Chapter 3 of 13

Scarce or Surplus

About 25 minutes to read

The robots in the care home

Between 2016 and 2020, the anthropologist James Wright spent eighteen months doing fieldwork in Japan, much of it inside a residential care home for older people. He had come to study a national experiment. Japan has the oldest population of any large country in the world, and for years its government had been warning that there would not be enough workers to care for the elderly. One answer, promoted by successive governments and supported with substantial public funding, was robots: machines that could lift residents out of bed, keep them company, lead their exercise classes and monitor them at night.

The care home Wright studied was trialling three such machines. One was a lifting robot, designed to help staff move residents between bed and wheelchair — a task that care workers perform many times a day, and that injures many of their backs. Another was a robotic baby seal called Paro, which responds to touch and voice and was intended to soothe residents with dementia. The third was a small humanoid robot called Pepper, programmed to lead recreational sessions.

What Wright found, and published in Robots Won't Save Japan (2023), was not what the robots' promoters had promised. The machines did not save labour. Very often, they created more of it. Staff had to move them, charge them, set them up, supervise residents while they were used, and deal with the confusion they caused. The robots could perform particular tasks, but they could not do what the work actually consisted of: noticing that a resident was frightened, judging whether she could stand that morning, joking with a man who had stopped speaking, holding a hand. The robots were not popular with the staff who were supposed to be relieved by them. Wright was not alone in his findings. According to a national survey cited in reviews of his book, by 2019 only around one in ten of Japan's more than nine thousand eldercare facilities had introduced care robots of any kind.

Japan's care homes are a small laboratory for the question at the heart of this chapter. The populations of most rich countries, and of China, are ageing. The number of people of working age is beginning to fall. At the same time, a new generation of machines — artificial intelligence above all — promises to take over much of the work that people do. Will the coming decades make human labour scarce, as the Black Death did in fourteenth-century Europe, and hand the 99 the strongest bargaining position they have had in centuries? Or will machines make human labour a surplus, and take away the oldest lever the many have ever had?

Under this book's assumptions — that AI has a wall, and that the wall is creativity — the answer leans towards scarcity. This chapter sets out the evidence for that, maps where the wall seems to be, and asks what the scarcity would be worth. It also faces the danger the book's assumptions create: if creativity becomes the scarcest good, the people who own it could become the next 1%.

The great demographic reversal

Start with the people.

For most of the past two centuries, the world's population grew, and with it the supply of workers. In the decades after 1980, as the first book described, the entry of China, India and the former Soviet bloc into the global economy roughly doubled the number of workers available to the world's employers. In their book The Great Demographic Reversal (2020), the economists Charles Goodhart and Manoj Pradhan argued that this enormous expansion of the global labour supply was one of the main reasons why wages in the rich countries stagnated, inequality rose and the bargaining power of labour collapsed from the 1980s onwards. The farm, in the language of the first book, had found a vast new herd.

That era, they argued, is ending. Birth rates have fallen below the level needed to replace each generation in most of the world's large economies. Japan's working-age population — people aged fifteen to sixty-four — peaked in the mid-1990s at around eighty-seven million and had fallen to under seventy-four million by 2024. China's working-age population peaked in the mid-2010s, and its total population began to fall in 2022. South Korea's fertility rate, at around three-quarters of a child per woman in 2024, is the lowest of any country in the world. Across the European Union, the number of people over sixty-five for every hundred people of working age is projected to rise steeply over the coming decades. The United Nations projects that the world's population as a whole will peak, probably in the 2080s, and then begin to decline.

At the same time, the demand for some kinds of work is set to rise sharply. Older people need more health care and more personal care. The World Health Organization has projected a global shortfall of around ten million health workers by 2030, concentrated in poorer countries. Goodhart and Pradhan's conclusion was striking: a shrinking workforce facing rising demand for labour would shift bargaining power back towards workers, push up wages, and reverse the trend towards greater inequality that had lasted for forty years. It was, in effect, a prediction that the force which levelled Europe after the Black Death was about to return — not through mass death, but through falling birth rates.

Japan's warning

Before accepting that prediction, there is a warning to consider, and it comes from the country where the reversal began earliest.

Japan has faced a shrinking workforce for three decades. Labour shortages have been reported across construction, transport, retail and above all care. If scarcity automatically produced bargaining power, Japanese workers should have been among the best placed in the world. They have not been. Real wages in Japan barely rose between the late 1990s and the early 2020s. A large and growing share of Japanese employees — around four in ten — work on "non-regular" contracts: part-time, temporary or dispatched through agencies, with lower pay, less security and little union representation. Employers responded to scarcity not by bidding up wages for everyone but by dividing the workforce into a protected core and a flexible periphery, keeping older workers on at lower pay, drawing more women into part-time work, and gradually admitting more foreign workers on restricted visas.

The first book has a name for what Japan illustrates. In Chapter 5 it described how the same labour scarcity after the Black Death freed peasants in western Europe and bound them more tightly in the east. Robert Brenner's explanation was that scarcity helps the many only where they are organised enough to use it. Where the owners can combine — with each other and with the state — scarcity becomes an incentive to tighten the fence. Japan's experience suggests that the same is true today. A shrinking workforce is a necessary condition for the return of labour's bargaining power. It is not a sufficient one.

The owners' three escapes

When labour becomes scarce, the owners of the farm have historically had three ways to escape the consequences. Each will be available in the coming decades, and each needs to be weighed.

The first is importing labour. The first book showed how Athens and Rome met their need for labour with captives and slaves, how the Atlantic plantations imported millions of enslaved Africans, and how the factories of the nineteenth century drew in waves of migrants who could be paid less and set against one another. Today, the richest countries meet much of their need for care workers, farm workers and construction workers through migration, often under visa schemes that tie workers to a particular employer and deny them the rights of citizens. In the Gulf states, where migrant workers form a large majority of the workforce, the "kafala" sponsorship system has long bound workers to their employers in ways that human rights organisations have compared to forced labour; some states have reformed it in recent years, with uneven results. Migration can be a breakout for the migrants themselves, many of whom escape far worse conditions. But as a response to scarcity, it can also be a way of importing a new, less protected herd — and the political backlash against it, in many rich countries, suggests that this escape will be contested.

The second is moving the work. Production that can be moved can follow cheap labour wherever it is, as the first book's account of globalisation showed. But much of the work that ageing societies need most cannot be moved. Care has to be done where the person being cared for lives. So does most construction, repair, transport, cleaning, cooking and teaching. The most offshorable work — manufacturing and some office tasks — has largely already moved.

The third is replacing the work with machines. This is the escape that matters most for the coming fifty years, and it is where this book's assumptions do their work.

Where the machines reach

What can artificial intelligence actually do to the demand for human labour? Estimates vary widely, and the technology is changing faster than the research can follow. But some careful studies give a sense of the scale.

In a paper published in Science in 2024, the economists and researchers Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock estimated how exposed different occupations in the United States were to large language models — the kind of AI that can read and write text. They concluded that around eighty per cent of the American workforce could have at least a tenth of their work tasks affected, and that around nineteen per cent of workers could see at least half of their tasks affected. Exposure was highest not in manual jobs but in well-paid office work: writing, analysis, programming, translation, administration. Studies by the International Labour Organization reached a similar conclusion for the world as a whole, finding that clerical work was by far the most exposed category, and that for most occupations generative AI was more likely to change jobs — to automate some tasks while leaving others — than to eliminate them entirely.

"Exposure" is not the same as replacement. As Chapter 1 noted, the most careful early study of AI used at work, by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, found that an AI assistant raised the productivity of customer-support workers by around fifteen per cent — by helping them, not by replacing them — with the largest gains for the least experienced. The economist Daron Acemoglu estimated that the AI technologies then available would add less than one per cent to total productivity in the United States over a decade. Others expect much larger effects as the technology improves. The honest summary is that AI will reach deep into many kinds of work, especially the routine cognitive work of offices — the work, it is worth noting, of the shepherds — and that how far it will go is unknown.

That last point matters for this book's design. The tasks AI handles most readily — keeping records, scheduling, coordinating, drafting, checking, allocating — are precisely the tasks that the 1% has performed for the 0.01% throughout history. The scribes of Uruk, the stewards of the medieval manor, the clerks of the industrial firm and the middle managers of the modern corporation all did versions of this work. If AI can do much of it, the shepherd class shrinks. That is the possibility on which the third element of the Federation hypothesis — AI replacing the shepherds — rests.

The long history of the end of work

Before going further, it is worth recalling how often the end of human work has been predicted, and how often the prediction has been wrong.

In 1930, as the first book recorded, John Maynard Keynes warned of a new disease he called "technological unemployment", even as he predicted that his grandchildren would need to work only fifteen hours a week. In 1964 a group of scientists, economists and activists sent a memorandum to President Lyndon Johnson, "The Triple Revolution", warning that "cybernation" — the combination of computers and automated machinery — would soon make large parts of the workforce unnecessary. Johnson's administration set up a National Commission on Technology, Automation and Economic Progress to investigate. Its report, published in 1966, concluded that technology eliminated jobs, not work: the demand for labour shifted from one kind of work to another, and the real danger was not a permanent shortage of work but the failure to help people move from the old jobs to the new.

In 2013, two Oxford researchers, Carl Benedikt Frey and Michael Osborne, estimated that around forty-seven per cent of American employment was at high risk of being computerised within a decade or two. Their study was widely cited as evidence of a coming jobs apocalypse. In the decade that followed, American employment rose to record levels. A study for the OECD in 2016 by Melanie Arntz, Terry Gregory and Ulrich Zierahn, which looked at the tasks within jobs rather than at whole occupations, put the share of jobs at high risk of automation in rich countries at around nine per cent — and many of those jobs changed rather than disappeared.

None of this proves that this time will be the same. The technologies of the 2020s can do things that those of the 1960s and 2010s could not, and some economists argue that AI is different in kind because it reaches into cognitive work that earlier machines left alone. But the history counsels humility in both directions: about the claim that AI will make human labour a surplus, and about this book's claim that it will not.

The shepherds' turn

If the evidence so far is a guide, the first people to feel AI's effects on their jobs will not be the care workers or the plumbers. They will be the shepherds.

In 2025, the economists Erik Brynjolfsson, Bharat Chandar and Ruyu Chen used payroll data from millions of American workers to look for the first signs of AI's effect on employment. In a working paper, they reported that since the spread of generative AI in late 2022, employment among workers aged twenty-two to twenty-five in the occupations most exposed to AI — such as software development and customer service — had fallen relative to less exposed occupations, while employment of older, more experienced workers in the same occupations had held up or grown. Their interpretation was that AI was substituting most readily for the codified, entry-level knowledge that young professionals bring to their first jobs, and less for the tacit knowledge that comes with experience. The findings are early and contested. But if they hold, they point to something important.

The first book described how the modern 1% is recruited through a meritocratic pipeline: elite schools, elite universities, and entry-level jobs in law firms, banks, consultancies and technology companies, where graduates learn the trade and rise. If AI erodes the bottom rungs of that ladder — the junior analyst, the trainee lawyer, the entry-level programmer — then the pipeline that produces the shepherds will narrow. The historian Peter Turchin, in End Times (2023), has argued that societies become unstable when they produce more aspiring members of the elite than there are elite positions to fill — a condition he calls "elite overproduction". A generation of highly educated young people finding that the jobs they were trained for have been taken by machines could become, in the terms of the first book, a body of disaffected shepherds. The first book noted that many of history's great breakouts were led by exactly such people: lawyers, clerks, junior officers, journalists, teachers.

That is a danger to the stability of the farm. It is also, for the 99, a potential opportunity. If the managing class shrinks and loses its confidence in the meritocratic story, the split between the owners and their shepherds — the fourth of the five forces — may widen. Chapter 5 will return to it.

The strongest case: agents that write code

The sharpest test of the wall is in software, the field where AI is most capable and where, as the payroll data above suggest, young professionals have felt it first. By 2025 the leading AI companies were selling coding "agents": systems that take a description of what is wanted, write the code, run it, read the errors and try again, often for long stretches without a person watching. The research group METR found that the length of software tasks such agents could complete half the time had doubled roughly every seven months for six years. Some engineers now describe their work as directing several agents at once rather than writing code themselves.

Two other findings complicate the picture. In a randomised trial published in July 2025, METR gave experienced open-source developers real tasks from their own projects, allowing AI tools on some and forbidding them on others. With the tools, the developers took about nineteen per cent longer, although they believed the tools had made them about twenty per cent faster. And when researchers at Carnegie Mellon built TheAgentCompany, a simulated small software firm in which agents had to browse, code, calculate and message simulated colleagues, the best agent completed only about thirty per cent of the tasks on its own. Both studies date quickly; the agents of 2026 do more than those of 2025. But both point to the same place.

What the agents do not do is decide what is wanted, judge whether what they produced is right, or answer for it when it is not. People set the goal, review the result and step in when the agent goes wrong. A "swarm" of agents working together needs someone to define the work and to check it. That is the refinement this book's assumptions now make: agents are powerful but not fully autonomous, and they remain driven by human beings. The wall, in software, is not a list of tasks the machine cannot do. It is the direction and the responsibility, the judgement and trust layers of the map below.

That has a consequence for the shepherds. The routine coordination that filled the days of many managers is exactly what agents can absorb. But directing agents is itself a managing function. A person who can direct a swarm of agents can do the work of a team, which could concentrate the managing role in fewer hands than ever, a new and smaller shepherd class, or spread it to anyone with access to capable agents. Which happens depends partly on who owns the agents, and partly, as later chapters argue, on whether open models keep capable agents within everyone's reach.

Mapping the wall

This book assumes that AI, however capable, meets a wall. Where is it?

The note that set out this book's plan named five layers, in order: creativity first, then care, judgement, physical work and trust. Each can be examined against the evidence, and each turns out to rest on a different kind of argument.

Creativity. The fourth standing assumption holds that AI cannot originate the genuinely new — new questions, new forms, new institutions — in the way that people can. Chapter 1 explained this in terms of Margaret Boden's distinction between combinational, exploratory and transformational creativity. Some recent evidence is suggestive. In a 2024 experiment published in Science Advances, Anil Doshi and Oliver Hauser asked several hundred people to write short stories, some of whom were given story ideas generated by an AI system. Independent readers judged the AI-assisted stories more creative and better written on average, especially those by less creative writers. But the AI-assisted stories were also more similar to each other than the stories written without help. Individually, the writers were better off; collectively, the pool of new ideas had narrowed. That is what one would expect if the machine were very good at recombination but drew all its users towards the same centre. It is consistent with this book's assumption, though it does not prove it. The law, for now, takes a similar view: in the United States, the Copyright Office and the courts — including the federal appeals court in Thaler v. Perlmutter in 2025 — have held that copyright protects only works with a human author.

Care. Care is the clearest case in the evidence so far, and James Wright's care home is its emblem. Caring for a sick child, a frail parent or a person with dementia is not only a set of tasks — lifting, feeding, washing, administering medicine — but a relationship, built on attention, trust and judgement about a particular person in a particular moment. Machines can help with some of the tasks. The relationship, so far, they have not provided, and the people receiving care, as well as those giving it, have generally not wanted them to.

Judgement. In 1966 the philosopher Michael Polanyi observed that "we can know more than we can tell": much human skill rests on tacit knowledge that people cannot fully articulate, and therefore cannot easily program into a machine. The economist David Autor later called this "Polanyi's paradox" and used it to explain why computers had automated routine tasks while leaving untouched many jobs that required adaptability, common sense and judgement. Modern AI, trained on vast amounts of human-generated data rather than on explicit rules, has eroded part of Polanyi's paradox; it can now do some things that no one could explain how to program. But judgement in novel, high-stakes situations — a doctor deciding between treatments for a patient whose case fits no textbook, a judge weighing a sentence, a teacher deciding how to handle a crisis in a classroom — remains, on present evidence, an area where people are reluctant to defer to machines, and often right to be.

Physical work. In the 1980s the roboticist Hans Moravec noted a paradox that has since borne his name: it is comparatively easy to make computers perform at adult level on intelligence tests or games, and very hard to give them the perception and mobility of a one-year-old. Forty years later, AI systems can write essays and solve mathematical problems, but robots that can reliably fold laundry, fix a leaking pipe in an old house or help an elderly person out of the bath remain far from ordinary use. Physical work in unpredictable environments — plumbing, electrical work, construction, repair, cleaning, care — is likely to remain human for longer than much office work. AI itself depends on such work. Its data centres, whose electricity use the International Energy Agency expects to more than double between 2024 and 2030, are built by electricians, construction workers and engineers, connected to grids that people maintain, and cooled with water that communities supply. The machine at the top of the wall stands on physical work at its foot.

Trust. Finally, there is work whose value depends on the fact that a human being does it and is accountable for it. Psychologists have documented what they call "algorithm aversion": in experiments by Berkeley Dietvorst, Joseph Simmons and Cade Massey, people who saw an algorithm make mistakes lost confidence in it far more quickly than they did in a human making the same mistakes. Laws are moving in the same direction. The European Union's AI Act, adopted in 2024, requires that high-risk AI systems be designed so that they can be effectively overseen by people. Someone, in the end, has to answer for a decision. As long as that is true, there will be human work that consists of being the person who answers.

This map is not a forecast that any of these layers will hold for ever. Some may fall sooner than this book assumes. But it gives the argument of this chapter its shape. If the wall holds roughly where the book's assumptions put it, then the work that remains human in the coming decades will be concentrated in creativity, care, judgement, skilled physical work and trust — and the demand for much of it, driven by ageing societies, will rise.

What scarcity would be worth

What bargaining power would that give the 99?

The answer depends on who is scarce. Some of the work behind the wall is already well paid and highly organised: senior doctors, judges, architects, engineers. Much of it is not. The care workers who are most clearly protected by the wall are among the lowest-paid workers in the rich world, disproportionately women and migrants. In the United States, home health and personal care aides have long been among the lowest-paid occupations in the country. The economist Nancy Folbre, in The Invisible Heart (2001) and later work, argued that care work is systematically undervalued because it was historically done unpaid by women within families, because its benefits spread widely and are hard to capture in a price, and because those who do it are often reluctant to withdraw their labour from the people who depend on them. Scarcity has not, so far, made care workers rich.

But there are signs of change where scarcity meets organisation. In December 2022, members of the Royal College of Nursing in England went on a national strike for the first time in the union's hundred-year history. In October 2023, around seventy-five thousand health care workers at Kaiser Permanente in the United States held what was described as the largest health care strike in American history, and won substantial pay increases. As Chapter 2 showed, the tight labour market after the pandemic raised the wages of low-paid workers faster than those of the well paid for the first time in decades. These are small signs. They are consistent with the first book's lesson: scarcity creates the possibility of bargaining power, and organisation turns the possibility into wages.

There is a further complication, which matters a great deal for this book's design. Much care, health and education work is paid for, directly or indirectly, by the state. When care workers bargain, they bargain with public budgets, and the people on the other side of the table are not the owners of the farm but taxpayers — including other members of the 99. Whether the scarcity of care work translates into better pay therefore depends on politics, and on whether the society is willing to tax the farm's owners to pay the people who look after its old.

The new aristocracy of creativity

The book's assumptions also create a danger, and it would be dishonest not to face it here.

If creativity is the scarcest and most valuable contribution of all, then those who possess it — or those who own its products — could become the next 1%, or even the next 0.01%. History offers a clear warning about how this might happen.

In 1981 the economist Sherwin Rosen published a paper called "The Economics of Superstars". He observed that in some markets — music, entertainment, sport, publishing — small differences in talent could produce enormous differences in income, because technology allowed the best performers to reach the whole market at little extra cost. A singer in 1800 could perform only for those who could hear her in person; a singer in 1980 could sell records to millions. The result was a winner-take-all market, in which a few stars earned fortunes and most performers earned very little. Digital technology has made such markets more extreme. In music streaming, a small fraction of artists receive the great majority of income.

AI could intensify this pattern dramatically. If machines can reproduce, adapt and distribute creative work at almost no cost, the value of a creative work will lie increasingly in its originality — and in the rights to it. The first book showed again and again that the people who captured the surplus were not usually the producers but the owners: not the farmers but the landlords, not the weavers but the mill-owners, not the musicians but the record labels. There is every reason to expect the same in an economy of creativity. The new aristocracy would be less likely to consist of creators than of those who own the rights to their creations, the platforms that distribute them and the models trained on them.

The law of intellectual property shows how such ownership can be extended. In the United States, copyright now lasts for the life of the author plus seventy years; the term was last extended in 1998, by a law nicknamed by its critics the "Mickey Mouse Protection Act" because it was strongly supported by Disney. The earliest version of Mickey Mouse, from the 1928 cartoon Steamboat Willie, finally entered the public domain on 1 January 2024 — ninety-five years after its release. In the language of the first book, intellectual property is a form of enclosure: the turning of a common resource — ideas, stories, images, music — into private property with a fence around it.

The danger, then, is not that people who create will be rewarded. Under this book's first assumption — reward weighted by contribution — they should be. It is that creativity, like land before it, will be enclosed: that the rights to it will be concentrated in a few hands and held for generations, and that the creative contribution of the many will be fed into the machines and the platforms of the few. Chapter 11 will have to answer this danger directly, through open access to tools and education, time for everyone to create, and limits on how long a creation can be owned.

The tests

This chapter tests two parts of the Federation hypothesis: the premise behind the fourth element — that every person can be a worker, because human work remains necessary — and the fifth, reward weighted by contribution, in an economy where creativity is scarce.

Precedent. Labour scarcity has shifted the ratio before, after the Black Death, but only where the many were organised to use it; where they were not, as in eastern Europe then and arguably in Japan now, scarcity tightened the fence. Winner-take-all markets for creative work, from the record industry to streaming, show that creative scarcity tends to enrich owners of rights and platforms more than creators.

Evidence. Demographic data make a shrinking workforce in most large economies close to certain. Studies of AI's effects suggest wide exposure but, so far, more augmentation than replacement, and the evidence from care, robotics and trust is consistent with a wall roughly where this book assumes it to be. The evidence on creativity — such as Doshi and Hauser's finding of individual gains but collective narrowing — is suggestive and early.

Capture. The farm has three escapes from scarcity: importing a less protected workforce, moving the work, and machines. If the wall holds, the third is limited and the second is largely used up, which leaves the first — and the enclosure of creativity through intellectual property, which would let a new owning class capture the scarce good itself.

Exit. Scarcity strengthens the worker's power to leave: to quit, to refuse, to move. That is the oldest form of breakout in the first book, and the one the post-pandemic labour market briefly revived.

The ledger

What survives. Under the book's assumptions and on the demographic evidence, human work remains necessary, and in important areas — care, skilled physical work, judgement, trust and creativity — it is likely to become scarcer. The premise that every person can be a worker survives. So does the possibility that the second of the five forces, scarcity, will be available in the 99's favour for the first time in centuries.

What has to change. Scarcity does not deliver power by itself. It must be combined with organisation (Chapter 4) and locked in (Chapter 6), or it will be absorbed by the owners' escapes, as Japan's experience suggests. The design must also be changed to meet the danger this chapter has identified: reward for contribution, in an economy where creativity is the scarcest good, risks producing a new aristocracy of creators and — more likely — of those who own the rights to creation. The Federation must treat the tools, training and time needed to create as part of the common floor, and must limit how long and how widely creative work can be enclosed.

What remains unknown. Whether the wall holds where this book assumes it does, and whether agents remain driven by people or learn to set and judge their own work. If AI proves able to provide care, exercise judgement or originate genuinely new ideas, the map in this chapter changes, as Chapter 1 explained. Also unknown: whether the politics of migration will allow the rich world to meet its need for care workers on fair terms, and whether societies will be willing to tax the farm to pay the people who look after their old.

Bridge: the scattered

The care workers who stand behind the wall are, for the most part, not gathered together in great factories. They work in private homes, small care homes and hospitals; many are employed through agencies or platforms; many are migrants, speaking different languages, on different visas. The delivery riders, drivers and freelancers of the digital economy are scattered further still, managed by algorithms rather than foremen, and often legally classified as their own bosses. The first book found that the factory, for all its cruelty, did one thing for the 99: it brought them together, where they could see each other and act together. The next chapter asks whether the scattered workers of the twenty-first century can be organised — and what the new forms of organisation have actually won.