. All four are inside his range of four to eight, with seven years still to go. An older set of historical estimates, the Maddison Project’s, puts the American multiple lower, at about 5.5, and the British at 4.4; the range survives that too.[8] The countries that were poor in 1930 overshot it: Portugal is about fourteen times richer per head than it was then, Japan about twelve.
This is worth pausing on, because economic forecasts over a century are usually not worth the paper. Keynes got it right for a reason he stated himself. He took capital to be growing at about two per cent a year and technical efficiency at better than one, and compounded them. The number was not a guess about which inventions would come; it was an assumption that the process that had been running since the eighteenth century would keep running. It did.
Where the fifteen-hour week went
Now the hours. Figure 2 draws the American working week since 1870 next to Keynes’s; Figure 3 shows annual hours per worker for five countries over the same span, drawing on the long series assembled by Michael Huberman and Chris Minns and continued with modern national accounts.[9] In 1870 a worker in these countries put in between about 2,750 and 3,300 hours a year: fifty-three to sixty-three hours a week, fifty-two weeks a year, with Sunday off and not much else. By 1929 that had fallen to about 2,100 to 2,300, and by 2023 to 1,790 in the United States, 1,520 in Britain, 1,490 in France and 1,340 in Germany.
So hours did fall, by between a little under a quarter and a little over a third per worker since Keynes wrote, and by two-fifths to three-fifths since 1870. The fall is steady on average, between 0.36 and 0.59 per cent a year across these countries over a century and a half, which is roughly a fifth of the rate at which productivity rose. That ratio is itself a regularity economists have tried to explain: Timo Boppart and Per Krusell showed that the same slow, steady decline appears across many countries and argued that it is what you get when people are slightly more inclined to spend a rise in their wage on leisure than on goods.[10] Slightly, not mostly. And Figure 3 shows the other thing plainly: the American line stops falling around 1980 and is nearly flat for forty years. Nowhere is within a mile of the dashed line.
Hours per worker, though, is the flattering measure. It counts how long the average employed person works, and in 1929 a large share of adults, mostly women, were not counted as employed at all. Their work at home, which was long and hard, does not appear in these series. Over the century they entered paid work in large numbers, so the number of workers per head rose while the hours of each worker fell. Figure 4 does the sum: total hours worked in the economy divided by the whole population, children and pensioners included.
France and Germany took a real share of their productivity as time: hours per head fell from about 1,230 to 670 in France and from 1,120 to 730 in Germany. In the United States the line ends about where it began: 836 hours per head of population in 1929 and 847 in 2023 on this series. Other estimates of hours in 1929 put that year somewhat higher, around 900 hours per head, which would make the fall about 6 per cent; on any estimate it is small. The reason is that the share of Americans in work rose from about 39 per cent of the population in 1929, 47.6 million people out of 121.8 million, to about 48 per cent in 2023, 161 million out of 335 million, mostly because women moved into paid work. That rise offset most or all of the fall in hours per worker. The Netherlands, with its short part-time weeks and high participation, lands in a similar place by a different route. In the aggregate, per person alive, the richest large economy in history works about as much as it did in the year of the Crash.[11]
Some of that is demography, since the share of children in the population fell and the share of pensioners rose, and some is the welcome fact that women’s work now shows up in the accounts and is paid. It would be wrong to read Figure 4 as saying nobody gained any time. But it does say, unambiguously, that the gains were not taken as time on the scale Keynes imagined.
Things, not time
Put the two pieces together. If productivity rises by some factor, a society can take that factor as more output, as fewer hours, or as any mix of the two. On a logarithmic scale the split is additive, so it can be written as a share: what fraction of the rise in output per hour between 1929 and 2023 was spent on working less? Figure 5 does this for thirteen rich countries with long records, once per worker and once per head of population.
Per worker, the share runs from about a tenth in Italy and the United States to just over a quarter in the Netherlands. Per head, it runs from nothing, in the United States, Australia and the Netherlands, to about a quarter in France. There is no country that took most of its gain as time. There is no country that took even a third. The typical rich country took something like one-fifth of a century of productivity growth as shorter working lives and four-fifths as more income.
The same numbers can be turned round to ask what Keynes’s world would have looked like. American output per hour in 2023 was high enough that the 1929 standard of living, everything the country produced in the year of the Crash divided by everyone in it, could be produced with the 1929 share of the population at work putting in about six hours a week, averaged over the year, or six and a half on the higher estimate of hours worked in 1929. The British figure is almost identical. Keynes was not too optimistic about productivity. If anything he was too cautious: at the living standard he knew, the fifteen-hour week was available twice over.
Run it the other way and the answer is more pointed. Suppose today’s share of Americans in work each worked fifteen hours a week, every week of the year, at today’s productivity. They would produce about $31,000 a head in 2020 dollars: three and a third times the 1930 standard of living, inside Keynes’s range, and exactly the American standard of living of 1977. The British equivalent is 1979, at about two and a half times the 1930 standard, just short of Keynes’s range.
A fifteen-hour week at today’s productivity buys the standard of living of the late 1970s. The rich world took the income instead.US and UK, output per hour in 2023
That is the whole of the Keynes question in one sentence. The fifteen-hour week was never technically out of reach. It was, and is, an offer on the table at a particular price: a house, a car, food, clothes, medicine and holidays at the level of 1977, with the phones, the computers and the drugs that did not exist then subtracted, and the time of two and a half days’ work a week added back. In aggregate the rich world has consistently turned that offer down. “In aggregate” is doing a great deal of work in that sentence, and the second part of this essay, Where the Hours Went, is largely about how much. The offer was made to the economy as a whole, not to any particular worker, and whether most workers ever saw it depends on who received the income, what it could buy, and whether they could choose their hours at all.
What the 1977 basket costs now
The counterfactual above values the 1977 standard of living with an average price index, which quietly assumes that people buy more of whatever has become cheap. The literal 1977 basket is a different thing. Taking the 197 categories of American consumer spending in 1977 and pricing each at its 2025 price, the basket comes to about $31,400 a head, 51 per cent of what Americans consume today; deflating with the chained index instead gives 40 per cent, which is where the fifteen-hour answer comes from.[12] At the 34 hours a week that American workers average over the year, the literal basket needs about 17.5 of them, the chained one about 13.6. Buying the actual life of 1977 costs about four hours a week more than the headline says.
The reason is in Figure 6, which measures each part of the basket in hours of work at the average hourly earnings of production and non-supervisory workers, the rank-and-file employees whose pay, as the second part of this essay shows, has followed productivity least. Television sets, audio and computers now take 4 per cent of the hours they took in 1977, clothing a fifth, furniture and appliances a quarter, cars under half and food at home three-quarters. Restaurants and fuel take about the same time as then. Housing takes 10 to 20 per cent more, childcare and health care half as much again, and education more than two and a half times as much.[12] Things became cheap; other people’s time and land did not.
Measured at that wage, the whole 1977 basket still takes 94 per cent of the hours it took in 1977, or 73 per cent if the cheapened goods are substituted in. The big items show why. The median new American house cost 8,974 hours of that wage in 1977 and 13,319 in 2025, 3.6 and 4.8 times median household income, though today’s houses are larger; a year of college, fees and board included, went from 418 hours to 950; health spending per person from 142 hours to 514, most of it now paid by insurers and the state.[13] Housing is dear because land is: rising land prices account for about 80 per cent of the rise in house prices across rich countries since the Second World War, and land’s share of the value of American homes doubled, from about a fifth to two-fifths, between 1970 and 1990.[14] That is the price set by other people’s incomes.
The indexes also miss things in both directions. The 1977 basket contains no mobile phone, no internet and none of the drugs developed since, and official price indexes are thought to overstate inflation by between about half a point and a point a year, much of it from new products and better quality.[15] Against that, the goods that fell furthest in price are the ones that are easy to give up, and the ones that did not fall, a home in a place with work, a school, a doctor, childcare, are the ones that are not. The honest answer to “what could be dropped to buy back a day a week” is: most of what got cheap, and almost none of what got dear.
Retirement: where much of the time went
Figure 5 counts hours in a working year. It misses the time that was taken at the two ends of a working life: people start later, because they study longer, and until recently they stopped earlier and then lived much longer. The OECD has tracked the second half since 1970.[16] In 1970 men in most rich countries left the labour market between 63 and 69. By 2000 the average had fallen to between 58 and 63: 58.4 in France, 59.5 in the Netherlands, 60.7 in Portugal, 62.5 in the United States. At the same time life expectancy at 65 rose by several years. Figure 7 shows the result: the number of years a man could expect to live after he stopped working went from between ten and fifteen in 1970 to between seventeen and twenty-three today. For women the figures are longer still, about 26 years in France and 25 in Spain.
Counted this way, Europe took much more of its productivity as time than the yearly ledger suggests. Take a man who starts work at 20 and leaves at his country’s average exit age, working the average annual hours of his time. Between 1970 and 2024 his career shrank by 32 per cent in France, 27 in Spain, 26 in Denmark, 22 in the Netherlands and 17 in Britain. Measured on careers rather than years, the share of productivity growth since 1970 taken as time comes to about 37 per cent in France, about 30 in Spain, Denmark and the Netherlands and 20 in Britain. In the United States it is about 5 per cent, because American men now leave work later than they did in 1970. And these figures understate the time gained: they ignore the later start, since the average American adult had about 8 years of schooling in 1930 and about 13 in 2020, and they leave out the extra years of retirement that longer lives added on top.[17] The rough rule of the ledger, a fifth as time and four-fifths as income, describes the typical rich country; America took a tenth or less. For much of Europe, close to a third or more was taken as time, a good part of it at the end of working life.
That part of the dividend is now being clawed back. Since about 2000 average exit ages have risen by between one and a half and six years: in Portugal from 60.7 to 66.5 for men, in the Netherlands from 59.5 to 65, in the United States from 62.5 to 67.3. Pension reforms raised the official ages and, in many countries, tied them to life expectancy. The OECD calculates that a man entering the labour market at 22 today will reach the normal pension age at 68 in Portugal and Britain, 70 in Italy, the Netherlands and Sweden, and 74 in Denmark. The retirement that productivity paid for in the twentieth century is being shortened in the twenty-first, for the demographic reasons in the outlook section below. Whether the extra working years are healthy ones, and whether the pensions at the end are enough to live on, is the subject of the companion essay, The Long Retirement.
A lifetime in hours
Both measures so far miss part of a life: hours per worker ignore who works, and careers ignore childhood, schooling, housework and the years after work. A fuller ledger follows one imaginary person from birth through today’s life tables, today’s employment rates at each age and today’s hours per worker, the same logic as life expectancy itself, and counts the waking hours of that life, at sixteen a day, and how many of them go to paid work.[18] Figure 8 does this for nine countries, from the earliest year with employment data by age to 2023.
Two things stand out. The waking life got longer everywhere, by 11 to 24 per cent, because people live eight to thirteen years longer. The hours of paid work in it hardly moved: about 58,000 in the United States in 1960 and 64,000 now, 62,000 and 51,000 in Germany between 1970 and 2023, 59,000 and 50,000 in France, 76,000 and 71,000 in Japan. In Portugal they rose from 46,000 to 62,000 as women moved into paid work. As a share of waking life, paid work fell from about 15 to 11 per cent in Germany, 14 to 10 in France and 18 to 14 in Japan, and not at all in the United States. Set against the rise in output per hour over the same years, the share of productivity growth taken as time over a whole life is about 37 per cent in France, 29 in Germany, 21 in Spain, 16 in Sweden, Norway and Japan, and 2 in the United States; in Canada and Portugal paid work rose faster than waking life. These are close to the career figures above, which is reassuring: the two methods agree that continental Europe took something like a fifth to a third of its gains as time, and America almost none.
The ledger leaves out unpaid work, and time diaries fill part of that gap. Across adults of all ages, including the retired, total work, paid and unpaid together, fell only a little between the 1970s and the 1990s or 2000s: from 56 to 54 hours a week for Norwegian men and 56 to 53 for women, 50 to 46 for British men and 46 to 46 for British women, 61 to 56 for French women, while in the Netherlands it rose, as women took paid jobs and men took more housework.[19] The shift of women from unpaid to paid work, and of men towards housework, reshuffled the total more than it cut it. Portugal has had no comparable diary survey since 1999, so its rise in paid work cannot yet be set against a fall in unpaid work.
Over longer spans the picture is clearer, and the longer life is the main event. In Britain the share of a worker’s disposable lifetime, after sleep, meals and childhood, spent at work fell from about half in 1856 to less than a fifth by 1981.[20] In the United States, Ramey and Francis estimate that lifetime leisure from age fourteen rose by about half between the generation born in 1890 and the one born in 2000, but as a share of the lifetime it moved only from 24 to 26 per cent: most of the extra leisure came from living longer, not from working less in a life.[21] Much of the time dividend of the twentieth century was paid out as more life, which is a real gain and partly a product of the same wealth. But it is not the fifteen-hour week; it is the forty-hour week, lived for longer.
Why the time was not taken
The explanations are not exclusive, and most of them have evidence behind them.
The wage cuts both ways. When the hourly wage rises, an hour of leisure costs more in forgone income, which pushes people to work more; but they are also richer, which pushes them to buy more of everything, leisure included. The two effects roughly cancel, which is why hours fall slowly rather than fast. Keynes assumed the second effect would dominate once basic needs were met. The data say it dominates only slightly.[10]
Keynes saw the catch himself. Halfway through the essay he divides human needs into two classes: those that are absolute, felt whatever other people have, and those that are relative, felt “only if their satisfaction lifts us above, makes us feel superior to, our fellows”. The first class, he says, can be satisfied. The second “may indeed be insatiable”. His whole prediction rests on the first class mattering more than the second. A century of consumption data suggests the reverse. A great deal of what rich people spend on is position: the better neighbourhood, the better school, the car that says something. Positional goods cannot be produced in greater quantity by any amount of productivity, because their value lies in being scarcer than someone else’s. Samuel Bowles and Yongjin Park found that across rich countries over several decades, working hours were longer where income inequality was higher, which is what you would expect if people are working to keep up with those above them; they called their paper, fairly, “Was Thorstein Veblen Right?”[22]
The long hours moved to the top. In the 1890s the lowest-paid men worked the longest days, because they had to. By the 1990s the pattern had reversed: the highest-paid worked longest.[23] Leisure, which had been the mark of the rich in every earlier society, became something the less educated had more of, much of it involuntary.[24] For the people whose hours set the tone of professional life, an extra hour of work is worth the most and a fifteen-hour week looks like giving up the race.
What we buy got more expensive in hours. Productivity has grown fastest in making things and slowest in services that consist of one person’s time spent on another: teaching, nursing, caring, performing. William Baumol pointed out that these services must still pay wages that keep up with the rest of the economy, so their relative price rises steadily; a string quartet needs as many musician-hours as it did in 1800.[25] As a society gets richer it spends more on health, education and care, exactly the sectors where an hour of work buys least, and the average productivity of what it consumes grows more slowly than the headline figure. Housing, whose price is mostly the price of land in the right place, is the extreme case and a positional one at that.
Institutions decided the split. Europe and America had similar hours in the 1960s and then diverged, with Europeans working about a third fewer hours per adult by the early 2000s. Edward Prescott attributed most of the gap to higher European taxes on labour, which make an extra hour of work less worth doing.[26] Alberto Alesina, Edward Glaeser and Bruce Sacerdote attributed it to unions and regulation: legal limits on the working week, mandated holidays and collective agreements, which coordinate a choice that no individual worker can make alone.[27] Their coordination argument is the important one. If everyone works less, nobody falls behind; if one person works less, they do. Taking productivity as time is partly a collective decision, and countries that made it collectively, through law and bargaining, took more of it. The institutions that won the shorter week have themselves weakened. Union membership, which fought for the eight-hour day, the weekend and paid holidays, fell from 31 per cent of American employees in 1960 to 10 per cent in 2024, from 52 to 22 per cent in Britain since 1980, and in Portugal from about 65 per cent in 1973 to 14 per cent; only in the Nordic countries, where unions help run unemployment insurance, does a majority of employees still belong.[28]
Some of the time came in forms the series do not show. Lives got longer and working lives did not get longer to match: people start work later, after more education, and stop earlier, with pensions, and spend a much smaller fraction of their waking lifetime at work than their great-grandparents did. Housework shrank with running water, washing machines and prepared food, which is part of why women could enter paid work in the first place. Valerie Ramey and Neville Francis, adding all of this up for the United States, found that leisure per person rose over the twentieth century, but by much less than the fall in hours per worker suggests.[21] Mark Aguiar and Erik Hurst, using time diaries, found real gains in leisure between 1965 and 2003, larger for less educated adults.[24] Something was gained. It was a fraction of what was possible.
Two further parts of the answer, who received the income and what it was spent on, and what the hours themselves cost, are the subject of the second part of this essay, Where the Hours Went.
Technological unemployment
If productivity rises sevenfold and the amount of work per head stays the same, then by arithmetic the economy is producing seven times as much. That is the answer to the third question, and it is why the fear Keynes named has, so far, been wrong about the aggregate. Figure 9 shows unemployment in the United States and Britain since 1870 from the Jordà–Schularick–Taylor macrohistory database.[29]
There is no trend. American unemployment averaged about 7 per cent from 1890 to the First World War, 11 per cent between the wars, 5 per cent from 1950 to 1973, 6.6 per cent from 1974 to 1999, and 6 per cent from 2000 to 2020. British unemployment averaged 4.5 per cent before 1914 and 5.6 per cent in the twenty-first century. The big episodes, the 1930s, the early 1980s, 2009, were caused by collapses in demand, by financial crises and by policy, not by a sudden jump in productivity; productivity was growing faster in the full-employment decades after 1950 than at any time before or since. France, Germany and the Netherlands have the same shape at a slightly higher level since the 1970s, for reasons economists usually trace to labour-market institutions rather than machines. Over a century in which an hour of work came to produce seven to ten times as much, the fraction of people who wanted a job and could not find one wandered and did not climb.
The reason is the one Keynes’s definition hinted at: the question is never only how fast machines destroy tasks but how fast new uses for labour appear. Four in ten American workers were on farms in 1900 and fewer than two in a hundred are now, and those thirty-eight in a hundred did not become unemployed; their grandchildren became electricians, nurses, programmers and the rest. David Autor and colleagues went through the census occupation lists and found that about six in ten Americans employed in 2018 worked in job titles that did not exist in 1940.[30] Daron Acemoglu and Pascual Restrepo formalise the race: automation displaces labour from tasks it used to do, and the creation of new tasks reinstates it; employment and the share of income going to workers hold up only when reinstatement keeps pace.[31] Autor’s own summary is that automation mostly complements labour, raises output in ways that lead to more demand for it, and that the fear of mass unemployment has been wrong because it counts only the first half of the process.[32]
This is not a story in which nobody lost. David Ricardo, who began by thinking machinery harmless to workers, added a chapter to the third edition of his Principles in 1821 admitting that it could make them worse off for a long time, and the handloom weavers of the next thirty years proved him right.[33] The Luddites were skilled men destroying machines that were destroying their trade, and they were correct about their trade. Acemoglu and Restrepo estimate that each additional industrial robot per thousand American workers between 1990 and 2007 reduced the employment-to-population ratio in the affected local labour markets by about 0.2 percentage points and wages by about 0.4 per cent.[34] The cost of technical change has been real and concentrated: on particular trades, towns and generations, and in the widening gap between what the highly educated and everyone else are paid. It has not shown up as a permanent rise in the unemployment rate. It has shown up as a question of who gets the output.
There is one well-known exception to the rule that displaced workers find new work, and it is the reason the argument is not settled. In 1983 Wassily Leontief pointed out that horses were once a central factor of production, that the arrival of the tractor and the car did not create new and better jobs for horses, and that the American horse population, having peaked around the First World War, fell by roughly nine-tenths over the next forty years.[35] Humans are not horses; they can learn new tasks, and they own the machines, and they vote. But the historical record that shows displaced workers always finding new work is a record of machines that were better than people at a limited range of things: lifting, spinning, calculating, repeating. The complement always lay in what machines could not do. The question for the twenty-first century is whether the range of things machines cannot do is still large enough.
What generative AI changes, and what it does not
The machines arriving since 2022 are aimed at the work that the last century moved people into: writing, summarising, coding, answering, drafting, advising. Studies that map what current models can do against occupational task lists find that a large majority of American workers have at least some of their tasks exposed, and that the exposure is highest among the educated and the well paid, the reverse of every earlier wave.[36] That is new. So is the speed at which the capability is spreading, which is closer to the speed of software than of steam.
The early evidence is mixed in the way the history suggests it would be. Inside firms, generative tools raise output per hour for some kinds of work substantially: in the best-known field study, customer-support agents with an AI assistant resolved about fourteen per cent more issues an hour, and the least experienced agents gained most, about a third.[37] At the edges of the labour market there are the first signs of displacement. Using payroll data, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen found that since late 2022 employment of workers aged twenty-two to twenty-five in the occupations most exposed to AI, software development and customer service among them, fell by about thirteen per cent relative to less exposed occupations, while older workers in the same jobs were unaffected.[38] Displacement at the entry level, where people learn the trade, is the form that matters most for the next twenty years. And in the aggregate statistics the signal is still faint: American productivity picked up in 2023 and 2024, Europe’s did not, and nobody can yet separate the machines from the post-pandemic rebound. Robert Solow’s line from 1987, that you could see the computer age everywhere but in the productivity statistics, applied for about eight years that time before the burst of the late 1990s.[39]
What the history of the last century predicts for the new machines is not reassurance and not doom but a reframing. The rate of technological unemployment has depended less on the machines than on demand, on how fast new work was invented, and on institutions: on whether the gains were spread widely enough that people had money to spend on the new things. The part of the gain that came as time was won by the people who worked, collectively, through law, bargaining and strikes. The rest arrived as income on terms they did not set: nobody asked to be paid instead of rested, and nobody who worked decided where that income went. If generative AI raises productivity growth back to the rates of the mid-century, the Keynes offer gets better every year: at three per cent a year, the output of a forty-hour week today can be had with a thirty-hour week in ten years. Whether anyone takes the offer is the same question it was in 1930, and the evidence of the last ninety-four years is that, left alone, people take roughly a fifth of it, and Americans about a tenth.
A short note from the other end of Europe
Portugal is the clearest example in Europe of the gap between where productivity went and where hours went. Output per hour rose about twelve-fold between 1930 and 2024, faster than in any of the large economies, most of it in the extraordinary years from 1950 to 1973, when it grew at seven per cent a year. It has grown at under one per cent since 2005. The Portuguese worker put in about 1,735 hours in 2023, some four hundred more than a German, ten working weeks, for an hour that produced a little under three-fifths as much.[2] The legal working week in the private sector came down from forty-four hours to forty only in 1996.[40] Unemployment went above sixteen per cent in 2013 and came down to between six and seven by the end of the decade, which is a story about the euro and the crisis, not about machines. And in 2023 the government ran a pilot of a four-day week with some forty private firms, coordinated by the economist Pedro Gomes, whose book makes the Keynesian argument that shorter weeks raise productivity and demand at once.[41] The firms volunteered, which is the main caveat on any result; the British pilot a year earlier had the same design and the same caveat, and most of its sixty-one firms kept the shorter week afterwards.[42] These are small, self-selected experiments. They are also the first serious attempts in half a century to take a larger share of the productivity dividend as time, by coordination rather than by each worker alone.
A choice, not a forecast
Keynes was right about the thing he thought was hard and wrong about the thing he thought would follow. The compound interest arrived on schedule; a century of productivity at about two per cent a year made the rich countries four to eight times richer, as he said it would, and the machines that were supposed to throw everyone out of work produced, instead, work nobody had imagined. What did not follow was the leisure. The rich world was offered the fifteen-hour week at the living standard of the late 1970s and took, by a margin of about four to one, the income. It would be more accurate to say the income was taken for it. In the United States, as the second part of this essay shows, the typical worker’s real wage barely moved for thirty-five years while the richest one per cent doubled their share of national income; the things that anchor a life, housing, health, education and care, grew dearer faster than wages; and the hours came as a package that few workers could refuse. Where the income was shared and the hours were limited collectively, people got more time, more of the income and longer lives. Keynes’s arithmetic was sound. What it left out was distribution.
It is easy to moralise about the result and Keynes did, with his line that the love of money would one day be recognised as “a somewhat disgusting morbidity”. It is more useful to notice what kind of choice it was. For most workers it was never an individual choice at all. Nobody can negotiate a fifteen-hour week alone and keep their place in a world of forty-hour weeks, and the gains that were taken as time were mostly taken together: the eight-hour day, the weekend, paid holidays, the forty-hour week, pensions. Each of them was fought for, legislated and bargained, as Figure 10 shows, and each looked at the time like a threat to prosperity that turned out not to be one. The next machine will make the offer again, and larger. How much of it is taken as time will depend, as it always has, on whether people decide collectively to take it.
Keynes ended his essay with a warning not to rush. “For at least another hundred years we must pretend to ourselves and to every one that fair is foul and foul is fair; for foul is useful and fair is not.” The hundred years are nearly up.
The next hundred years, roughly
Keynes gave himself a hundred years, and there are four left. On his first prediction the result is already in. If GDP per head keeps growing at its pace since 2005, by 2030 it will be about 8.3 times its 1930 level in the United States, just past the top of his range, 7.2 times in Germany, 6.2 in France and 5.3 in Britain. Every one of them will have landed between four and eight, or a whisker above.
On the second, the honest forecast is that the fifteen-hour week will not arrive by 2030, and on current trends not for a very long time after. Figure 11 extends each country’s hours per worker at the rate they have fallen since 1980. Germany, which has cut hours fastest, reaches Keynes’s 780 hours a year around 2107. France gets there around 2180, Britain around 2370. The United States, whose hours have hardly fallen in forty years, would take until the thirty-third century. At the long-run pace since 1870, which for every country but Germany is steeper than the recent one, the dates come closer together, between 2114 and 2255, but they all belong to Keynes’s great-great-grandchildren. These are extrapolations, not forecasts. Their point is that nothing in the present trend delivers his week; only a change in the trend would.
Artificial intelligence is the obvious candidate for such a change, and the range of serious estimates is wide. Daron Acemoglu, working up from task-level studies, puts the gain in total factor productivity over the next ten years at no more than about 0.7 per cent in all, and probably less.[43] Economists at Goldman Sachs estimated in 2023 that generative AI could add about 1.5 percentage points a year to productivity growth for a decade, about 16 per cent more output per hour by the end of it.[44] The arithmetic of this essay turns those numbers into hours. If Americans took the whole of the optimistic gain as time, their average week, 34 hours counted over the year, would fall to about 30. If they split it the way the typical rich country has split its gains since 1929, a fifth as time and four-fifths as income, it would fall by about one hour; on the American split, a tenth as time, by about half an hour. On Acemoglu’s estimate it would fall by minutes. The machines, even on generous assumptions, do not bring the fifteen-hour week by themselves. The IMF reckons that about 60 per cent of jobs in advanced economies are exposed to generative AI, roughly half of them in ways that could complement the worker and half in ways that could replace tasks, which makes the distribution question of the second part more pressing, not less.[45]
Demography pushes the other way. Across the European Union, Eurostat’s baseline projection has the number of people aged 65 and over per hundred of working age, 20 to 64, rising from 38 today to 55 by 2050 and 59 by 2070.[46] Hours per head of population will fall as that happens, but not as leisure for workers: the same output has to be produced by relatively fewer of them, which is why pension ages are rising almost everywhere. The coming decades may well give workers less time, over a lifetime, not more.
Politics will decide the split, as it always has. The legal working week has barely moved in most rich countries since the 1990s. The most direct recent attempt to move it, the Spanish government’s bill to cut the legal week from 40 hours to 37.5 without loss of pay, was voted down by the Spanish Congress in September 2025, by 178 votes to 170; the minister behind it has promised to try again.[47] The four-day-week trials in Britain, Portugal and elsewhere are small and self-selected, but they are growing, and most participating firms have kept some form of the shorter week.
So the things worth watching over the next few years are not the ones that make headlines. Whether productivity growth, not AI investment, actually rises in the statistics. Whether the median wage, which has grown faster since 2014 than in any decade since the early 1970s, keeps closing on productivity or falls behind again, as it did for thirty-five years before. Whether the price of housing, care and education keeps rising faster than pay. And whether any country legislates a shorter week and makes it stick. If all four go the right way, the fifteen-hour week stops being a curiosity of 1930 and becomes a political programme. If none does, Figure 11 is roughly what to expect.
Where this could go
Unpaid work across a life. The lifetime ledger here counts paid work only. The harmonised time-use microdata, which need registration, would allow unpaid work, care and schooling to be added age by age, for men and women separately, and would show whether the shift of women into paid work cost leisure or replaced housework.
The Portuguese basket. The repricing of 1977 is American. INE’s household budget surveys since 1980 and the consumer price index by category would allow the same exercise for Portugal, where housing and the minimum wage have moved very differently.
This piece was written collaboratively with Claude Opus 5.5 (Anthropic): human specification, editorial direction and critical review; machine data analysis, research and drafting. The series in the figures, and the numbers derived from them, are computed by scripts/keynes.py from public datasets: the Bergeaud–Cette–Lecat Long-Term Productivity Database (output per hour, hours per head and GDP per head, 1890–2024), Our World in Data’s series of hours per worker and the Maddison Project’s GDP per head, the Jordà–Schularick–Taylor Macrohistory Database for unemployment, and the OECD’s Pensions at a Glance database for exit ages. Results are in docs/keynes-results.json and the figures in docs/keynes-figures.html. The photographs in the body are public-domain images from Wikimedia Commons, credited in their captions; local copies are in assets/images/2026/. The shares in Figure 5 are ratios of log changes between 1929 and 2023. The career calculation in the retirement section assumes a man who starts work at 20 and leaves at the OECD’s average effective exit age, working the average annual hours per worker of the year; Portugal is left out of it because its hours series before 1990 is unreliable. The counterfactual weekly hours average annual hours over fifty-two weeks, holidays included, so they compare with the 34 hours a week that American workers average on the same basis today, not with a contracted working week. The outlook section’s dates are straight-line extrapolations of log hours, and its AI scenarios apply published estimates to American hours arithmetically; neither is a forecast. Hours before 1950 rest on a few benchmark years and should be read as approximate. The two follow-up sections, the repricing of the 1977 basket and the lifetime ledger, are computed by scripts/fifteen_followups.py (results in docs/fifteen-followups-results.json); the repricing uses fixed 1977 quantities at 2025 prices, and the ledger is a period calculation, like life expectancy, not the history of a real generation.
The cover photograph is a Bundy time clock, of the kind bus crews once used to record their times, preserved in West Bromwich; photograph by Andy Mabbett, via Wikimedia Commons, licensed CC BY-SA 3.0, cropped.
Authored by: Luis Matos Ferreira — Physicist, Developer, Writer
- Where the Hours Went — Part II: who got the income, what it bought, and where the hours of work actually go.
- The Long Retirement — exit ages, years after work, how healthy they are, and whether pensions are enough.
- The Two-Thirds Country — the same questions for Europe and Portugal: productivity, hours, the minimum wage, emigration and rent.
- The Treadmill — why modern economies seem to need growth: jobs, debts, pensions and politics.
- Growth of What — what GDP measures, what it misses, and Portugal measured other ways.
- The Fifth Day — Portugal’s four-day-week pilot and what the evidence from abroad says.
- The Manufacture of Wants — how the consumer economy taught people to want more, and whether it worked.
- The Portuguese Basket — the 1980 household basket repriced in hours of work.
- Bullshit Jobs — pointless jobs, and the slower colonisation of real jobs by pointless tasks.
- The Great Inversion — how AI inverted the relationship between producing and verifying.
- The Arithmetic of Bigness — Geoffrey West on why cities grow faster than exponentially and firms do not.
- The Second Derivative — another quantity whose growth rate matters more than its level.
- O Grande Debate — the argument over markets, states and who gets the output.
- The Sustainability Story — Portugal’s pension system: a contributory surplus, a closed civil-service scheme and thirty years of gloomy forecasts.
- The Duty to Work — how work became a duty, an identity and a source of status, and what it does to happiness.
- Earned and Unearned — the split of income between work and capital, who owns, and how each is taxed.
- Keynes, “Economic Possibilities for our Grandchildren”, The Nation and Athenaeum, 11 and 18 October 1930; reprinted in Essays in Persuasion, Macmillan, 1931. All quotations are from this essay. Its reception and a century of commentary are collected in Pecchi & Piga (eds.), Revisiting Keynes: Economic Possibilities for Our Grandchildren, MIT Press, 2008.
- Bergeaud, Cette & Lecat, “Productivity Trends in Advanced Countries between 1890 and 2012”, Review of Income and Wealth 62, 420 (2016); database version 2.7, February 2026, longtermproductivity.com.
- Maddison, Contours of the World Economy, 1–2030 AD, Oxford University Press, 2007; Jones, “The Facts of Economic Growth”, in Handbook of Macroeconomics vol. 2A, Elsevier, 2016.
- Solow, “Technical Change and the Aggregate Production Function”, Review of Economics and Statistics 39, 312 (1957).
- Goldin & Katz, The Race between Education and Technology, Harvard University Press, 2008.
- Dimitri, Effland & Conklin, The 20th Century Transformation of U.S. Agriculture and Farm Policy, USDA Economic Research Service, Economic Information Bulletin 3, 2005.
- Gordon, The Rise and Fall of American Growth, Princeton University Press, 2016.
- Bolt & van Zanden, Maddison Project Database 2023, via Our World in Data.
- Huberman & Minns, “The times they are not changin’: Days and hours of work in Old and New Worlds, 1870–2000”, Explorations in Economic History 44, 538 (2007); Feenstra, Inklaar & Timmer, “The Next Generation of the Penn World Table”, American Economic Review 105, 3150 (2015); combined series from Our World in Data.
- Boppart & Krusell, “Labor Supply in the Past, Present, and Future: A Balanced-Growth Perspective”, Journal of Political Economy 128, 118 (2020).
- US employment and population: 47.6 million employed in 1929 (Lebergott, in Historical Statistics of the United States) and a resident population of 121.8 million (Census Bureau); 161.0 million employed in 2023 (BLS, Current Population Survey, annual average) and 334.9 million residents (Census Bureau, vintage 2023). Hours per head are from the Bergeaud–Cette–Lecat database; the higher 1929 estimate multiplies Huberman and Minns’s hours per worker by the same employment share.
- US Bureau of Economic Analysis, National Income and Product Accounts, Tables 2.4.4U and 2.4.5U (price indexes and expenditures for personal consumption expenditures by type of product), August 2026 vintage; US Bureau of Labor Statistics, Current Employment Statistics, average hourly earnings of production and nonsupervisory employees, total private; repricing computed in the research notes kept with scripts/fifteen_followups.py.
- US Census Bureau, median sales price of new houses sold, 1977–2025, and Historical Income Table H-5; National Center for Education Statistics, Digest of Education Statistics, Table 330.10 (2024); Centers for Medicare & Medicaid Services, National Health Expenditure Accounts, 1960–2024.
- Knoll, Schularick & Steger, “No Price Like Home: Global House Prices, 1870–2012”, American Economic Review 107, 331 (2017); Davis & Heathcote, “The price and quantity of residential land in the United States”, Journal of Monetary Economics 54, 2595 (2007), and the updated land-price data of the AEI Housing Center.
- Boskin et al., Toward a More Accurate Measure of the Cost of Living, final report to the Senate Finance Committee, 1996; Moulton, “The Measurement of Output, Prices, and Productivity: What’s Changed Since the Boskin Commission?”, Brookings Institution, 2018, Table 5.
- OECD, Pensions at a Glance database (OECD Data Explorer, DSD_PAG@DF_PAG): average effective age of labour market exit and expected years after labour market exit, by sex, 1970–2024; current and future normal retirement age for a person entering the labour market at 22.
- Lee & Lee, “Human capital in the long run”, Journal of Development Economics 122, 147 (2016), combined with later Barro–Lee and UNDP estimates by Our World in Data (mean years of schooling, long run).
- United Nations, World Population Prospects 2024, abridged life tables; OECD, Labour Force Statistics by five-year age group and sex (employment-to-population ratios); Our World in Data, annual hours per worker; ledger computed in the research notes kept with scripts/fifteen_followups.py.
- Gimenez-Nadal & Sevilla, “Trends in time allocation: A cross-country analysis”, European Economic Review 56, 1338 (2012), Tables E3–E4 (Multinational Time Use Study, adults 21 and over including the retired).
- Ausubel & Grübler, “Working less and living longer: long-term trends in working time and time budgets”, Technological Forecasting and Social Change 50, 195 (1995), Table 1 and p. 204.
- Ramey & Francis, “A Century of Work and Leisure”, American Economic Journal: Macroeconomics 1(2), 189 (2009).
- Bowles & Park, “Emulation, Inequality, and Work Hours: Was Thorstein Veblen Right?”, Economic Journal 115, F397 (2005).
- Costa, “The Wage and the Length of the Work Day: From the 1890s to 1991”, Journal of Labor Economics 18, 156 (2000).
- Aguiar & Hurst, “Measuring Trends in Leisure: The Allocation of Time over Five Decades”, Quarterly Journal of Economics 122, 969 (2007).
- Baumol & Bowen, Performing Arts: The Economic Dilemma, Twentieth Century Fund, 1966; Baumol, The Cost Disease, Yale University Press, 2012.
- Prescott, “Why Do Americans Work So Much More Than Europeans?”, Federal Reserve Bank of Minneapolis Quarterly Review 28(1), 2 (2004).
- Alesina, Glaeser & Sacerdote, “Work and Leisure in the United States and Europe: Why So Different?”, NBER Macroeconomics Annual 20, 1 (2005).
- OECD/AIAS ICTWSS, trade union density (OECD Data Explorer, DSD_TUD_CBC@DF_TUD), % of employees, 1960–2024.
- Jordà, Schularick & Taylor, “Macrofinancial History and the New Business Cycle Facts”, NBER Macroeconomics Annual 31, 213 (2017); Macrohistory Database release 6, macrohistory.net. US 2019: Bureau of Labor Statistics, Current Population Survey, annual average.
- Autor, Chin, Salomons & Seegmiller, “New Frontiers: The Origins and Content of New Work, 1940–2018”, Quarterly Journal of Economics 139, 1399 (2024).
- Acemoglu & Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor”, Journal of Economic Perspectives 33(2), 3 (2019).
- Autor, “Why Are There Still So Many Jobs? The History and Future of Workplace Automation”, Journal of Economic Perspectives 29(3), 3 (2015).
- Ricardo, On the Principles of Political Economy and Taxation, 3rd ed., John Murray, 1821, ch. 31, “On Machinery”.
- Acemoglu & Restrepo, “Robots and Jobs: Evidence from US Labor Markets”, Journal of Political Economy 128, 2188 (2020).
- Leontief, “Technological Advance, Economic Growth, and the Distribution of Income”, Population and Development Review 9, 403 (1983).
- Eloundou, Manning, Mishkin & Rock, “GPTs are GPTs: Labor market impact potential of LLMs”, Science 384, 1306 (2024).
- Brynjolfsson, Li & Raymond, “Generative AI at Work”, Quarterly Journal of Economics 140, 889 (2025).
- Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, Stanford Digital Economy Lab working paper, August 2025.
- Solow, “We’d Better Watch Out”, New York Times Book Review, 12 July 1987, p. 36.
- Lei n.º 21/96, de 23 de Julho, which reduced the normal working week in Portugal to forty hours.
- Gomes, Friday Is the New Saturday: How a Four-Day Working Week Will Save the Economy, Flint, 2021; Gomes & Fontinha, reports on the Portuguese four-day week pilot, IEFP and Ministério do Trabalho, 2023–2024.
- Lewis, Stronge, Kellam, Kikuchi et al., The results are in: The UK’s four-day week pilot, Autonomy, February 2023.
- Acemoglu, “The Simple Macroeconomics of AI”, Economic Policy 40, 13 (2025); NBER Working Paper 32487 (2024).
- Briggs & Kodnani, “The Potentially Large Effects of Artificial Intelligence on Economic Growth”, Goldman Sachs Global Economics Analyst, 26 March 2023.
- Cazzaniga, Jaumotte, Li, Melina, Panton, Pizzinelli, Rockall & Tavares, “Gen-AI: Artificial Intelligence and the Future of Work”, IMF Staff Discussion Note SDN/2024/001, January 2024.
- Eurostat, EUROPOP2023 population projections, demographic balances and indicators (proj_23ndbi), baseline, old-age dependency ratio (65+ to 20–64).
- “Spain’s Congress rejects law to shorten work week”, The Local Spain, 11 September 2025; Reuters, “Spain’s Parliament Kills Shorter Work Week Bill in Blow to Government”, 11 September 2025.
Comentários
Enviar um comentário