Who owns the machine?

A few things stood out. In all scenarios AI delivers economic growth, but the tool is frank about a downside for society: a bigger slice of the pie for capital and higher unemployment are hardly an advertisement for AI. The most extreme scenario takes it for granted that AI can do almost half of today's cognitive work within four years. Not a word, of course, about energy, water or ethics. I leave those aside here too, because my subject is distribution: who gets the profits, and what are the consequences of that.
What the scenarios show

However you torture the parameters, capital celebrates and labour loses out. In the extreme scenario, the share of national income going to workers falls (in the US) by a quarter in four years, from 60 to 45 percent. Most of that loss is knowledge work. How many jobs disappear depends on how quickly wages fall, and on how easily AI replaces ‘whole’ workers rather than single tasks. When wages fall fast, knowledge workers keep their jobs at a sharply lower salary. If they don't, nearly a quarter of knowledge workers end up unemployed. Not easy to choose. Ironically, wage inequality may shrink, because manual work gets more rewarded and knowledge work less. The electrician gains a third, the lawyer loses ground. The economy grows by a third while the total wage bill stays flat. All of the extra prosperity goes to those who own capital.

Not a bold prediction

Let us assume the direction of the scenarios is right, not least because this is a forecast of a trend that already exists in many countries. The share of value added going to labour has been falling in many countries for decades. In Germany, France, Italy and Spain, it fell by roughly a tenth between 1980 and 2000, and then stabilised. In the United States, it drifted down slowly for decades before falling much more sharply after 2000. The causes for this decline differed between the US and Europe. In Europe, the decline came mainly from labour market reforms and demographic change. In the United States, automation did most of the work. AI is the next chapter of that American story, also coming to Europe. The same research also concludes that technology alone does not decide who benefits from growth. Institutions and policy choices do.

This distribution question is a classic. In his Grundrisse, Marx described how the accumulated knowledge of a society eventually congeals in machines. A language model is almost literally that: our books, ideas and knowledge, compressed into capital. What the knowledge worker used to sell by the hour is now bought in one go. Workers have always defended themselves against this through trade unions and collective agreements. But if the work is no longer needed, and no new work takes its place, there is little left to bargain over. The only escape is to move the fight from the payslip to ownership.

Ownership is on the table

At least one AI company draws the same conclusion. In a policy paper, Anthropic proposes, among other things, to give every American a capital account, partly funded with shares in AI companies. It builds on the Trump Accounts, which give every newborn 1,000 dollars.

Why would an AI company propose this? The motives are probably mixed. There is genuine concern about what its own technology does to work. There is also self-interest. A company that is visibly thinking about how to share the gains is better placed to shape regulation, and to weather a public backlash, than one that is not. Anthropic published the paper in June, together with a pledge of 350 million dollars for research and training and a call for federal AI legislation. Both motives can be true at the same time.

Bernie Sanders goes further. He wants half of the shares of Big AI, including OpenAI and Anthropic, placed in a public fund, with voting rights. And Donald Trump let it be known that his voters and Sanders's voters are not that far apart. Even in the United States, public ownership of AI is no longer unthinkable. 

Europe's problem

Americans could at least share in the profits of their own AI companies. In Europe we have a problem: we don’t have those companies. We will get the effects on the labour market all the same. So what is the answer?

Start with what we once expected the labour market to deliver: economic security for everyone. Europe has answered that question before. In 1942, William Beveridge laid the foundations of the British welfare state, with a flat-rate floor of security that did not depend on how much you earned. Across Europe, similar floors followed. In 1957, the Netherlands introduced the AOW, a state pension for everyone who has lived in the country, regardless of their work history. Those floors were paid for collectively, through contributions and taxes, largely on income from work. Exactly the base AI erodes.

We can create the same kind of economic security with wealth: a basic endowment. Capital of your own, for everyone, with or without a job, independent of what you earn or ever earned. The idea has European roots as well. The British economist Tony Atkinson proposed a capital endowment for every young adult. Thomas Piketty went further and argued for an inheritance for all. I see three ways to fill it.

Three ways to fill the pot

First: earn from the AI machine that already exists. Europe may lack AI champions, but it does not lack capital. Take the Netherlands. Dutch pension funds manage around 1,720 billion euros, and some 150 billion of that was invested in tech last year. Through their pensions, Dutch workers already own a slice of the machine. So do Australians through their superannuation and Danes through their labour-market pensions. Invest in AI: own it, literally.

The catch thereafter is who gets that slice. Those who do not work build up nothing. Those who do cannot touch it until retirement. A knowledge worker who loses her job to AI at 45 gets nothing from the very technology that replaced her. So give everyone a basic endowment, invested the same way. Equal for all, with or without a job, and available for retraining or for bridging the gap between jobs. Part of the money is already on the table: the tax breaks that now flow only to pension saving through employment.

Second: if Europe builds its own AI, build it with different ownership. The Draghi report called for massive public and private investment to close Europe's technology gap. Fine. Since 2024, the EU has been turning its publicly owned supercomputers into AI Factories for European start-ups and researchers. Public money, public hardware. 

Yet the question of who owns what comes out of them is hardly asked. That matters. Building on Elinor Ostrom's work on the commons, a recent taxonomy of commons-governed AI shows that pooling computing power alone is fragile. If a model trained on public machines ends up behind a private license, the effort has been shared but the power has not. Shared ownership has to run through the whole chain: the data, the computing power, the model. It can be done. The open language model BLOOM was built by around a thousand researchers on a French public supercomputer, and released for anyone to use. 

There is a democratic side as well. Every major AI company says it serves the common good. In practice, as the philosopher Mark Coeckelbergh observes, a handful of executives and their boards decide what that good means. 

So attach shared ownership to that capital. An AI factory built partly with taxpayers' money can just as well be partly owned by taxpayers, through a public or cooperative stake that pays its returns into the basic endowment and gives citizens a voice in what gets built. What we build together should belong to all of us.

Third: skim the profits of companies that use AI intensively. AI profits do not only arise in Silicon Valley. They also land at the insurer in Frankfurt or the bank in Amsterdam that automates its back office. And those profits rest on something we produced together: the books, ideas and knowledge compressed into a large language model.

Europe has skimmed unearned profits before. In 2022, when energy prices soared, the EU imposed a temporary solidarity contribution on the windfall profits of fossil fuel companies. Profits that companies had done little to earn were partly returned to society. The same logic applies here.

Exactly which profits are due to AI cannot easily be determined. It does not need to be. Tax high profits more heavily in sectors that are automating fast, and pay the proceeds into the basic endowment. Do it at European level, so that companies cannot play member states off against each other. That has been done before as well: since 2024, large multinationals across the EU pay a minimum tax of 15 percent.

Room to say no

Back to the knowledge worker who loses her job to AI at 45. The scenarios gave her two options: a sharply lower salary, or no job at all. Either way, the gains go to whoever owns the machine. Capital of her own gives her a third option. She can afford to wait for a job that fits, or take a year to retrain for a new one. And it changes her position at the table: someone who can walk away gets a better offer. A basic endowment does what trade unions once did for the payslip. It gives people back some bargaining power.

On Anthropic's web page, you can move sliders for how capable AI becomes, how fast wages fall and how many jobs disappear. There is no slider for who owns the machine. I wouldn't have put it in either, if I worked at Anthropic. That is the one parameter we have to set ourselves.