Digital Socialism in the Age of AI

Proposals for digital socialism usually begin with ownership: put the models, the data and the compute in public hands. They rarely begin with the state that would hold them. Owning a technology and governing it are different skills, and the second is the scarcer one. A public body that runs a frontier model would have to set its limits, audit its outputs, answer for its failures and resist the pull to use it for its own ends, all at a speed no ministry was built for. The case for public ownership of artificial intelligence may be right. But it rests on an assumption it seldom examines: that the institution receiving the power is able to carry it.

I come to this from political science and public administration, fields that ask less who should hold power than whether the holder can use it without breaking things. That puts me on neither side of the usual argument.

What is digital socialism?

Digital socialism, as I use the term here, is the idea that the core inputs of artificial intelligence (the models, the data and the compute) should be owned publicly or collectively rather than by a few private firms. It is less a single programme than a family of diagnoses that share a cure.

The diagnoses are familiar. Nick Srnicek's Platform Capitalism (2016) warned, in its publisher's words, that "the fundamental foundations of the economy are rapidly being carved up among a small number of monopolistic platforms". Shoshana Zuboff's The Age of Surveillance Capitalism (2019), on its publisher's account, describes a power that "has deepened inequality, sown societal chaos, and undermined democracy." Yanis Varoufakis's Technofeudalism (2023) "shows how the owners of big tech became the world's feudal overlords", its publisher says. Ben Tarnoff's Internet for the People (2022) is the plainest: the internet is broken, in Verso's summary, "because it is owned by private firms and run for profit", and the remedy is to "deprivatize the internet."

As I read them, each locates the problem in who owns, and the cure follows: change the owner. Tarnoff goes furthest, calling for "publicly and cooperatively owned alternatives that encode real democratic control." I want to stay with that verb, encode. It treats democratic control as something written into an institution at its founding, like a clause. I think it is more like a muscle: built slowly, kept up by use, lost through neglect.

Isn't this just the old socialist calculation debate again?

Partly, and that is where the trouble starts. The twentieth-century argument over socialism was mostly about knowledge, not administration, and the data era has revived it on the same terms.

In 1920 Ludwig von Mises argued that a socialist economy, without market prices, could not solve the problem of economic calculation. Oskar Lange answered that a central planning board could set prices by trial and error. Friedrich Hayek went deeper in "The Use of Knowledge in Society" (American Economic Review, 1945). The knowledge a planner would need, he wrote, "never exists in concentrated or integrated form"; much of it is "the knowledge of the particular circumstances of time and place", held by the people closest to the work.

For decades that looked decisive. Then the platforms seemed to gather exactly the dispersed knowledge Hayek said no centre could hold. Evgeny Morozov's essay "Digital Socialism?" in New Left Review (116/117, 2019) took the opening: "I will go on to suggest ways in which the development of digital ‘feedback infrastructure’ offers opportunities for the left to propose better processes of discovery." Leigh Phillips and Michal Rozworski's The People's Republic of Walmart (2019) pointed to Walmart and Amazon as some of the largest planned economies in the world. If a retailer can plan, why not a society?

Notice what both sides argue about. Can the information be gathered? Can it be processed? These are questions about calculation, and even Hayek's objection, for all its depth, is about where knowledge lives. What kind of organisation would sit at the centre, with what staff, what rules, what interests of its own and what record of failure, is left as a detail. In public administration, that detail is usually where a project lives or dies.

Why is governing AI harder than owning it?

Because ownership is a legal status and governing is a continuous practice. A parliament can transfer title to a model, a dataset or a data centre in a single vote. Nothing in that vote creates the people, procedures and habits needed to run the thing well the next morning.

Francis Fukuyama gave the distinction a working definition in 2013: "As a starting point, I am going to define governance as a government’s ability to make and enforce rules, and to deliver services, regardless of whether that government is democratic or not," as Foreign Policy quoted him. The last clause matters. Capacity is not legitimacy. A state can be elected and incapable, or capable and unaccountable, and a digital-socialist proposal needs one that is neither. Ownership debates rarely measure either.

Jennifer Pahlka, who founded Code for America in 2010 and served as US Deputy Chief Technology Officer under President Obama, makes the point from inside government. Recoding America (2023) rests on the observation, as the book's site puts it, that "policy is disconnected from the details of implementation", and its own summary could hang over this whole debate: "While factions fight about policy, it's delivery we need now."

Apply that to a frontier model. Ownership gives you the right to set its limits, title to its outputs, liability for its failures, the power to use it and a budget line. Governing needs a procedure for setting and revising those limits, auditors the operator does not control, someone named who answers to the people harmed, a body outside government that stops the state using it for its own ends, and engineers who are hired, paid and kept. The first list can be acquired by statute. The second has to be built, slowly.

Can governments run AI today?

On the evidence of the most careful public audit I have read, one large and established democracy was not yet ready to use AI at scale, let alone run a frontier model. In March 2024 the UK's National Audit Office reported on AI across the British government, drawing on a survey of government bodies as at autumn 2023.

The line I keep returning to is not a number. The draft strategy for AI adoption in the public sector, the auditors wrote, "does not set out which of these departments has overall ownership and accountability for its delivery." Note the word. In administration, ownership means being the one who answers, and that is exactly what the strategy lacked. The government's own digital office, the report adds, noted that "there is currently limited capacity within the system to fully exploit and scale the opportunities presented by AI."

Three cautions. This is one country. The survey describes autumn 2023 and I have not read a follow-up, so the gaps may have narrowed. And these bodies were trying to use AI, not build or steer a frontier model. That last caution cuts my way: if using AI stretched an established civil service, steering one would stretch it further.

What does public AI look like in practice?

The public AI efforts I have read settle the ownership question first and leave the governing question for later, sometimes in so many words.

The first is Switzerland's Apertus, announced on 2 September 2025 by researchers from EPFL, ETH Zurich and CSCS. It is a real achievement, and I want to say so plainly. Its makers describe "Switzerland's first large-scale, open, multilingual language model", released with "its architecture, model weights, training data and methods". Joshua Tan, lead maintainer of the Public AI Inference Utility, called it "our best proof yet that AI can be a form of public infrastructure like highways, water, or electricity."

Then look at how its behaviour on contested questions was set. The project's technical report describes a Swiss AI Charter: "We develop a set of precepts for LLM behaviour informed by Switzerland’s constitutional values, including neutrality, consensus-building, federalism, multilingualism, and respect for cultural diversity." To test public agreement, the team "recruited a sample of 163 Swiss residents through Prolific and through the ETH Decision Sciences Lab." And the report looks ahead: "The Swiss AI Initiative plans to open the Swiss AI Charter for further refinement through a democratized process."

This is careful, candid work, and it is the governing problem in miniature. Ownership was settled completely: weights, data and code are public. Who decides what a public model says on contested matters was settled by a research team, tested on 163 recruited residents, with the democratic process in the future tense. I do not fault the researchers; a university consortium cannot be a constituent assembly. My point is that no body with a democratic mandate was there to set the model's values, so its builders did.

Advocacy shows the same order. Mozilla's 2024 vision of Public AI, as its authors summarised it, calls for "a robust ecosystem of initiatives that promote public goods, public orientation and public use." That says what public AI is for, not who governs it, though the paper, which I could not open, may.

The second example is the European Union. On 11 February 2025 the Commission launched InvestAI, "an initiative to mobilise €200 billion for investment in AI", including "a new European fund of €20 billion for AI gigafactories". The announcement's own image was a CERN: "This unique public-private partnership, akin to a CERN for AI". On 30 July 2026 the call went out for "up to seven AI Gigafactories", closing on 12 November 2026.

Read the call, and the CERN comparison changes shape. The facilities will be "Led by industry". Public money from the Union and member states will act as "an anchor customer", and the call expects "more than €20 billion" in private investment. The public share is computing time, procured. That may be sensible, and I have not read the tender documents, which may set conditions the call page does not describe. But on that page, the public is a buyer, not a governor.

CERN is worth the comparison because it is governed. Its Council "is the highest authority of the Organization", it "approves programmes of activity, adopts the budgets and reviews expenditure", and "Each Member State has a single vote and most decisions require a simple majority". That is a governing body with a known membership and a known rule for deciding. A CERN for AI would need one, and the call page I read describes none.

What can Project Cybersyn and Seeing Like a State teach us about public AI?

That a planning technology lasts only as long as the institution holding it, and that a state with powerful tools tends to simplify the society it governs. Both lessons are about governing.

Chile's Project Cybersyn is the best-known ancestor of digital socialism, and Eden Medina's history of it is Cybernetic Revolutionaries (MIT Press, 2011). In the book's description, the system was to feature "holistic system design, decentralized management, human-computer interaction". It "was never completely implemented", and Allende's government "ended with a violent military coup". Cybersyn is often cited as proof that socialist planning by computer was possible. I read it as proof of something plainer: the technology did not outlast the government that held it.

James C. Scott's Seeing Like a State (Yale University Press, 1998) supplies the darker half. "Centrally managed social plans misfire," in his publisher's summary of the argument, "when they impose schematic visions that do violence to complex interdependencies". Among the conditions for disaster it lists "administrative ordering of nature and society by the state" and "a prostrate civil society that cannot effectively resist such plans."

A frontier model is the most powerful instrument of administrative ordering ever offered to a state: it can read, sort and summarise a population's records at a scale no census office imagined. That is why resisting the pull to use it for its own ends is on my opening list. Under private ownership, the worry is a company using AI on its customers. Under public ownership, the same model sits inside the institution that also taxes, polices and counts. Public ownership does not remove that risk. It moves it to where Scott found the worst failures.

Does Elinor Ostrom prove collective ownership can work?

Yes, and that is why she belongs on my side of this argument. Ostrom showed that people can hold resources in common without handing them to a state or a market, and that the arrangements that lasted shared a recognisable set of governing institutions.

Her design principles were first set out in Governing the Commons (1990), the Ostrom Workshop at Indiana University notes. Matthew Rimmer's summary in The Conversation lists them: "clearly defined boundaries; congruence between rules and local needs and conditions; collective-choice arrangements; monitoring; graduated sanctions; conflict-resolution mechanisms; minimal recognition of rights to organise; and nested enterprises." Her 2009 Nobel Prize in economics was, in the citation he quotes, "for her analysis of economic governance, especially the commons".

Of the eight, only the first, boundaries, comes near the question of who holds title, and even that is about membership. The rest are about governing: who makes the rules, who watches, what happens to those who break them, how disputes are settled, how small units fit inside larger ones. An AI commons would need its own answers. Who are its members: citizens, users, the people whose data trained it? Who monitors a model that changes every few months? What is a graduated sanction when a model harms someone? In what I read of the ownership-first texts for this essay, these questions barely appear.

Is Anthropic's public benefit fund a step toward digital socialism?

Not yet, and the reason is this essay's argument in miniature. On 18 October 2026 Anthropic announced that, "Subject to stockholder approval, 5% of Anthropic's shares will be committed today, with an additional 2% added at the board's discretion for every 50% increase in the company's value above a set baseline." Counting its co-founders' and employees' earlier pledges, it says, "roughly 20% of Anthropic shares will be dedicated to the public good." The new Anthropic Public Benefit Fund "might explore large donations to traditional nonprofits or direct distributions to specific populations", and the company names AI's effect on jobs as one example. I covered the terms in our news item on the fund.

This is a real transfer of value, and a large one: CNN put the opening 5% at $48 billion at the company's latest valuation. But look at the order of the announcement. The shares come first. The rules come later: "In the coming months, we will share more about the Anthropic Public Benefit Fund's structure and governance." The fund is set up as "a wholly-owned LLC", so for now it answers to the company whose shares it holds. The public receives a stake in the value of a frontier lab. It does not receive a say in how the lab is run, and not yet in how its own share will be spent.

Elon Musk's reply made the gap easier to see. "I recommend giving the shares directly to the public… This is what I will be doing," he wrote on X, and on Monday he added: "Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many." His proposal, CNN noted, "was thin on detail": it is not clear how the shares would reach tens of millions of children, or who would be eligible. Even taken at its word, a share in a child's name is ownership in its purest form and governance in its thinnest: one small holding in a company whose control stays exactly where it was.

So the first large moves to put AI wealth in public hands have come not from states but from the companies themselves, and both arrived in the order I have been describing: ownership announced, governing promised. To its credit, Anthropic says the governing part is coming, and that is the part to watch. If the fund publishes who decides, by what rule, under whose audit and answering to whom, it will have done something most digital-socialist proposals have not. If it does not, a fifth of a frontier lab will have been dedicated to the public without the public gaining any way to govern it.

Isn't "capacity first" just an excuse to leave AI to private companies?

It can be, and the strongest version of that objection has three parts.

First, capacity is built, and states have built it. Mariana Mazzucato's The Entrepreneurial State argues, as her book's page puts it, that "the private sector only finds the courage to invest after an entrepreneurial state has made the high-risk investments." DARPA was "Created in response to the launch of Sputnik in 1957", with a mission "to create and prevent technological surprise for our national security." CERN, with 25 member states, shows that states can govern a vast shared technical facility together. None of these inherited its capacity. Each built it after being created and given a job.

Second, the private firms that own frontier AI today were not chosen for their governing capacity either, and nobody elected them. If my test applies to a ministry, it applies to a lab. Holding the public to a standard the market never had to meet is not a neutral position.

Third, waiting is itself a decision. Every year spent building capacity before acting is a year in which the field is shaped by whoever is already in it.

I accept most of this, and my answer is narrower than the objection fears. I am not saying public bodies cannot govern AI. I am saying the capacity has to be part of the proposal, not assumed by it. DARPA was created with a mission, and CERN has a council that votes; their capacity grew inside a structure set up for it. Mazzucato's state takes risks as a funder, and funding research is a different job from governing a deployed system the public talks to every day. Moving ownership from one actor that cannot yet govern AI to another that cannot yet govern it does not solve the problem. It relocates it.

What test should a digital socialist proposal pass?

Before asking who should own AI, ask who can govern it, in a form that can be checked. This is the test I would put to any proposal: public, collective or private.

A proposal that passes most of this deserves a serious hearing, whoever would own the model. A proposal that skips it is asking for power before showing it can carry it.

So who should own AI?

I don't know yet, and I suspect the question is being asked too early. The record I read is lopsided. Swiss researchers built an open model meant as public infrastructure and left the democratic process for later. Europe promised a CERN for AI and wrote a call in which public money is an anchor customer. A British audit found that the government's draft strategy for AI in the public sector did not say which department owned it and would answer for it. Anthropic dedicated a fifth of itself to the public good and promised the rules for spending it later. In each case the hard part was not getting the technology. It was governing it.

Digital socialism may yet be the right answer for parts of AI: open public models, shared compute, data held in trust. If it is, it will be because someone built the institution first and gave it rules, staff, auditors and a public it answers to. The calculation debate asked whether a socialist economy could know enough. The question now is whether a public institution can govern enough. Ownership is the easy part. Governing is the part the argument still owes us.

---

This is an opinion piece. It argues a position and it is signed. The record comes from: Hayek's "The Use of Knowledge in Society" on Econlib; the Mises Institute's page on Mises's 1920 essay; Wikipedia's account of the calculation debate for Lange's reply; Morozov's "Digital Socialism?" in New Left Review; Novara Media on The People's Republic of Walmart; the publishers' pages for Platform Capitalism, The Age of Surveillance Capitalism, Internet for the People and Technofeudalism; Foreign Policy for Fukuyama's definition; the site of Recoding America and Pahlka's Stanford RegLab profile; the UK National Audit Office's Use of artificial intelligence in government and its summary; ETH Zurich's Apertus announcement and the Apertus technical report; Mozilla's announcement of its Public AI paper; Anthropic's announcement of its Public Benefit Fund and CNN on Elon Musk's response; the Commission's InvestAI announcement; the EuroHPC Joint Undertaking's AI Gigafactories call; CERN's governance page; MIT's page for Cybernetic Revolutionaries; Yale University Press's page for Seeing Like a State; the Ostrom Workshop and The Conversation on Ostrom; Mazzucato's book page; and DARPA. Where the essay leaves the record and enters argument, it says so. Our writing workflow runs on Anthropic models, and Anthropic is one of the companies this piece discusses. The interpretation is mine.