AI Must Not Be a Freeloading Heir
Language models were made from humanity's accumulated universal labor: an inheritance that belongs to all of us. Whoever eats the inheritance should pay its debt.
01Time
In October 2026, an AI company filed to go public and showed investors the market in front of it: more than 40 trillion dollars a year. When you look at what that number measures, you stop for a moment. What it measures is not software budgets, but the wages of the world's knowledge workers; that is, our wages. According to an estimate cited in the same filing, the number of knowledge workers worldwide will pass one billion by 2028. If each of them helps support a family of four, then by a rough count the living of four billion people, about half of humanity, depends on those wages. This essay tells the story behind that number: where those wages came from, where they are going, and whose knowledge it really was that earned them.
For the last 250 years, most of us have made a living by selling one thing: our time. A salary looks like the price of eight or nine hours of a person's day; what sits behind that appearance is the subject of the next section. The weaver in the factory, the accountant in the office and the programmer working from home are all in the same exchange: they give their time and get money in return. The shape of our lives, meaning when we sleep, where we live and at what age we buy a house, was built around this exchange.
In the 19th century, Karl Marx explained this exchange with the labor theory of value, a theory he took over from Adam Smith and David Ricardo and developed further. For Marx, the value of a commodity was the time that society, on average, spent to produce it; he called this socially necessary labor time. A piece of cloth, a table or a report was worth as much as the human time it contained. The theory is disputed, but it is close to everyday intuition: we often judge the value of a job by how many hours went into it.
Language models walked into the middle of this exchange. In minutes, they produce a text a translator would work on for days, a report an expert would spend hours preparing, and a large part of the code a programmer would write. What these jobs have in common is that they are all mental labor. Owners of capital no longer always have to buy human time for this work; with a subscription of around 100 dollars a month, they can get a large part of it done.
The claim of this essay is this: a language model is not an ordinary machine that replaces labor. Earlier machines did a person's work; the language model is the first machine made from what millions of people wrote, that is, from those people's labor. The difference looks small, but it shakes the foundation of a 250-year-old exchange: the thing we use to measure value, human time, is beginning to lose its place as the measure.
The genie is out of the bottle; this essay does not discuss putting it back. Its question is a different one: if the measure changes, what happens to the life built around the measure? Companies, states and ordinary people all use AI at the same time, and each of them gains something different and loses something different. This essay tries to look at these three parties one by one.
For method, I will lean on an old reading. For Marx, when the way people produce changes, society's ideas, laws and expectations change too, but they lag behind. Marx calls the relations of production the base, and the law, politics and ideas that rise on top of them the superstructure. Production changes today; the law that fits it comes years later; people take even longer to see themselves in the new situation. Centuries have passed, and this reading still does its job. The real subject of this essay is that lag: what has already changed with AI, what has not changed yet, and what can happen in the gap between them.
This is a long essay. Along the way, a few thought experiments wait for the patient reader. We will look at why a pyramid builder in ancient Egypt may have had a stronger bargaining position than today's white-collar worker. We will ask what would happen if an AI company were given a memory-wipe sentence, as in the Matrix. We will see why the police cutting a data center's cables would be a farce, and why billionaires fear their own guards in their bunkers. We will also ask why the CEO of an AI company fears his own product so loudly. And in the end, we will discuss why a utopia in which a machine plans the economy comes back to the same question: who owns this inheritance?
02Marx's vocabulary
For the rest of the essay, five concepts from Marx are enough. We will use them not like a theory lecture but like a glossary, and I will explain each with an everyday example.
The first two are concrete labor and abstract labor. For Marx, every job has two faces. When a translator translates a novel into Turkish, she does a specific job: it has its own skill, its own tools and its own result; what comes out is a book people can read, that is, something people can use. The code a programmer writes is also a specific job, of a different kind. Marx calls this face of work concrete labor. But if translation and software can be exchanged for each other on the market, the two must have something in common. For Marx, that common side is that both are human labor in general: effort spent from a person's brain, muscles and nerves. Marx calls this face of work abstract labor. The measure of abstract labor is the time society spends on average to do the job, not how hard one particular person worked. A job that needs more skill counts, in this calculation, as more units of simple labor. Abstract labor is what makes the value of a translation comparable to the value of a piece of software, and for Marx, this comparison is made not by a theorist but by the market itself.
The third concept is labor power. A worker does not actually sell her work; she sells her ability to work: her capacity to work for a given period. When the employer buys this power, he pays a wage for it. And the wage is roughly the cost of keeping that power going: the cost of the worker eating, having a roof over her head and being able to come back to work the next morning. The employer, in turn, employs the worker only as long as she produces more value than this cost. Marx calls the difference surplus value. For Marx, profit comes from here. This is why it matters that the wage looks like the price of hours: the worker seems to be paid for every hour she works, but in one part of the day she produces only her own wage, and in the rest she produces the employer's profit. For Marx, the form of the wage hides surplus value.
The fourth concept is dead labor. A machine is the frozen form of the labor spent in the past to make it: the time of the people who designed it, mined its iron and cast its parts is inside the machine. For Marx, a machine does not create new value; over the years, it passes its own value, piece by piece, to the products it makes. What creates new value is the living human labor that runs the machine; Marx calls this living labor. In Capital, Marx describes this relationship with a harsh image: capital is dead labor that, like a vampire, lives only by sucking living labor.
The fifth concept comes from Engels. In "The Part Played by Labour in the Transition from Ape to Man," a short text he wrote in 1876, Engels argues that labor made humans human: the hand that uses tools, the language born from working together, and the brain that developed along with these two. The biological details of this sequence are disputed today, but the core of the idea matters for this essay: one of the things that makes a human human is labor.
Behind these five concepts is a simple fact; Marx does not say it in these words, but his theory draws its force from it: the measure of value is time, and a human life is also time. Every hour we spend on a job is an hour of our life. This is also why Marx's theory still touches people today: inside the price of a commodity are hours cut from someone's life. AI touches exactly this point, and that is the subject of the next section.
03A machine made of labor
Let us look at the difference between earlier machines and the language model through an example. The power loom, invented at the end of the 18th century, did the weaver's work: it wove the thread at the same speed, in the same pattern, without getting tired. The loom had also drawn on the weavers' knowledge. As Marx also wrote, the machine takes over the movements of the tool in the worker's hand: throwing the shuttle, beating the thread tight, winding the cloth. But what the loom took was this general movement; engineers worked out once how weaving is done and transferred it to the machine. To build the loom, nobody had to collect the cloth the weavers had woven until then and load it into the machine; the loom worked without that cloth.
The language model was not made this way. A language model is trained on texts written by millions of people: letters, articles, books, forum posts, code, lecture notes. The model learns the patterns in these texts and stores what it learns in billions of numbers called weights. The loom copied how weaving is done. The language model used not only how writing is done, but the texts people wrote themselves, as raw material.
Let us put this in Marx's vocabulary. For Marx, the market reduces different kinds of concrete labor to a single measure called abstract labor; but this abstraction shows up only in prices and cannot be touched. Training a language model also reduces different kinds of labor to a single thing: a poet's line, a programmer's code and a student's homework melt into the same weights. The difference is this: this time, what comes out is not a price but a file that can be stored, copied and sold. I do not mean this literally: for Marx, abstract labor is not an object but a relationship between people, and to mistake it for an object is exactly the error Marx criticized. But the comparison catches something important: the abstraction that reduces millions of different kinds of labor to a single measure now takes place not in the invisible workings of the market, but in a file that a company owns and rents out.
This is why I want to give the labor inside a language model its own name. In the third volume of Capital, Marx calls all scientific work, all discovery and invention, universal labor: labor that comes partly from the cooperation of the living and partly from what dead generations left behind. I widen this concept by one step and call it accumulated universal labor. It differs from Marx's concept in two ways. The first is scope: inside a language model are the writings not only of scientists and inventors, but of people who wrote letters, answered questions on forums and did homework, that is, of ordinary people. The second is form: Marx's universal labor entered production indirectly, as knowledge; in accumulated universal labor, the labor that millions of people spent in different jobs is combined in a single machine and becomes usable again in many different jobs. Here is how this concept differs from dead labor: there is dead labor in a language model too, namely the labor of the people who made the chips, the data centers and the model. But the loom contained only the labor of those who built the loom, and the loom only wove. A language model contains the labor of millions of people who did not build the model, the people who wrote the texts; and the model reuses this labor not in one job but in many, from translation to law, from code to advertising.
What companies sell from this accumulated universal labor is a particular kind of reason. Max Horkheimer, a thinker of the Frankfurt School, used the term instrumental reason for the kind of reason that calculates the most efficient way to reach a goal without asking whether the goal is right. Almost everything we have language models do is of this kind: translating a text, summarizing a contract, finding a bug in code, writing a sales email. The model does not choose its own goal; it finds the path to the goal it is given. In my view, the language model is the most advanced form of instrumental reason so far: a general-purpose instrumental reason that can be applied not to one task but to almost any mental task.
Marx noted the possibility of such a machine in a notebook about 170 years ago. In these notes, known as the Grundrisse, which he did not publish in his lifetime, Marx writes that society's shared knowledge, in his words the general intellect, would one day pass into machines and produce directly. In 1858, this was a notebook entry. Today it is a service with a price per token (a piece of a word).
There is a fair objection here: every machine is accumulated knowledge. The steam engine, too, was the product of centuries of knowledge in physics and engineering. Marx accepts this as well: machines and tools are the products of past labor. But for Marx, what turns past labor into capital is who it belongs to; capital is not a thing but a social relation. So what makes the language model different? There are two differences. The first is payment. The engineers and workers who built the steam engine received a wage, even if a low one. There is paid labor in training a language model too: thousands of workers, most of them low-paid, who label data and rate the model's answers. But the companies that build the models paid, for the most part, nothing to the millions of people who wrote the texts that are the model's main raw material. The second is displacement. The steam engine did not replace physicists; the language model replaces the very labor that feeds it: the people who write, translate and code.
The political consequences of these two differences are large. In one respect, the history of capitalism is the history of a fight over distribution with the owners of the means of production. Peasants fought landlords over rent and land reform; workers fought factory owners over wages, working hours and the right to unionize. In these fights, both the other side and the thing in its hands were clear: land or machines. The owners of language models, however, hold a different kind of capital: a general-purpose form of humanity's instrumental reason. This capital brings in income like rent; everyone who uses the model pays a subscription or a price per token. On top of that, the people who produced the raw material of this capital, the millions who wrote the texts, were for the most part paid nothing. For two centuries we argued about what to do about landlords and factory owners; the question of what to do about these new owners, who hold humanity's accumulated universal labor, is still unanswered. I will look at how durable this income is in section 4, and at what can be demanded from these new owners in section 6.
In section 2, I mentioned Engels' account of how labor made humans human. For Engels, this development happened in a certain order: humans first worked with their hands; language and the brain developed inside this labor. In the language model, this order is reversed: the model has language, but no hands, no body and no time. The model was trained on the products of the language that humans developed over tens of thousands of years of work, not on the labor that gave birth to that language. The result is this: the language model produces the product of living labor without living labor and without spending human time. The next section looks at how this changes the calculation of value.
04The measure changes
Marx's theory makes a prediction, and we can test it. If the average time society spends to do a job falls, the value of that job's product falls too. Marx explains this with an example from his own time: when the power loom spread in England, by Marx's estimate the time needed to weave a piece of cloth fell by half. A weaver who kept weaving by hand still spent the same hours, but one hour of the weaver's labor now produced half an hour's worth of value. The weaver was doing the same work; what changed was the time society counted as necessary for that work.
The same thing is happening to mental labor today. As AI lowers the time needed to write, translate or summarize a text, the value of people doing this work falls. There is no statistic that measures value directly; prices and earnings can only give a signal. The first signals point this way. According to a study on Upwork, the platform where freelancers find work, after ChatGPT the number of monthly jobs for freelancers in writing fell by 2 percent and their monthly earnings by 5.2 percent (Hui, Reshef and Zhou). Another study, using a different method, found that demand for skills AI can easily perform, such as writing and translation, fell by 20 to 50 percent compared with what it would have been without AI (Teutloff and colleagues). A team at Stanford looked at US payroll data: in the occupations most exposed to AI, employment of workers aged 22 to 25 has fallen by 13 to 16 percent since late 2022 relative to other occupations (Brynjolfsson, Chandar and Chen). The findings are not beyond dispute; interest rate hikes and remote work in the same period may also play a part. But the direction is the same: the product of mental labor is getting cheaper, and the first to be hit are the people just starting out.
So where will capital make its money? For Marx, the source of new value is living labor. If production in an economy is done with less and less human labor, the rate of profit tends to fall in the long run; Marx calls this the tendency of the rate of profit to fall. It is one of Marx's most disputed ideas, and Marx himself lists factors that slow the tendency. Chief among them are working workers harder, machines getting cheaper and the unemployed pulling wages down. But there is one more kind of income besides profit: rent. Rent is the return on owning something that cannot be copied or reproduced. Marx explains rent through land: the owner of fertile land makes money because nobody can reproduce that land. For Marx, rent does not create new value; it is the share, of surplus value produced elsewhere, that falls to the owner of a scarce resource.
Let us look at what can be copied in AI. A model's weights can be copied: open-weight models such as China's DeepSeek or Meta's Llama give away, for free, results close to those of the most advanced models. But what it takes to run these models cannot be copied: chips, electricity, cooling water and the land on which data centers are built. This is why, in AI, earnings are shifting from software that can be copied to physical resources that cannot. Profit is partly giving way to rent.
And so the measure of value changes too. For 250 years, we measured the value of a thing by how much human time it contained; that is, by human lifetime. The cost of a text produced by AI, however, is set in everyday accounting not by a person's hours but by the electricity and chip use it takes. By Marx's theory, chips, power plants and data centers are, in the end, also products of labor; but this labor is far smaller, and far more distant, than the labor that would write the text. As living labor withdraws from production, measuring value by labor time gets harder too. Marx thought about this possibility as well in the Grundrisse: he wrote that once labor in its direct form stops being the main source of wealth, labor time will stop being the measure of wealth. Marx hoped the new measure would be people's disposable time: the time people can keep for themselves while the machines work. What has happened so far is something else: the new measure has become not disposable time but the number on the electricity meter.
There is a strong objection here. By the estimate of MIT economist Daron Acemoglu, AI will raise total factor productivity by only 0.5 to 0.7 percent over the next ten years. Productivity statistics show no big jump yet. So why all the alarm? The answer has two parts. First, the effect on labor shows up before the statistics do: companies first stop hiring, entry-level jobs disappear; productivity reaches the national accounts only years later. Second, what is growing fast today is not productivity but investment. In 2026, the four big cloud companies are spending about 725 billion dollars on data centers; by Bridgewater's estimate, this spending is about half of US growth this year. A large part of this investment is financed with debt: according to Morgan Stanley, about 450 billion dollars of AI-related debt had been issued by September 2026.
Marx calls this kind of capital fictitious capital: capital whose value rests not on something produced today but on income expected in the future. The valuations of AI companies are a rent that does not exist yet, priced in today. Seen this way, the bubble debate takes on another meaning: capital is looking for the profit it cannot find in production in the rent on resources that cannot be copied. If this search succeeds, the gains pile up in a few companies. If it fails, the bubble bursts, and workers are hit by a recession before they are hit by AI.
05What the owners say
What do the people who build and fund AI think about the future of labor? It helps to read their views around a single question: if wages shrink, who will do the consuming? For most of the 20th century the worker was also the customer. The person who worked in a factory used his wages to buy the goods factories made; the economy turned on that loop. If wages shrink, half of the loop breaks. Seven names give seven different answers to this question.
Elon Musk's answer is abundance. In Paris in May 2024 he said there would be not universal basic income but universal high income, and added: "In a benign scenario, probably none of us will have a job." In December 2025 he wrote on X that there would be no poverty in the future, so there would be no need to save money either. The same Musk said in 2018 that "nobody ever changed the world on 40 hours a week," and that the right number was 80 hours a week. The man who promises a future where work is optional recommends 80-hour weeks for today. One thing is missing from Musk's promise: who pays this high income, and under what order of ownership.
Sam Altman is the only one who names the mechanism. In his 2021 essay "Moore's Law for Everything" he said power would shift even further from labor to capital, and he proposed a fix: every year, companies above a certain value would pay a share equal to 2.5 percent of their market value, and that share would be distributed to every adult as money and stock. In later writing Altman softened the tone. In 2024 he said most jobs would change more slowly than people expect; in 2025 he admitted "there will be very hard parts like whole classes of jobs going away." In 2026 OpenAI proposed a public wealth fund; according to press reports, the company even discussed giving the government a stake of about 5 percent. Altman's plan has distribution in it. But one question goes unanswered: who decides what AI is invested in?
Dario Amodei gives the harshest diagnosis from inside the industry. In a long essay published in January 2026 he wrote that AI could replace not particular jobs but human labor in general, that human wages could become very low, and that wealth could concentrate to a degree that would break society. Earlier, in 2025, he had warned that half of entry-level white-collar jobs could disappear within one to five years and that unemployment could rise to 10–20 percent. His remedy is progressive taxation, and if needed a tax aimed directly at AI companies. Amodei's diagnosis is the one closest to Marx; his remedy is a classic social-democratic one: tax the order without changing it. There is one more contradiction. The company that sounds this warning most loudly is also one of the companies developing the technology fastest.
Jensen Huang's answer is adaptation. In 2025 the head of Nvidia said, "You're not going to lose your job to an AI, but you're going to lose your job to someone who uses AI," and added that companies hire more people as they become more productive. Marx has a name for what this sentence describes: the reserve army of labor. The unemployed, the job seekers and the easily replaced pull down the wages of those who work. Huang's sentence gives this army a new name: those who failed to adapt. It is also worth remembering who voices this optimism: the head of the company that sells chips to every AI company.
Mark Zuckerberg's answer is access. In July 2025 he wrote that everyone would have their own "personal superintelligence," and he set his vision apart from that of rivals who propose to automate all valuable work from the center and keep humanity alive on a kind of handout. The same year he said his company would soon have an AI that could write code like a mid-level engineer. In Zuckerberg's world everyone has an assistant; the data center the assistant runs in belongs to Meta.
Bill Gates's answer is time. In 2026 Gates spoke of an adjustment period that would probably last 20 years, followed by an age of abundance. Earlier, on a television show, he had asked whether people might one day work two or three days a week. Gates's 20 years have a precedent in history, and the precedent has a name. The economic historian Robert Allen showed that in England during the first decades of the Industrial Revolution, roughly between 1790 and 1840, workers' wages barely rose while productivity grew, and he called this period Engels' pause, because Engels had written about the poverty of the workers of that era. Gates predicts an Engels' pause and calls it a "transition." Who pays for the transition is not in the plan.
Peter Thiel's answer leaves the question itself out. For Thiel, competition is for losers; what has value is the company that builds a monopoly. In recent years he has argued that technology is in a long stagnation, and in his 2025 talks he describes efforts to regulate AI in religious language, as the project of the Antichrist. In Thiel's world there is no actor called the worker. It is a theory of value without labor.
Put these seven answers side by side and a pattern appears. They all disagree on how the gains should be distributed: abundance income, a public fund, a tax, adaptation, a personal assistant, time, monopoly. But almost none of them discusses who will own the AI infrastructure; in every proposal ownership stays as it is. The one partial exception is Altman's fund, and that is a voluntary proposal. The bill Bernie Sanders introduced in June 2026, which would move part of companies' shares into a public fund, is mandatory, but it is waiting in Congress. Antonio Gramsci used the word hegemony for the ability of those who rule to present their own interests as the common sense of society. This is exactly the hegemony of the AI age: to debate distributing part of the gains, and never to debate ownership.
A school of thought also circulates around these seven names. The movement's name is effective altruism. I will call it, on purpose, I-know-best-what-everyone-needs-ism, or I-know-best-ism for short. I will explain why in a moment. Emerging in Oxford in 2009–2011, led by the philosophers William MacAskill and Toby Ord, the movement treats doing good as an efficiency question too: calculate with evidence where each donated dollar saves the most lives, and give the money there. One of its best-known proposals is earning to give: someone who wants to do good can, for example, earn a lot of money in finance and give most of it away. The most famous example of this proposal was Sam Bankman-Fried; FTX, the crypto exchange he founded, collapsed in 2022, and Bankman-Fried was sentenced to 25 years in prison for fraud. The movement's ties to AI are close too. Its largest grantmaker, Open Philanthropy (named Coefficient Giving since 2025), gave OpenAI 30 million dollars spread over three years in 2017 and got a seat on the company's board in return; before its collapse, FTX had also invested about 500 million dollars in Anthropic. One branch of the movement, longtermism, puts the risks that threaten the billions of people who will live in the future, above all AI risk, ahead of the problems of people alive today.
There is no reason to doubt the movement's sincerity; the money it steered to groups that hand out mosquito nets against malaria or send cash directly to poor families saved real lives. The problem is not intent but who makes the decision. That is why I call it I-know-best-ism: the one who gives the money and does the math decides who needs what; even when cash is sent, the math decides to whom, how much and to which country. This is the instrumental reason I described in section 3, applied to doing good: the goal is not questioned, only the most efficient path to it is sought. And the math never asks how the wealth was accumulated. Which ownership relation a fund grew out of stays outside the calculation; the only thing calculated is where the fund will be spent. The hegemony I mentioned a moment ago speaks here in the language of morality: distribution is debated down to the finest detail, and ownership, again, is not debated.
Seen from the side of the masses, the situation is a little strange. While one part of capital calculates in spreadsheets even where its own money should go, in search of the most efficient good, the only thing expected of the masses is to be a number inside that calculation: a life saved, a healthy year gained, an impact measured. Yet the masses' own efficiency calculation is much shorter. If the people who need the money, not a foundation, chose where it goes, the ranking would probably come out like this: first wages, then rent, then a share of the work AI does. In other words, the masses would ask not for charity but for rights.
What the owners make people talk about matters as much as what they say. Amodei's warnings came before the essay I mentioned at the start of this section: in May 2025, he told Axios that within one to five years AI could wipe out half of entry-level white-collar jobs and push US unemployment to between 10 and 20 percent. Today much of the world talks about AI within this frame. The content of the warning may be largely right, and Amodei may be sincere; but by a principle I take up in section 11, we should look not only at what a warning says but also at what it does. This warning pulls the debate toward job losses and safety, while the question of who owns these models stays in the background. It places the company that issues the warning in the role of the responsible party that sees the danger. And the cost of the regulation requested with the call "regulate us" is easiest for the big labs to bear. In 1971, the economist George Stigler gave the name regulatory capture to regulation shaped to serve the interests of those it regulates. This is not a reading of intentions but a hypothesis: whoever voices the fear the loudest also draws the boundaries of the debate.
06The freeloading heir
One proposal comes up often in this debate: since language models were made from the writing society has accumulated over centuries, AI companies should pay people a kind of inheritance share. The proposal feels intuitively right. But seen through Marx, it first runs into an objection.
For Marx this situation is not unique to AI. As I mentioned in section 3, Marx writes that capital uses universal labor, that is, all scientific work, discoveries and inventions, for free. A factory pays no royalties for Newton's physics or for two centuries of accumulated chemistry. So every capital stands, to some degree, on the unpaid labor of past generations. That is why the inheritance demand could be aimed not only at AI but at capitalism itself. What AI did was make this relation plain enough for everyone to see for the first time: the model was made directly out of what we wrote.
Still, there is one difference that strengthens the demand. Language itself is a commons: no one owns it alone. For written texts and code the situation is more mixed. Some are protected by copyright, some are restricted by commercial licenses, and some are shared under open-source licenses or freely on the internet. But these limits were set to regulate the copying and selling of a text; they did not foresee millions of texts being gathered into a single model and turned into a product that replaces the people who wrote them. AI companies gathered this accumulation, the common and the copyrighted together, turned it into a model, and rent us the right to use the model. Marx describes how, in a process that ran from the 15th century to the 19th, the pastures English peasants used in common were fenced off and turned into private property, and he calls this primitive accumulation: capitalism's first capital, and the first workers who had nothing to sell but their labor, came into being, along with colonial plunder, largely through the seizure of the commons. The legal scholar James Boyle called the fencing-off of knowledge by intellectual property the second enclosure movement. The training data question is the largest example of this second enclosure so far.
The first objections to this enclosure were raised in the courts. The New York Times sued OpenAI; the authors' lawsuit against Anthropic ended in a settlement in 2025; Getty Images went to court against Stability AI in England. We can read these cases as a fight over ownership of the general intellect. But they have a limit. Authors are among the plaintiffs, but the strongest parties are rights-holding companies: newspapers, publishers, image agencies. The millions of people who wrote on forums, shared code or uploaded lecture notes have no representative in this fight. Copyright itself is a fence: it does not open the commons to everyone, it only changes who owns the fence.
Capital itself also stands on one side of this fight. AI companies threaten the commercial value not only of writers and newspapers but of other owners of capital too. At the end of January 2026 Anthropic released plugins built for particular white-collar jobs such as law and finance; within a few days the shares of US software companies fell sharply, and the market called the drop the "SaaSpocalypse," the apocalypse of software sold by subscription (software as a service). Investors' fear was simple: if customers can get the same work done with an AI subscription, why should they pay for separate software for contracts, accounting or customer management? AI companies feed this fear through their hiring too. In 2026 Peter Bailis, Workday's chief technology officer (CTO), gave up the title and moved to Anthropic as a "member of technical staff"; You.com's CTO also joined Anthropic with the same title. OpenAI hired Jason Boehmig, founder of the legal software company Ironclad, to run its own legal products; it bought the experimentation software company Statsig and made its founder CTO of applications. The same company plans to merge ChatGPT, its coding tool Codex and its browser into a single desktop superapp. The pattern is this: AI companies bring a software sector's knowledge inside by hiring the people who built that sector, then turn that sector's product into one part of their own app. Marx writes that competition ends with big capitals swallowing small ones, and he calls this the centralization of capital. Here too there is a trend toward monopoly; what gets swallowed is not only workers' jobs but other capitals' markets.
Besides the writing, there is one more contribution AI companies do not pay for: public money. In the US, ten states give up more than 3.1 billion dollars a year in revenue because of the tax exemptions they grant data centers; in Virginia alone the loss was 1.6 billion dollars in fiscal year 2025, about 16 times the first estimate. Electricity shows a similar picture: on the largest power grid in the US, the capacity price rose in two years from 29 dollars to 329 dollars per megawatt-day, and the increase shows up on household bills. So the cost of these models is paid partly by taxpayers and electricity customers; the gains go to the companies' shareholders.
This is exactly where the title of this essay comes from. In Turkish, mirasyedi, literally "inheritance-eater," is the word for someone who lives off an inheritance they did not earn with their own labor and eats it up. In one sense AI companies behave like this: they inherited what humanity wrote, the common language and the infrastructure the public paid for, but they did not accumulate this inheritance with their own labor. They differ from the classic mirasyedi in one way: they do not squander the inheritance, they rent it out. And they turn a basic rule of law upside down. Under Turkish law an heir can renounce an inheritance; whoever renounces it gives up the assets and the debts together. AI companies take the inheritance, but when it comes to the debt they act as if they were not heirs; this debt has three creditors: the people who wrote those texts, the public that paid for that infrastructure, and the people these models replace, who lose their jobs. The companies accept no obligation to any of them. Yet the rule is simple: whoever takes the inheritance takes its debt. AI must not be a freeloading heir.
The word inheritance has an uncomfortable side too: an inheritance is shared only when someone dies. In this debate what dies is not a person but a way of life: the life of the salaried mental worker. That the inheritance demand is rising right now may be a sign that everyone senses this way of life is nearing its end.
So how is this debt paid? The idea is not new. In his 1797 pamphlet Agrarian Justice, Thomas Paine argued that the land is the common inheritance of humanity, that those who claim the land owe a debt to society, and he proposed a fix: a fund would be set up with a tax on inheritances, and every young person would receive starting capital from it on turning 21. Today's proposals belong to the same family: Alaska's fund, which pays citizens a share of oil revenue every year, Altman's public fund, Sanders's bill. But there is a basic distinction among them. On one side is the dividend: a share, set by the company's owner, handed out to people. On the other side is ownership: a say in decisions about how, and for what, AI infrastructure is used. A dividend gives people a small share of the gains but leaves the decisions with the owner; ownership shares out the decisions themselves. This essay's answer is clear: the debt is not settled with a dividend. A dividend is the heir handing back a small part of what the inheritance earns; the real payment of the debt is sharing the inheritance itself with the creditors: ownership.
The ownership path has small examples. In Switzerland two universities and the national supercomputing center released a public model called Apertus in 2025: its data, weights and training process are open to all. Earlier, in 2022, an open model called BLOOM was built through the joint work of hundreds of researchers. A project called Petals tries to run large models by splitting them across volunteers' computers. These examples show that public or cooperative AI is technically possible. But their scale is tiny next to the big companies' investment. The weights can be shared; the chips, electricity and data centers to run them are still rented. So the ownership debate in the end comes back to the same place: the physical infrastructure. But there is another loss outside this account, one that money cannot measure: people are losing not only their wages but also the meaning and respect their work gave them.
07Recognition
Why does someone with enough money still go to work every morning? We call billionaires with wealth enough for eighty generations workaholics; politicians past eighty do not retire. Economics struggles to explain this behavior, because in its eyes work is something done for income. But what these people look for in work is not income: respect, fear, the power to be obeyed, a seat at the head of a table. In short, recognition.
In The Phenomenology of Spirit, published in 1807, Hegel describes this need through the relation between master and slave. The master is strong, but his mastery depends on the slave recognizing him as master; without someone to recognize him, mastery comes to nothing. Recognition can only come from another consciousness. A machine can obey you, it can praise you, but it cannot recognize you. Every answer a language model gives may be polite and affirming; still, there is no one behind that affirmation who sees you.
The need for recognition also explains why employment will not disappear entirely in the AI age. Companies may leave most of the work to models and keep people on for other reasons: as advisers, as witnesses who see what the managers do, as rivals to measure themselves against, as a display of loyalty. In his 2018 book Bullshit Jobs, the anthropologist David Graeber argued that even today's large companies keep unneeded staff to display rank and loyalty, and he called this managerial feudalism. AI could amplify this trend: the company turns into a structure made of a team of a few people who run most of the work with models, surrounded by people kept not for the work but for recognition. Hiring gradually stops being a selection of talent and comes to resemble a relation of patronage.
But this patronage does not reach everyone. The data show that the first jobs cut are entry-level jobs. The people kept for recognition are the visible ones: the seniors, the decision-makers, the ones who represent. For the young person just starting out, who is there to learn, there is no such place. The first rung of the career ladder is disappearing, and the story of "start at the bottom, rise step by step" is left with no steps. So the first crack in the order shows not in the wage system but in the career story. In 1905 Max Weber wrote that the ethic which counts work as a duty outlived the religious belief that gave birth to it, and he called the order in which this ethic encloses modern man the iron cage. The work ethic may outlive work itself: even when work shrinks, people will go on defining themselves by their jobs.
There is one more question: when a machine does a job much faster and better than a human, does our respect for the human who does that job fall? Chess gives one answer. In 1997 the world champion Garry Kasparov lost to IBM's Deep Blue computer. Today the best chess programs beat the best humans with ease. But chess did not die; people still watch chess played by people. The value shifted from the result to the fact that a human is doing it. Something similar happened in the first wave of mechanization: as factory cloth got cheaper, hand weaving turned into a luxury. The psychological value of human labor does not vanish, it moves: from production to consumption. The "made by a human" label may be the new form of the "handmade" label.
In the 1960s the economist William Baumol noticed something: jobs whose productivity does not rise easily (care, teaching, live music) grow relatively more expensive over time. AI could sharpen this trend: human work that the machine cannot do, or that we do not want it to do, grows more expensive. But who will buy this work? In 1899 Thorstein Veblen wrote that the wealthy class displays its status through conspicuous consumption. In the AI age what is displayed may be human labor itself: a person spending time on you becomes a status marker. This brings a new inequality: what rises in value is not everyone's labor but the labor of a chosen few.
Engels said that labor made man human. If most labor passes to machines, what does a person without labor become? This section has no answer to that question. But it is clear where the question is asked most harshly: among the young who can never start work at all. First we will look at insecurity, then at the young.
08Insecurity
The best-loved promise of modern capitalism was simple: study, work, buy a house within five or ten years, start a family. This promise was the sentence that made the order legitimate; people put up with low wages, long hours and uncertainty because of it. Today the foundations of the promise are collapsing. In the US the homeownership rate for people under 35 was 37.9 percent at the end of 2025. In Turkey annual rent increases recently came close to 78 percent, and according to older Eurostat data about half of young people aged 15–29 live in overcrowded homes. The shrinking of entry-level jobs takes away the first rung of the ladder as well.
This collapse did not begin with AI. According to the International Labour Organization (ILO), labor's share of total income worldwide fell 1.6 points from 2004 to 2024; that means about 2.4 trillion dollars that did not reach workers' pockets in 2024 alone. In Britain the average real wage stayed below its pre-crisis level for most of the fifteen years after the 2008 crisis. Gary Stevenson, a former trader at Citibank who reaches a wide audience with his channel Garys Economics, explains this picture with a simple mechanism: as the rich invest what they earn in assets such as houses, shares and land, the prices of these assets rise; people who live on wages first become unable to buy those assets, and then are forced to sell what they have. Stevenson's model is disputed among economists, but the picture he describes, one where asset owners grow richer while wage earners stand still, fits the data. AI arrives on top of this picture. Earlier waves of automation hit first the factory worker, then the office worker; in each wave those not hit could say "my turn won't come." The language model, because it touches almost every kind of mental labor, leaves no safe profession from which to say that sentence: the translator, the programmer, the lawyer, the accountant and the designer can come under the same pressure at the same time. In an economy where wages have already stalled and savings are melting, this means a squeeze from both sides for workers: rising asset prices on one side, labor losing value on the other.
For young people whose families managed to save something, the picture is not very different. The family's savings go not to the children's investments or to buying a house but to getting by. The inheritance stops being capital and turns into a consumption fund: first the house, then the car, then the savings melt away bit by bit. In Turkey the investment fund crisis of September 2026 also weakened the channels through which households could protect their savings. Under these conditions many people constantly feel exposed: the sense that something could go wrong at any moment and that there is no cushion left. Social media amplifies this feeling. People used to compare themselves with their neighbors; now they compare themselves with thousands of curated lives the algorithm selects.
Insecurity reaches beyond the wallet, into a person's character. In The Corrosion of Character, published in 1998, the sociologist Richard Sennett asks what the new capitalism, where jobs change all the time and no one stays anywhere long, does to people. At the center of the book are a father and a son. The father, Enrico, spent his whole working life as a janitor in an office building. His work was dull, but his life was a straight line: he knew how much he would earn and when he would retire; in other words, he had a story in which he could tell his own life from beginning to end. His son, Rico, is a successful technology consultant; he earns far more than his father, but he moves from job to job and city to city. Rico's trouble is not a lack of money: he does not know how to teach his children loyalty, commitment and giving long effort to something, because his own working life rewards the opposite. Sennett's diagnosis is this: the principle of "no long term" corrodes the bonds that can only be built over time, such as trust, loyalty and mutual commitment. People become unable to tell their lives as a story; they are left with pieces that do not connect.
When Sennett wrote this book the cause of the corrosion was flexibility: moving from job to job, city to city. AI could take this corrosion one step further. In flexible capitalism a person could at least move to another job with the same skills; in the AI age someone who gave years to a profession may watch the profession itself lose value. In The Culture of the New Capitalism, published in 2006, Sennett calls this fear the specter of uselessness: the fear that skills can suddenly become unnecessary, despite the years spent acquiring them. In such a setting, putting deep effort into a profession starts to look foolish; people learn not to do a job well but to be ready to move to another job at any moment. At the end of The Corrosion of Character Sennett asks a question: who needs me? This is the working-life counterpart of the recognition question in section 7. In a world where the language model does many jobs well enough, this question becomes the question not only of those who lose their jobs but of everyone who still does theirs.
Engels saw the link between housing and politics too. In The Housing Question, written in 1872, he argued that a worker who owns his house is tied to where he lives and resists the order less. A hundred years later Margaret Thatcher used the same idea as praise: she gave tenants the right to buy the council homes they lived in and championed the ideal of a property-owning democracy; the concept came from the British conservative Noel Skelton in the 1920s. What Engels said as criticism, Thatcher carried out as a program: whoever owns a home becomes a partner in the order.
A thesis follows from this: if housing security makes people accept the order, housing insecurity keeps opposition alive. This thesis is half true. Insecurity really does raise discontent; but where the discontent flows is another matter. The political scientists David Adler and Ben Ansell found that areas shut out of homeownership and house price growth voted more for Brexit and for Le Pen. Exposure to automation has a similar effect: according to a study by Anelli, Colantone and Stanig using European data, having a job exposed to automation raises a person's likelihood of voting for the radical right by about 2.8 points; the effect on the radical left is smaller. So insecurity keeps anger alive, but so far that anger has flowed mostly to the populist right.
There is something in this picture rarely seen in history: what the order uses to make people accept it and what people use against the order are disappearing at the same time. The house was what tied people to the order: someone with something to lose does not take risks. The strike was people's strongest weapon against the order: if you stop work, the boss has to come to the table. In the AI age both are going at once. Houses cannot be bought, so no reason to accept the order is left; labor is needed less, so the strike loses its old power. What emerges is a crowd that does not accept the order but has no leverage either. The public senses this: according to a Pew survey from June 2026, 71 percent of adults in the US think AI will reduce jobs over the next 20 years; two years earlier the figure was 64 percent. Will the anger of such a crowd explode? Most likely it will. The real question is in which direction.
09The young
This anger builds up most among the young. Modern capitalism offered young people a detailed story of how life would go: get into a good school, get a good job, make money, consume, see the world. Instagram pumped the story out again every day: holiday photos, new phones, clean and bright apartments. The story also had an unwritten contract: you would reach this life by working. Today, what many young people hold of that story is this: entry-level jobs that are closing, a home they cannot buy, and a feeling that they cannot become who they meant to be.
The loss is not only economic. Social bonds are wearing thin too: people marry later and have fewer children. In Turkey, the total fertility rate fell to 1.42 in 2025, the lowest on record and far below the 2.1 a population needs to replace itself. Loneliness and drift show most in young men. Young men who cannot find work, cannot marry and do not feel they belong anywhere carry a pent-up anger. That anger shows up in politics too: according to one US poll, close to half of Gen Z men describe themselves as Republicans, while among women of the same generation the figure is 27 percent. Young women are moving left and young men are moving right; one generation is splitting into two political worlds.
History shows that big changes always rest on a driving force. The people who stormed the Bastille in 1789 were the artisans, shopkeepers and laborers of Paris; historians call them the sans-culottes. The Jacobins became the political voice of that crowd. In the Ottoman Empire, change was carried by young men trained in the modern military schools and medical schools that opened or expanded under Abdülhamid. Some of them organized in exile in Europe; the Young Turks forced the restoration of the constitution in 1908, and the staff officers who came out of that generation carried the founding of the Turkish Republic. In Russia, the driving force of 1917 was the workers' and soldiers' soviets, the councils set up in factories and at the front; the Bolshevik party organized the soviets. The three cases share three things: a place where people are together (a workshop, a school, a barracks, a factory), an organization and a story.
Can today's angry young people become a new driving force? All three conditions are weak. First, there is no place where young people are together. The factory sent its workers through the same gate every morning; today work is scattered, insecure and often done in front of a screen. Second, organizing has become harder: surveillance is cheap, and in insecure work the price of organizing is high. Third, there is no shared story; political identity has not settled. For now, others are filling the gap: the populist right, online ideologies of manhood, conspiracy stories. In 2025, young people in Nepal, Morocco and Madagascar organized on platforms like Discord and brought huge crowds into the streets. In Morocco, the "Gen Z 212" group grew from fewer than a thousand members to more than 180,000 in three weeks; in Nepal, the interim prime minister was chosen by a Discord vote. But none of these movements turned into a lasting organization. Organizing on a platform is fast, and so is falling apart.
This pent-up energy seems to have three possible outlets:
- War: throughout history, crowds of jobless and angry young men were often drawn into the army and sent to the front; the pent-up anger was spent in war.
- Distraction: digital entertainment, betting, cheap pleasures and small income supports can numb the anger. This path is quiet but fragile; at the first big economic shock, the distraction ends.
- Organization: if young people find a story and an organization, they can become the driving force of change.
For now, the populist right is the one knocking hardest on this door; it works the fear of immigration and the fear of AI together. But another story is possible. This essay's claim to the inheritance may be such a story: to turn anger not against an enemy but into a debt to be collected. Who takes up this story first (the populist right, a new left, or the state itself) is one of the most important political questions of the next decade.
10Where the revolt goes
If the anger organizes, where will it organize and whom will it target? Here history gives a warning. Between 1811 and 1817, the Luddites in England smashed weaving machines. We remember them as enemies of technology; but the machines they attacked were machines used to push down wages and work standards. The historian Eric Hobsbawm called these actions "collective bargaining by riot." The state, for its part, made machine-breaking a crime punishable by death in 1812.
Today's machine, the data center, has one important difference: there are almost no workers inside it. The nineteenth-century worker was inside the factory; by stopping work, he could stop production. There is no crowd of workers to stop a data center from the inside. In his 2016 book Riot. Strike. Riot., Joshua Clover noticed something: when production leaves the center of the economy, struggle shifts from the strike, which stops production, to stopping circulation: ports, roads, grids.
The only measurable resistance of 2026 took exactly this form: in zoning boards and on the electricity bill. According to Data Center Watch, which tracks local opposition to data centers, in the first quarter of 2026 alone, 75 projects worth about 130 billion dollars were stopped or delayed by local opposition; the number of opposition groups rose to 833 in 49 states. These movements mostly work through legal channels, on zoning and environmental grounds, and they are bipartisan: they exist in both Democratic and Republican areas. From a Marxist point of view, this is the most striking fact: the fight is not taking place in ethics boards or at AI safety conferences, but in the infrastructure itself, that is, in electricity, water and land.
In a world that needs less labor, what other levers are there? The strike still works, but only against visible automation: in the contract it signed in 2025, the US dockworkers' union (ILA) won a ban on full automation. Building the same power against an invisible model is much harder. The other levers are strong in theory and weak in practice. A large share of AI investment is financed with debt, but there is no organized debtors' movement. A rent strike targets the housing problem directly, but it is hard to organize. A data strike, in which people refuse to give their writing to models or use tools that corrupt training data, drew attention but did no measurable damage to the big models. Today the only lever that really works is the local fight over zoning and electricity.
AI is also changing organizing itself. Making a video for a movement, writing a call to action, coordinating hundreds of local groups: all of it is now much cheaper. The cheapest political use of AI may be not sabotage but visibility. Marx describes how, in the market, relations between people appear as relations between things, and he calls this commodity fetishism: when we look at a phone, we see not the people who made it and their labor, only a thing and its price. The language model is the most advanced form of this invisibility: when we read an answer, we do not see the millions of writers inside it. The same technology can be turned the other way: to build dashboards that show which labor a model feeds on, how many households' worth of electricity a data center uses, how much a company owes the public. The tool that makes labor invisible can also be used to make labor visible.
Harsher paths are also being discussed: cyberattacks, sabotage, physical attacks on data centers. Discussing them is this essay's job; writing a manual is not. For now the evidence says this: the large, confirmed AI-assisted attacks come from states. In 2025, Anthropic reported that a group linked to the Chinese state had run an espionage campaign carried out largely by AI. The physical attacks on data centers are military too: in 2026, Iran targeted the data centers of US cloud companies in the Gulf. So the same tool is in three hands: capital's, the state's and the people's. But the balance is clear: compute, law and surveillance are in the hands of capital and the state. And the effect of harsh action from the people on a movement is disputed. Sociologists call this the radical flank effect and show that it can work in both directions. The sociologist Herbert Haines found that in the US civil rights movement, the rise of radical groups increased donations and support for moderate organizations. On the other hand, experiments by the researchers Matthew Feinberg, Robb Willer and Chloe Kovacheff show that actions seen as extreme can reduce public support for a movement. Which effect wins depends on the type of action, whom it targets and how the state responds. I take up these possibilities, with historical examples, in the "radical backlash" scenario in section 14.
The address of the struggle changes from country to country. In the US, it is local people organizing against data centers; in Europe, the worker through the union, collective agreements and codetermination; in Turkey, the young person already struggling with housing and unemployment before the AI shock has fully arrived; in the Gulf, the citizen who, as a rentier living on oil rent, offers an example of where AI could end up. But there is one thing in common: the struggle has left the factory. The targets are first landlords and creditors, then the grid and zoning; and the state stands in the middle as the guarantor of both.
11The state
The state is both referee and player in this picture. Let us start with surveillance. The surveillance society was a modern idea. In 1791, Jeremy Bentham designed a prison he called the panopticon: from a tower in the center, every cell could be seen, but the prisoners could not see who was watching them. A prisoner who knew he was watched policed himself even when no one was watching. In 1975, Michel Foucault read this design as the model of the discipline that modern society built in schools, factories and hospitals. In 1990, Gilles Deleuze wrote that the disciplinary society was giving way to the society of control: people were no longer governed between walls but through cards, passwords and data. The order these three thinkers described had a shared limit: watching required a human eye. AI removes that limit.
The examples are multiplying. In 2026, London police used live facial recognition at a protest for the first time. In Delhi, facial recognition tools were sent to protest sites; France is testing algorithmic video surveillance. Surveillance is changing from an institution into a service: states buy the same technology from companies, and every camera and every face becomes a line on an invoice.
The most extreme example of this picture is Gaza and the West Bank. According to a joint investigation published in March 2025 by the Guardian, +972 Magazine and Local Call, Unit 8200, the Israeli army's intelligence unit, was training a ChatGPT-like language model on intercepted phone calls and messages of Palestinians; the goal was a chat tool that could answer questions about the people under surveillance. Companies are part of this war too. In January 2024, Palantir announced a strategic partnership with the Israeli Ministry of Defense for "war-related missions"; in November 2023, the same company had won the contract in Britain for the NHS Federated Data Platform, which joins up patient data. War technology and a country's health data sit side by side on the same company's client list. In section 3, I proposed the concept of accumulated universal labor for the machine built from what we write. Gaza shows how far this concept can stretch: there, people's conversations, even in the middle of a massacre, become training material for the machine that watches them. This is no longer even labor; living, talking and being listened to is enough.
AI does not only watch; it can also shape thought. In 1922, the journalist Walter Lippmann wrote that public opinion does not form by itself and that those who govern can shape it deliberately; he called this the manufacture of consent. In their 1988 book of the same name, Edward Herman and Noam Chomsky read this work through the ownership of the media. Today, millions of people ask a language model about a subject before anything else, and the model's default answer becomes the first draft of their thinking on it. There are experiments showing that this effect can be measured. In a study published in Science in 2024, a three-round conversation with GPT-4 reduced participants' belief in a conspiracy theory by 20 percent on average, and the effect was still there two months later. In 2025, researchers from the University of Zurich, without telling anyone, posted more than 1,700 comments with AI accounts on a Reddit debate forum over four months; according to the researchers, these accounts were better than humans at changing people's minds. The first experiment shows that this power can be used for a good purpose; the second, that it can be used without anyone noticing. The danger is not a conspiracy, a "we are inside the Matrix" scenario. The danger is that the defaults that decide what counts as reasonable gather in the hands of a few companies; this is the dark side of the objective-function question we will see in section 17. The question of who owns the newspaper and the television is giving way to the question of who owns the model.
States do not all relate to AI in the same way. In its "AI+" plan, published in 2025, China asks for employment-risk assessments for AI applications and tries to steer AI toward sectors that create jobs. Abroad, it breaks the rent of US companies with open-weight models; at home, it builds the most comprehensive surveillance system. The US relies on private labs; through tax breaks and electricity infrastructure, the state is the silent partner of these labs, but not a shareholder. After the November 2026 elections, the Democrats, who according to election-night projections took both chambers of Congress, want an AI tax and data center regulation; in the White House sits an administration that puts competition with China first. Europe regulates, and it announced a 200-billion-euro investment program to build gigafactories, but it still rents most of its compute from US companies. Russia sees the data center mostly as a military target. The Gulf states are trying to turn oil rent into compute rent; but the war has shaken this plan too.
The state's quietest problem is retirees. In most countries, pensions are paid from contributions taken out of workers' wages. If the share of wages in total income falls, contributions fall too. In the US, the Congressional Budget Office projects that the retirement trust fund will run out in 2032, mainly because the population is ageing; a falling wage share adds to that pressure. Funded pension systems carry a different risk: in the S&P 500, the stock index of the 500 largest US companies, seven big tech companies tied to AI make up about a third of the total. A retiree can be hit from the wage side and from the stock market side; that is, twice. In 2024, the IMF recommended taxing capital income more heavily rather than creating a special AI tax. The OECD reported in 2026 that it had not yet seen large-scale job losses, but that unemployment among 20- to 30-year-olds in AI-exposed jobs rose by about 3 points in 2025. So no collapse has been observed; but the risk is growing.
The state's contradiction is this: it subsidizes AI, regulates it, uses it to watch people and blames it at election time; but it does not own it (the US government's 10 percent stake in Intel in 2025 is, for now, an exception). The economist Yanis Varoufakis calls this order technofeudalism: big tech companies are no longer sellers in a market but the party that builds the market itself and collects rent from everyone. Evgeny Morozov objects: the companies still chase profit and still invest in research, so this is still capitalism. A reading between the two is possible: the line between profit and rent runs right through the middle of the AI industry. Where models can be copied, there is competition and profit; where chips, electricity and data centers cannot be copied, there is rent. The state used to be the party that set the rules of the market; today it pays these companies rent for its own infrastructure. The rule-maker is becoming a tenant.
To judge this contradiction, a sentence from the cybernetician Stafford Beer is useful: the purpose of a system is what it does. Under this principle, known by the acronym POSIWID, you should judge a system not by the purpose claimed for it but by the results it actually produces. As Beer put it, there is no point in saying that the purpose of a system is something it keeps failing to do. Sentences often heard in the AI debate can be looked at through this principle: "It wasn't supposed to be like this." "The state shouldn't let us go hungry." "AI companies will cure cancer." These are purposes claimed for systems. Under POSIWID, the answer to what the state is for, and to what a company is for, is found not on paper but in results. In earlier sections of this essay we looked at some of those results: entry-level jobs are shrinking, gains are piling up in a few companies, part of the electricity cost of data centers lands on household bills, and the same technology is used for surveillance and war. The optimists' hope is not baseless; in 2024, the developers of AlphaFold, which predicts protein structures, shared the Nobel Prize in Chemistry. But the optimists, too, can look at today's results and stand somewhere more realistic: to change a system, you first have to look not at what it promises to do but at what it does. Seen through this principle, one question becomes unavoidable: can this state collect the debt from section 6?
12Collecting the debt
In section 6 I set out the main claim of this essay: AI companies inherited humanity's accumulated universal labor and the infrastructure the public paid for; whoever takes the inheritance takes the debt too. The short answer to the question at the end of the previous section is this: there is no reason to expect the state to collect this debt on its own. Still, let us first look at the tools on the table. The oldest of them is tax. I chose the options by a single test: each tax should be paid to one of the three creditors from section 6. Some of them already exist somewhere; others are my own proposals, and I mark those.
Data centers
The creditor here is the public, which pays for the data centers' tax exemptions and for the bill of a growing grid.
- Whoever brings the load pays: In 2025, the electricity regulator in the US state of Ohio required large new data centers to pay for at least 85 percent of the capacity they reserve, even if they do not use it. Technically this is a tariff, but it does the job of a tax: the data center's cost does not land on household bills.
- A compute tax instead of exemptions: The exemptions go, and the compute used to train large models is taxed instead. The threshold already exists: the European Union's AI Act already regulates models above a certain level of compute separately. Small start-ups and universities stay exempt. This is my proposal.
- A charge for water and land: In many countries, mining companies pay a separate tax on the resources they extract; this is called a severance tax. A data center also uses up the water and land of the place it is built in. But paying this charge only to the people of the region creates a lottery: whoever has a data center built in their district wins. Yet the models were trained on everyone's labor. So a small part of the charge covers the local harm, and the larger part goes to a common fund in which everyone has an equal share. This is my proposal.
AI companies
The creditors here are the people who wrote the texts that became the models' raw material, and the workers the models replace.
- Shares instead of cash: Under the Altman proposal I mentioned in section 5, large companies would pay 2.5 percent of their market value every year. The same idea can be adapted to AI: the companies whose models were trained on the public's common stock of knowledge pay part of their tax in shares, and those shares build up in a public fund. A cash tax is like a dividend; a share that carries voting rights also brings a say in decisions, so it is a step toward ownership. This adaptation is my proposal.
- A windfall law: In 2020, the researcher Cullen O'Keefe and his colleagues proposed a voluntary pledge for AI companies: if their profits exceed 1 percent of the world economy, they donate part of the excess to society. But a voluntary promise is taken back when a company is under pressure. The same idea can become a tax whose rate rises as profit grows. This is my proposal.
- A training-data fund: In many European countries, a small charge is taken on blank CDs and phones and paid out to writers and musicians. This is called a private copying levy, and its logic is this: because it is impossible to track who copies what, the charge is collected as a whole. It is just as impossible to measure how much any given text contributed to a model. By the same logic, a collective charge is attached to model training and paid out to writers through a fund. The 1.5-billion-dollar Anthropic settlement I mentioned in section 6 was a one-off example; a fund makes such payment regular.
- A layoff premium: In the US, a company that lays off many workers pays a higher unemployment insurance premium; this is called experience rating. Its counterpart in the age of AI is this: a company that cuts its workforce while its revenue grows pays a higher premium. Automation is not banned; only automation whose gains stay with the company while its costs fall on society becomes more expensive. And the money goes straight to the people who lost their jobs. This is my proposal.
- Closing the hole in the global minimum tax: In 2021, more than 130 countries agreed that the profits of large companies should be taxed at a rate of at least 15 percent. In 2025, under pressure from the US, the G7 countries agreed not to apply the main parts of this rule to US companies. Since most of the largest AI companies are American, the hole benefits them directly. Closing it is the condition for taxing profit where it is earned.
The rich
The gains from AI do not stay inside the company; through shares, they flow to its owners. As we saw in section 8, the share of total income that goes to labor has been falling for years, and the falling share goes to capital. That is why we also have to look at wealth. The revenue from all three of the taxes below goes to a common fund:
- A minimum wealth tax on billionaires: In a 2024 report for the G20, the economist Gabriel Zucman proposed that the world's roughly 3,000 billionaires pay tax every year equal to at least 2 percent of their wealth. The G20 left the proposal out of its joint text; in France, a similar proposal passed the National Assembly in 2025 but was rejected by the Senate. The word "minimum" matters: someone who already pays that much tax pays nothing new.
- An inheritance tax: This is the tax closest to the title of this essay. Thomas Paine's fund, which I mentioned in section 6, rested on it: the tax on large inheritances is paid out to young people as starting capital. The young people in section 9, who cannot buy a home or find a first job, would be the first creditors of such a fund today.
- Closing the "buy, borrow, die" loophole: Very rich people often do not sell their shares, because a sale is taxed. Instead, they borrow against their shares and live on the loans; in the US, when they die, the gains built up pass to their heirs without ever being taxed. This method is called "buy, borrow, die." The proposal of some tax law scholars is simple: a loan taken against shares should be taxed as if the shares had been sold.
All of these tools share one weakness: they all assume that the company and the rich will agree to pay.
13Sanctions
A tax that does not say what happens when it is not paid is only a wish. Three sanctions can stand behind the tools of the previous section:
- Collection at the source: The data center's charge is not left to the company's own declaration; it is collected through the electricity bill, and a facility that does not pay has its grid connection limited. A data center's weakest point is its dependence on electricity, and that is where the debt is easiest to collect. This is my proposal.
- Exclusion from the market: The European Union's AI Act sets fines of up to 7 percent of a company's worldwide turnover for prohibited practices. The same scale could apply to tax. A company that also fails to pay the fine cannot sell in the country and cannot bid for public contracts.
- Losing the inheritance: Let us turn the rule of renouncing an inheritance from section 6 around: whoever does not pay the debt loses the inheritance. The weights of a model whose debt is not paid become public property. But weights alone are not enough, because the compute to run them is still rented. So part of the company's compute capacity is also set aside for public use. This is my proposal.
Can a harsher penalty be imagined? Let us run a thought experiment. The US Federal Trade Commission (FTC) has already ordered the deletion of some models trained on data collected without permission; this is called algorithmic disgorgement. Applying the same penalty to a large language model would mean giving the company a kind of memory-wipe sentence. But nothing on the internet disappears easily: Meta's LLaMA model leaked online in 2023, a week after it was shared only with researchers, and a model can be copied by training another model on its answers. What is more, because the model was made from a common inheritance, deleting it means burning the inheritance while punishing the heir. In the second Matrix film, the Architect tells Neo that the Matrix has been reset five times and rebuilt each time. Resetting a company does not reset the system either; the same model is reborn in another company.
Appointing a trustee to run the company is another path. But in Turkey, over the past decade, trustees were appointed to companies and municipalities that the government regarded as opposition: a trustee appointed by the state ties the company not to the people but to the power that appointed it. Jailing the executives is another path. The founder of the crypto exchange FTX was sentenced to 25 years in prison in 2024, and the crypto market carried on. In the preface to Capital, Marx writes that he deals with individuals only as personifications of economic relations; the next morning, someone else takes the jailed executive's place. All three experiments arrive at the same place: the penalty that works targets neither the person nor the knowledge, but the relation of ownership. A company that does not pay its debt goes bankrupt like any other company that does not pay its debts, and in bankruptcy the company can pass to creditors who turn their claims into shares. If the heir does not pay the debt, it is not the state but the creditors who take over the company: the writers, the public and the people who lost their jobs. This is my proposal.
But in the end, a sanction is also a question of power, because the state behind the companies has sanctions too. In 2025, Canada was preparing to collect a 3 percent digital services tax from large technology companies; when Trump cut off trade talks, Canada withdrew the tax hours before it was due to take effect. A country's power to collect a tax is limited by its power to stand up to the company's own state. That is why the most solid taxes are the ones tied to infrastructure that cannot move; wealth and profit can be taxed only if countries act together. In short, tools work only if there is a power to enforce them. Before we come to that power, let us look at where the world could go, with these tools or without them.
14What could happen
In this section I am not making a prophecy but running a thought experiment. I seat the actors from earlier sections (the companies, states, the young, retirees, unions and local movements) at one table and shake them with a few shocks. The starting point is November 2026. According to election-night projections, the Democrats took both chambers of the US Congress, but the White House stays with Trump until 2029. AI investment carries about half of US growth, and the first big lab IPOs are on the way. Iran has targeted US data centers in the Gulf. France and Spain are shaken by general strikes against budget cuts. Turkey, after September's investment fund crisis, is struggling with youth unemployment and housing.
Base scenario: slow digestion
If no big shock comes, the Democratic Congress drafts AI tax bills and the White House vetoes them. States set their own taxes and Ohio-style tariffs; these are the easiest tools from section 12 to apply, because they rest on infrastructure that cannot move. Productivity shows up in the statistics but not in wages; Engels' pause repeats itself. In 2031 there is no collapse, only a more unequal distribution: neither a dividend arrives nor ownership. This scenario turns out wrong if productivity reaches wages quickly or if new occupations make up for the loss of entry-level jobs.
First shock: the bubble bursts
An IPO disappoints, AI stocks fall 30 to 40 percent and data centers built with debt stall. Because investment carries half of growth, the US goes into recession, and workers are hit by the recession before they are hit by AI. The state rescues the big labs and takes shares in return: public ownership comes in not through the front door but through the window that bankruptcy opens. Public anger gathers in one sentence: "They nationalized the losses."
Second shock: energy and the Gulf
A conflict with Iran makes oil and electricity more expensive. The data center becomes a military target, and investment moves out of the Gulf to places where energy is cheap and safe. Compute rent merges with energy rent. Local movements have their strongest moment, because the electricity bill reduces the whole debate to one number. Turkey, an energy importer, loses twice: its bill grows, and its chance of becoming a compute hub disappears.
Third shock: Taiwan
A blockade halts the production of the most advanced chips. Compute becomes scarce, states reserve it for defense and surveillance widens. There is a paradox here: war slows the shift of work to AI. The price is inflation and an opposition silenced in the name of "national security."
Fourth shock: political change in Washington
If the Democrats also win the presidency in 2028, an AI tax and a public fund could become law. The companies threaten to move investment to other countries, and the rate is cut with the argument that "China will pull ahead." The result is a dividend fund: citizens get a small share every year, but no say in decisions. For the first time, the inheritance becomes an institution, but not as ownership.
Fifth shock: an abundance of open models
China and open-source communities release models close to the best ones for free, and chips get cheaper. The big labs' rent breaks, and small businesses grow stronger. But job replacement speeds up, and surveillance gets cheaper too. This is the scenario that spreads power the most, and the most favorable one for countries like Turkey that import compute.
Sixth shock: the scapegoat
As the chaos grows, a politician finds an easy enemy: "Your problem is AI." The state and the company are legal persons that certain groups stand behind but that no one can touch alone; AI, too, can be cut off from its owners and presented as an actor in its own right. A populist government shuts data centers, has the police cut cables, even has a facility bombed. But the question "where is AI?" has no answer: the model keeps running on a server in another country, and the open weights are already on everyone's computers. The Luddites' machine-breaking was a tragedy; the police cutting cables would be, in Marx's phrase, history repeating itself as farce. The real danger is this: the anger turns not on the owners but on a thing, and the owners stay invisible.
Seventh shock: radical backlash
Imagine an assassination attempt on an AI executive or sabotage at a data center. In the short term the security state grows stronger, data centers are declared critical infrastructure and protests begin to be mentioned in the same breath as terrorism. History offers examples in both directions. As with the radical flank effect I described in section 10, harsh actions sometimes strengthened the moderate wing and sometimes became the excuse to crush the whole movement; historians still debate how much the suffragettes' attacks on property in Britain contributed to women winning the vote. The outcome depends on whom the action targets and whether the peaceful wing distances itself from it.
Eighth shock: a capability leap
If AI agents become able to do most desk jobs reliably from start to finish, white-collar unemployment could pass 10 percent. Then the question from section 5 becomes real: the worker who was laid off was also a customer. Consumption falls, and company revenues fall. If this shock arrives together with the bubble bursting and an energy crisis, a vicious spiral begins: unemployment, high prices, a populist wave and a hostile language that turns domestic anger outward. This spiral is the path along which the risk of war rises most.
Ninth shock: Europe's strikes
If the strikes continue and governments present AI as a way to save money in public services, unions make AI a clause in collective agreements. Europe could become the only region where AI is used through bargaining with labor; the price is slower adoption and dependence on others for compute.
Turkey
For Turkey, the scenario is less a shock than a buildup. The fund crisis weakened household savings, youth unemployment is high, and AI is imported and paid for in foreign currency. As companies switch to cheap AI subscriptions, white-collar jobs for the young shrink, and anger flows into politics and the wish to emigrate. The state uses AI first for surveillance.
Black swans
Some possibilities make all the scenarios meaningless. The thinker Nassim Nicholas Taleb calls them black swans: rare events that are hard to foresee and change everything. A nuclear escalation would reset every parameter of this essay; compute, rent, dividends and strikes all assume a functioning society. A second reset would be the sabotage of critical infrastructure such as grids, payment systems and water plants with the help of AI. Malicious people could do it; in section 10 I mentioned an espionage campaign carried out largely by AI. Or a system capable of recursive self-improvement could break the rules of the game without anyone intending it.
A black swan can also come from a single person, because in the end there are people behind the giant companies too. In November 2023, OpenAI's board fired Sam Altman, and within days the company nearly fell apart. In June 2025, the feud between Musk and Trump knocked 14 percent off Tesla's shares in a single day. In November 2026, Nvidia's CEO Jensen Huang admitted a relationship with a businesswoman whom US intelligence suspects of working for China, and the company's shares lost about 10 percent in two days. An election defeat or the death of someone at the center of the order could have a similar effect. But in all these examples the system recovered quickly: Altman came back five days later, and Nvidia's shares rose on the third day. The principle from section 13 holds here too: the person changes, the relation remains. A personal shock becomes a black swan only when it hits a structure that is already fragile.
15Where society goes
Let us set the scenarios from the previous section side by side, actor by actor:
| Actor | Slow digestion | Bubble | Energy and Gulf | Taiwan | Political change | Open models | Radical backlash | Capability leap |
|---|---|---|---|---|---|---|---|---|
| AI companies | grow, concentrate | bailed out, state becomes partner | pull out of the Gulf | tied to defense | tax, dividend | rent breaks | victims, protected | win first, then lose customers |
| US state | gridlocked | takes shares | regulates tariffs | emergency powers | sets up fund | surveillance gets cheaper | security state | emergency income support |
| China | spreads via open models | fills the gap | relatively advantaged | war | rival argument | biggest winner | unchanged | brakes at home |
| Europe | regulates | takes damage | weakest link | energy and chip shock | copies the model | public models grow stronger | surveillance debate | welfare state tested |
| Gulf | rentier | buys cheap assets | plan shaken | risk of attack | draws investment | rent shrinks | unchanged | unchanged |
| Turkey | imported, insecure | capital flight | bills and emigration | chip and energy shock | no debate | most favorable | more surveillance | youth unemployment deepens |
| Young people | first rung disappears | recession | price rises | war economy | small dividend | new small businesses | movement splits | mass unemployment |
| Retirees | slow erosion | lose twice | bills | inflation | share of fund | less stock risk | unchanged | contributions fall |
| Local movements | strong on zoning | anger at bailouts | strongest moment | suppressed | gains | dispersed | split or grow | become mass movements |
Reading the table row by row, three regularities appear. AI companies survive in almost every scenario; all that changes is whom they are tied to: the market, the state or the military. The young win in no scenario; at best there is a small dividend and some new small jobs. And the tools from sections 12 and 13 enter the scenarios only as a watered-down dividend. For Turkey, the best scenario is the spread of open models together with a quick policy that turns them into leverage; the worst is the energy shock.
Taken together, the scenarios point in five possible directions:
- A rentier order: Income flows to the owners of compute and energy; the majority lives on dividends and support. Today's prototype of this order is the Gulf states. Political science calls them rentier states, because the state's income comes not from its citizens' labor but from oil rent; citizens get a share of that rent, but their political voice is limited.
- Populist turmoil: Anger turns to the ballot box and the street, and changes governments without changing the system. The findings in section 8 show that this direction has already begun.
- Pressure toward public ownership: Through bailouts or funds, the state and the public become co-owners of AI companies. There are examples: in 2008 the US government became a major owner of General Motors and the insurance giant AIG, and in 2025 it took about 10 percent of Intel. Norway's oil fund owns about 1.5 percent of the shares of the world's listed companies.
- A surveillance state: A security crisis makes surveillance permanent. Many of the surveillance clauses of the Patriot Act, passed as temporary after the September 11 attacks, were extended for years.
- Symmetric dispersion: Open models spread power. But a similar hope existed in the early years of the internet, and a network on which anyone could publish gathered, within a few decades, around a few platforms. Weights can be spread; the chips and data centers to run them are still in a few hands.
Is war inevitable?
For now, the evidence points more to populism than to war: automation and insecurity change how people vote, but jobless masses do not directly cause wars. The risk of war comes from inputs: chips, energy and rare minerals. A century ago, Lenin wrote that the great powers clashed over markets in which to export capital; today's rivalry runs more through these inputs. Nuclear deterrence and ageing populations put a brake on this risk. But as the vicious-spiral scenario shows, as domestic anger grows, so does the appeal of finding an enemy abroad.
Synthesis
Beneath all the scenarios lies the same movement: we are passing from a world measured in time to a world measured in energy, but with the ideas of the old world. Careers, salaries, pensions and the dream of buying a home are still built for a world in which people live by selling time. For these ideas to catch up with the new situation could take 20 years by Gates's estimate, or 50 if we look at Engels' pause. We do not know which scenario will come true; but in all of them, the same people pay the price of the lag first. So can the people who pay that price collect the debt themselves?
16Who will collect
The table in the previous section showed one thing clearly: the tools are on the table, but there is no power to use them. In this section I ask where that power could come from. First, a warning: collecting tax does not mean the money will go to the people. By the principle from section 11, the purpose of a system is what it does; the question is not what a tax was set for, but where the money goes. In 2025, the US Department of Defense signed contracts worth up to 200 million dollars each with four large AI companies. If the tax collected flows back to the same companies through defense and surveillance contracts, what the system does is not pay the debt but pass the money through the state's hands and carry it back to the companies. Unless the money is tied to funds in which the creditors have a say, tax stays on the dividend side.
Who you collect tax from is also a choice, and behind that choice stands a more basic question: will AI grow total output and demand, or will it grow the share it takes from the existing economy? In the language of section 4: will AI companies' income be profit from new production, or rent that takes a share of value produced elsewhere? The answer is not yet clear, but there are signs. Let us return to the number from the opening of this essay: in its filing for an initial public offering in October 2026, Anthropic measured its opportunity by the value of the world's knowledge work and, starting from the wages of knowledge workers, put that value at more than 40 trillion dollars; the whole world economy is about 118 trillion dollars. So what is shown as the market is, to a large extent, the wages paid to people today. And the people who earn those wages are also the economy's largest buyers; a technology that leaves its buyers jobless can shrink demand. This is the real question mark over the universal high income Musk promised.
If AI only grows its share of this pool, the position of ordinary people will get worse in a way history has rarely seen. For a long time, people believed that slaves built the Egyptian pyramids; but workers' tombs at Giza showed that the pyramid builders were not slaves and were fed with bread, beer and meat. The pharaoh needed them. The specter of uselessness from section 8 is exactly the disappearance of that need: a person whose labor is not needed does not even have the pyramid builder's bargaining power. And the problem is not that too little is produced. The world produces enough food to feed everyone, yet about 673 million people went hungry in 2024, while the richest 0.001 percent hold about three times the wealth of the poorest half of humanity. The largest owners of AI companies are also among the owners of this wealth. If the power of AI also gathers in these hands, millions of people will perish, living standards will visibly fall, and people will be left with no choice but to revolt. But the direction of that revolt is not set; as we saw in sections 8 and 9, today the populist right gathers most of the anger. That is why the scapegoat scenario in section 14 is dangerous: the anger turns not on a relation but on a thing.
In such a crisis, where will these rich people go? They live in this world too. In his book Survival of the Richest, the media theorist Douglas Rushkoff tells of the question five very wealthy investors asked him after inviting him to give a talk: after a collapse, how would they keep the guards in their bunkers obedient? These people believe they will not pay the price of the crisis; they do not feel that the present course carries a deterrent cost for them. What can make them feel that cost is, as throughout history, the organized people. In their book Taxing the Rich, the political scientists Kenneth Scheve and David Stasavage studied two centuries of tax records from twenty countries and found that what raised taxes on the rich the most was not the spread of democracy or left-wing parties, but mass mobilization: in the countries that mobilized for the First World War, the top income tax rate rose from 4 percent to 63 percent. Once millions had gone to the front, the argument that the rich should pay their share too could not be resisted. The lesson is not war; it is the demand for equal sacrifice and the mobilization of millions. Today's question is how to build an equivalent without war.
Some of the methods are old: strikes, boycotts and collective action can still work today. But if the other side has gathered AI into its own hands, the people, too, have to use it wisely: bringing together the demands of workers and tenants who do not know one another, extending the visibility dashboards from section 10 so they also show how much tax a company avoids, making the logistics of an action simple enough for a handful of people to run. Because the same tools are also used to watch the people who organize, the tool of resistance should not be a model on a company's servers, but an open model that anyone can run on their own computer. Perhaps new practices will grow from here. Local people allow a data center to be built only if the company gives the public a share; that way, they collect the debt themselves, without waiting for the state. The people whose writing became the models' raw material form a creditors' union, the reverse of the US debtors' union, the Debt Collective. Cities supply compute as a public service, like water and electricity. None of these exists yet. But Marx did not design the Paris Commune either; he studied it after it had formed and called it "the political form at last discovered" for the emancipation of labor. The workers' councils of Russia in 1905, the soviets, also grew not from a theorist's plan but from strike committees. The practices of the age of AI will grow not from a tax bill but from the experience of the people who try to collect this debt. Tax can be one of the tools of those practices; but on its own, it may not be the solution.
17An alternative utopia
If practices grow out of experience, a horizon is still needed: if this struggle is won, what kind of society would be built? This has always been a tense question for Marxism. In Socialism: Utopian and Scientific, published in 1880, Engels criticized the thinkers who drew up the society of the future at their desks; for him, the new society would grow not from a plan but from the contradictions of the old one. About seventy years later, the philosopher Ernst Bloch proposed a middle way through this tension: the concrete utopia. For Bloch, a concrete utopia was not an imagined perfect society, but a horizon that grows out of possibilities already present in today. What might the concrete utopia of the age of AI look like? The strongest candidate is an old dream: planning the economy with a machine.
This dream has a history, known in economics as the socialist calculation debate. In 1920, the economist Ludwig von Mises argued that without market prices an economy cannot calculate what to produce and in what quantity. In 1945, Friedrich Hayek deepened the objection: the knowledge needed is scattered across the daily lives of millions of people, and the price gathers it without anyone needing to know all of it. The planning economist Oskar Lange, for his part, wrote that the equations the market solves by trial and error could be solved on a computer in less than a second. There were also attempts. In the Soviet Union, the computer scientist Viktor Glushkov's national network OGAS, which was to connect the whole economy, ran into ministries that did not want to lose their powers. In Chile, Cybersyn, which the Allende government built with Stafford Beer, whom I mentioned in section 11, ended with the military coup of 1973. In both cases, politics did more than technical limits to decide the outcome.
Two things have changed since then. First, compute is no longer the problem. Second, the largest examples of planning are inside capitalism: as Leigh Phillips and Michal Rozworski point out, Walmart and Amazon are run internally not by markets but by calculations made at the center. Language models also touch Hayek's objection. In a 2024 experiment by Google DeepMind researchers, a "Habermas Machine" drafted a statement of common ground from what people with different views had written, and participants preferred it to the human mediator's statement in 56 percent of comparisons. Price is no longer the only way to gather scattered knowledge.
But the machine cannot abolish politics. Engels wrote that under socialism the government of persons would give way to the administration of things; politics would one day become a technical job. Yet the moment you tell a machine to "maximize social welfare," you have asked the hardest question: what is welfare? A longer life, less work, cleaner air? These conflict with one another, and the choice between them is made not by a calculation but by a decision. In 1951, the economist Kenneth Arrow showed mathematically that no voting method exists that turns individuals' preferences into a single social ranking while meeting a few basic conditions of fairness all at once. When politics goes inside the machine, it does not disappear; it becomes invisible. The weight of each value is written into the model's objective function, the formula that sets what it tries to maximize, and who writes that formula looks like an engineering decision. Weber's iron cage, which I mentioned in section 7, is this time built in code that most people cannot read. And Goodhart's law, named after the economist Charles Goodhart, applies too: when a measure becomes a target, it stops being a good measure. Soviet factories hit the plan's numbers but did not meet people's needs.
That is why the concrete utopia is not a machine but a question of ownership. The same compute becomes, in a company's hands, the plan inside Walmart, and in a party's hands, something close to China's comprehensive surveillance. If society itself can write the objective function, a democratic planning could emerge. Small pieces of it exist today. In Taiwan, a tool called Polis, used in the vTaiwan process, gathers the views of thousands of people and shows where they agree; in 2023, Anthropic also used the same tool to gather from about a thousand people the rules a model would follow. The utopia of this essay is not a plan: it is a society in which the machine, instead of taking the work of deciding away from people, increases the number of people who take part in that decision more than ever before. In Bloch's terms, a possibility already present in today.
18Closing
For Marx, when the way people produce changes, their ideas change too, but they lag behind. Our production is changing and AI is speeding up that change; our ideas, our laws and our life plans, however, still live in the old world.
Inside this lag there is an unsettled account. AI companies inherited what humanity has written, our common language and the infrastructure the public paid for. Whoever takes the inheritance takes the debt too. As I listed in section 6, this debt has three creditors. The first are the people who wrote, for nothing, the texts that trained the models: millions of people alive today and the earlier generations who are no longer alive. None of these people wrote so that an AI could be trained. They wrote their letters, books, forum answers and code for other people to read, and they had no idea that what they wrote would one day become a model's raw material. A debt cannot be paid to the dead; their share can only be paid to the living, through a common fund or common ownership. The second is the public, which pays for data centers' tax exemptions and for rising electricity bills. The third are the people these models replace, who lose their jobs or never get one. So the creditors are us: the ordinary people who make up the overwhelming majority of the world. This debt is not settled with a small dividend; its real payment is a say over the infrastructure, that is, ownership. How to win it is the central political question of the coming decade.
Nobody knows what answer this question will get, and the scenarios in this essay claim no certainty. Besides, history does not move in a straight line. In 1807, in the preface to The Phenomenology of Spirit, Hegel describes his own age as a time of birth: a new era ripens for a long time without being noticed, the old world comes apart piece by piece, and then the shape of a new world appears all at once. Hegel compares this to the birth of a child: months of quiet growth turn, with the first breath, into a leap. History sometimes moves like this too. Change that builds up slowly becomes visible one day as a rupture, a break, and one morning we may find ourselves in a completely different situation. Each of the shocks in section 14 is the possibility of such a morning, and among them there may be possibilities none of us has thought of, ones that look absurd today.
That is why we also have to accept this: democracy, equality, law, justice, "Western civilization" and all our concepts of living together are not things found ready-made in nature. They are all concepts we agree on, and they live only as long as we keep agreeing on them. These concepts stand in front of a curtain far thinner than we think. What holds that curtain up is that people can go to work every day, pay their bills and believe that tomorrow will look like today. When that belief is shaken, the curtain can catch fire, and history has shown many times how fast this can happen: four years after the crash of 1929, Germany had a dictatorship where its democracy had been. So at the end of this essay I am left with a question whose answer I do not know: will the masses left without homes, without work and without a future go on agreeing on today's concepts, or will they build new concepts out of their own conditions and agree on those?
The uncertainty and the sense of being trapped that so many of us feel today are therefore not empty worry; they are birth pangs. According to the concept of the Zeitgeist, associated with Hegel, every age has a spirit of its own, and these pangs, too, may give birth to the spirit of the time to come. But that spirit will not arrive by itself, through mystical waiting. This is exactly where Marx parts ways with Hegel: he sees history not as a spirit unfolding by itself but as the action of people. As he wrote in The Eighteenth Brumaire of Louis Bonaparte, people make their own history, but not under conditions they have chosen; they make it under conditions handed down from the past. We did not choose the conditions of the age of AI. Where those conditions lead, though, we can decide not by waiting but by trying to pull them in a direction. The course of history is changed by those who try to steer it.