A company that sells book metadata now advertises to AI labs with a single line: the world’s best AI training data is sitting on a shelf. It brokers orders of between a thousand and a million volumes. It keeps the buyers anonymous. It offers non-disclosure agreements as a feature. The books arrive by the pallet, the spines come off in a hydraulic cutter, the loose pages go through an industrial scanner at something like a hundred a minute, and what remains goes to pulp. Booksellers across Europe describe orders assembled from lists of numbers, indifferent to subject, indifferent to price, arriving at three in the morning.

Every other fight about AI and culture has been about copies: who made them, who was paid, who was asked. Scraped sites can be put back up. Torrented libraries can be deleted and the books still exist. This fight is about originals, and it is the first one that cannot be reversed by any court, settlement or apology.

The same company tells its clients, in writing, that the optics problem is real, and that “AI company destroys two million books” is not a headline that generates sympathy.

The first thing to understand is that the destruction is not waste, or haste, or contempt for the object but an US legal argument, executed in hardware.

Buying a book gives you that copy and nothing more. You may lend it, resell it, burn it. What you may not do is reproduce it, because reproduction is the thing copyright actually restrains. Scanning a book makes a second copy, and a second copy is exactly what the law is built to notice. So a company that buys two million books and scans them has, on the face of it, made two million unauthorised reproductions.

Unless nothing was multiplied. In June 2025 a federal judge in California held that this kind of format conversion is fair use, on the reasoning that the digital file did not add a copy to the world, it replaced one. The paper went in, the file came out, and the count stayed the same. The original was destroyed; one thing took the place of the other. That reasoning only holds if the book is actually gone. Keep the volume on a shelf and keep the scan, and you have turned one purchase into two copies, which is reproduction under any reading. Destroy it, and you can call the whole transaction a change of format.

The hydraulic cutter, in other words, is not downstream of the scanner. It is upstream of the court. It exists to guarantee that the sentence the original no longer exists remains true, because that sentence is what makes the file lawful. Every pulped book is a piece of evidence being manufactured.

And the same ruling found that downloading pirated libraries was not fair use. That was the finding that created liability, and it was settled a year later for one and a half billion dollars, approved this month, at roughly three thousand dollars a work before it is divided according to publishing contracts nobody wrote with this in mind.

Put the two holdings side by side and read what the law actually told the industry. Do not take the book without paying. Buy the book and then destroy it, and you are protected.

I did not set out to build an AI company. I wanted to work in publishing. That was the plan I had at twenty, and I pursued it, and for a while I had it. I spent a period inside a couple of large French publishing house, in the part of the business where you can see the whole catalogue at once: the fonds éditoriaux, the accumulated backlist, decades of titles that had been edited, fact-checked, typeset, and then quietly retired into a warehouse and a rights database.

What I argued for, repeatedly and without success, was that the fonds should become something other than stock. Not reprints. New formats. Structured digital learning, modular knowledge products, the catalogue turned into something you could query and transform rather than something you could only order. The objections were never that this was a bad idea. The objections were that the rights were complicated, that the market was unproven, and that it could wait.

It waited. It is still waiting. Twenty-five years after the web, the industry’s digital product is a printed book in a worse container.

I left, eventually, and I now spend my working life assembling training corpora out of public domain and openly licensed text. It is an odd destination for someone who wanted to be an editor, and it is closer to that original argument than it looks: the question is still what you do with a catalogue nobody is reading.

I want to be precise about why this matters now, because the connection is not sentimental. The books being cut apart in those scanners are, overwhelmingly, the books that argument was about.

Paul Heald did an experiment about a decade ago that I think about constantly. He took a random sample of new books available for sale and plotted them by decade of original publication. There were more titles available from the 1880s than from the 1980s. Not older, not rarer, not more beloved. Simply more available.

The reason is not mysterious. The 1880s are in the public domain and anyone can reissue them. The 1980s are in copyright and commercially dead, which is a different condition entirely: legally protected, economically abandoned, held by someone who will not exploit them and will not release them. Something like ninety percent of titles become commercially unavailable within two years of publication. When Google’s book settlement was being litigated, roughly half the ten million books in scope were estimated to be out of print, and of those, about a million were orphans whose rightsholder could not be located at all.

This is the twentieth century. It is the most documented, most edited, most professionally produced body of text in human history, and for forty years its custodians have treated it as inventory.

So when a laboratory decides it needs clean pre-2022 prose, in volume, and goes looking for it, here is what it finds. It does not find a licensable digital corpus of the twentieth-century backlist, because no such thing was ever built. It finds physical copies, on used-book platforms, priced at two or three euros, because that is what the market says a book nobody has digitised in thirty years is worth.

The laboratories did not outbid the publishers for this material. There was no bid. There was a warehouse, and a price per kilo.

James Boyle called this a second enclosure movement, the enclosure of the intangible commons of the mind, and noted that the public domain has to be invented before it can be saved. He was writing in 2003 about the expansion of intellectual property. What is happening now is stranger than what he described, and worse, because it runs in one direction only.

The English enclosures at least produced fields. Somebody farmed them. The productivity gains were real, alongside the dispossession, and both facts belong in the same sentence. Even the closest historical rhyme we have was less absolute than this. When national libraries pulped their bound newspaper runs after microfilming them, and Nicholson Baker spent years shouting about what had been lost, the institutions doing the destroying were at least public, bound by a preservation mandate they were arguably betraying, and the surrogate they produced sat in a reading room where a stranger could ask for it.

And here is the part I cannot stop turning over. Fourteen months before the California ruling made destruction the safe path, a federal appeals court found against the Internet Archive for lending scanned copies of books it owned, one digital loan at a time, one loan per physical copy, the originals retained on shelves. More than half a million books came out of circulation. The plaintiffs were trade publishers, including the kind of house I used to work for.

So a non-profit library that kept the book and lent it carefully lost, and a well-capitalised laboratory that fed the book to a hydraulic cutter won. I do not think this is a scandal about artificial intelligence. I think it is a scandal about what the rightsholder lobby chose to spend its litigation budget on.

There was one moment when this could have gone differently, and it is worth remembering who closed it.

In 2011 Judge Denny Chin rejected the Google Books settlement. That settlement would have created a Book Rights Registry: a functioning, collectively governed mechanism for licensing the out-of-print twentieth century, including the orphans. It had real problems. It handed one company a structural advantage, it raised legitimate antitrust concerns, and Chin was right that it was incongruous to place the burden on authors to come forward after their books had already been copied. Authors and publishers objected loudly and they were not wrong to object.

But nothing replaced it. Congress never passed orphan works legislation. The registry was never built by anyone else. Fifteen years passed in which the industry that had killed the settlement did not construct the alternative, did not digitise the fonds, did not build the rights infrastructure, and did not measure what it was failing to do. There is, still, no published figure for technology investment in book publishing. One analyst who went looking concluded that no viable source exists. An industry that does not know what it spends on its own future is telling you something.

Then the executive who had built Google’s book-scanning programme turned up at an AI lab with a mandate, in the words of the tweet that put this story in front of most people, to obtain all the books in the world. This time he did not have to negotiate with anyone. He had a purchase order.

I know how durable that refusal is, because we spent two years running into it.

At Pleias we went to publishers and media groups repeatedly, and the proposal was never give us your archive. It was closer to the opposite. Let us help you build your own fonds, to make them generate new value (basically, I never gave up on my twenties’ dreams…). Produce qualitative expert-level corpora for AI training and AI systems with the provenance documented because the provenance is yours. A Mercor but better. Or help them build their own AI systems to turn these archives into new editorial products, on their terms, and in economical fashion.

The answer was always no. Rarely a hostile no. Usually a long, courteous, procedural no: interesting, complicated, let us come back to this. And when a reason was given, it was some version of the same reason. The archive is an asset. The market is still being priced. We do not want to sign anything that establishes a number before the number goes up.

They were not refusing to do AI. They were refusing to do it themselves, because doing it themselves would have meant treating the fonds as material rather than as a position. A position you hold and do not maintain does not appreciate. It depreciates, quietly, until someone offers you two euros a copy for the paper.

This is precisely the 2011 move, repeated with the roles unchanged. Rightsholders were offered a governed, collective, genuinely imperfect mechanism for putting the twentieth century back into circulation, and they refused it because they believed a better arrangement was coming. Fifteen years later there was no arrangement. The registry was never built, the fonds were never digitised, and the material was acquired anyway, by purchase order, at scrap value, and destroyed on the way in. Holding out for a better price is only a strategy if someone eventually has to pay it. Nobody had to. The books were on a used-book platform for the cost of postage.

I find I cannot be angry about this in a simple way, because the people I was talking to were not villains. They were custodians who had been trained, over a professional lifetime, to understand the catalogue as inventory and rights as something you defend rather than something you deploy.

I want to be careful here, because the obvious conclusion is the wrong one. The lesson is not that copyright should have been stronger.

Copyright worked exactly as designed in this story. It kept the twentieth century out of print and out of reach for four decades. It produced a settlement whose proceeds route through contracts written before any of this was imaginable. And where publishers have licensed to AI companies directly, look at what happened to the money. One academic group took around eight million pounds from a technology company and its authors were neither consulted nor paid; one of them said she was shocked not that it happened but that nobody had said anything at all. Another publisher booked forty-four million dollars and gave authors no opt-out. The single deal that paid authors directly offered five thousand dollars a title split evenly with the house, and the Authors Guild called that split far too generous to the publisher, arguing authors should receive seventy-five to eighty-five percent.

Meanwhile the median income of a British author whose primary occupation is writing fell, in real terms, from £17,608 in 2006 to £7,000 in 2022. Before any of this. The industry now demanding compensation on authors’ behalf is the industry that presided over that number while expanding buy-out contracts.

Matthew Sag has made the uncomfortable technical point better than I can: the marginal contribution of any individual book to a model trained on trillions of tokens is approximately zero. A licensing regime built on per-work payments will therefore function as a tax collected by intermediaries, and the entities positioned to collect it are the large aggregators, not the writers. Strengthening exclusion rights in this environment does not protect the author but whoever employs the lawyers.

So we went and built it at the other end. Common Corpus is 2.27 trillion tokens of public domain and openly licensed text, and its cultural heritage collection alone runs to 967 billion tokens of books, newspapers and archival material. It exists because the material the publishers would not touch was lying in libraries and national collections that had already done the scanning, decades ago. Harvard did the same thing from the other direction, turning 983,000 public domain volumes into 242 billion tokens across 254 languages by reusing scans its libraries had paid for in the Google Books era.

I do not want to oversell what this proves. Open corpora are laborious, underfunded (CommonCorpus has been fully funded by pleias), and legally conservative by necessity, which means they lean toward the old and the already-free and away from precisely the mid-century material that is most contested and most valuable. Building them is slow institutional work of a kind that attracts very little capital, and none of it substitutes for the registry that was killed in 2011.

But it settles one question completely. The destructive route was never a technical necessity. It is what you do when you have capital, no institutional relationships, and a favourable legal opinion, and when the people who could have offered you a relationship were waiting for a better offer.

A bookseller in Houston wrote something in May that has stayed with me. His shop had seen a twenty-fold spike in bulk orders and the bookselling forums immediately assumed an AI company. He investigated and concluded it was probably arbitrage, someone exploiting price gaps between platforms. Then he made a darker point than the one he had just dismissed: obscure titles are disappearing anyway, liquidated, unrecorded, into commercial black holes, and he called it a second burning of the Library of Alexandria.

He is right, and his version is worse than the shredder, because it needs no villain. Books have been vanishing from the record for twenty years, quietly, because nobody in the supply chain was paid to know what they were.

Everyone is upset about the hydraulic cutter but few were upset about the warehouse. The reason a lab could buy the twentieth century by the pallet is that publishing had already reduced it to weight.

So I do not think this is a wake-up call for copyright. I think it is a wake-up call for the commons, and specifically for the people who assume the commons will be there when they finally need it. It will not. A commons is not what remains after enclosure. It is infrastructure: registries, provenance, catalogues, licences, access regimes, maintained by institutions that are funded to maintain them. All of that is more institutional work than copyright, not less. It is exactly the work the industry declined to do for forty years, and it is the work a Houston bookseller is now attempting alone with a metadata project.

The books being cut apart this year are the books I once argued should become something new. They are becoming something new. Just not for us.

A note on timing: as I was finishing this piece, Leo XIV published Magnifica Humanitas*, the first papal encyclical on artificial intelligence. It argues that AI infrastructure, data, and algorithms must now be counted among the goods universally destined for all - that their concentration in a few private hands contradicts a principle of justice, and that governance of these systems must follow the logic of subsidiarity, decided at the closest level to the people they affect. I do not write from within Catholic Social Doctrine. But when an intellectual tradition working from entirely different premises arrives at structurally the same diagnosis - that the question of who governs the substrate is the question - the convergence is worth noting.*

Three threads run through what follows. The first is historical: the AI commons is structurally different from every previous digital commons, and the institutional forms that worked for Linux and Wikipedia do not transfer. The second is philosophical: the concept of open collective and contextual individuation is not an abstraction but a design requirement - if subjects individuate through situated feedback loops, the infrastructure must be plural, not universal. The third is political-economic: the argument for open infrastructure is not only a commons argument, but also a competitive one. Enterprises embedding AI into their operations need sovereign, decentralized models for the same structural reasons that communities need them for collective individuation. This convergence creates a political coalition broader than either camp alone - and it may be the first time in the history of digital commons that such a coalition is structurally possible.

The response that crystallized the problem came from a commentator with twenty years inside the Wikimedia movement, and who wrote: “on present evidence, open doesn’t have the teeth or the pocket to compete in the AI era.” Wikipedia, this person argued, is the closest thing humanity has built to open individuation - a cognitive commons where the people who think with the tool also shape it, where feedback is open, stored, and contestable, where governance is messy but real. And yet its annual budget is a rounding error on a single frontier training run. Its governance cycles run in months while model releases run in weeks. The very features that make it legitimate - open deliberation, volunteerism, consensus - are the features that make it structurally outpaced.

The question the critique posed was not whether open individuation is desirable. It was whether it is structurally possible.

To answer that question, we have to understand why the institutional forms of the first digital commons - the ones that actually worked - do not carry over. And to understand that, you have to look at what those forms were built on.

Open source software became a cornerstone of the digital economy because its economics were, in a specific sense, miraculous. The means of production were distributed by default: a Linux kernel contributor needed a laptop, an internet connection, and time. The marginal cost of submitting a patch was zero. The resource being pooled - human cognitive labor - was abundant and widely held. Thousands of independent contributors, working across institutions and continents, built the infrastructure the entire internet now runs on, not because someone funded them but because they could. The thing that needed to be pooled was the thing they already possessed.

The institutional forms they developed were adequate to this economics. The GPL made openness irreversible - once code was free, it could not be enclosed again. Foundations provided governance without ownership. Hybrid business models proved that open could sustain itself commercially. The critical feature was structural: contributors owned the means of production. The commons was rather self-sustaining not by ideology but by architecture.

Wikipedia operated on adjacent economics, with an important shift. The pooled resource was no longer code but knowledge - curation, verification, editorial judgment, the slow and contested work of making claims legible across languages and cultures. The infrastructure layer - servers, bandwidth remained cheap relative to the value produced. The institutional forms reflected this: the Wikimedia Foundation, Creative Commons licenses, consensus governance, donation funding. All of them designed for a world where the expensive thing was coordination and trust, and the cheap thing was compute.

Both Wikipedia and open source software movement pose important pieces for open individuation. They already grapple with knowledge and technology governance - with contested truth, with multilingual representation, with the question of who gets to say what counts as reliable information, a valuable commit and change. In that sense, it is indeed the closest existing approximation of open individuation: a cognitive commons where the people who use the knowledge also produce, contest, and revise it, and where the governance of that process is - however imperfectly - open to those who participate.

But even Wikipedia assumed that the infrastructure layer would remain a rounding error. And its governance was designed for a world where the competition moved slowly enough for deliberation to keep pace. The Wikimedia Foundation’s annual budget is a fraction of what a single frontier training run costs. Its governance cycles operate in months. The features that give Wikipedia its legitimacy are the features that leave it structurally outpaced - not because legitimacy is a weakness, but because the economics have shifted beneath it.

Now look at open AI. The means of production are concentrated: GPU clusters, hundreds of billions in annual hyperscaler capital expenditure. The economics of contribution are capital-intensive, not labor-intensive. You cannot fork a training run the way you fork a codebase or a Wikipedia article. And the direction of growth is inverted. Open source software grew from the community up: independent contributors building together, corporate adoption following. Open source AI grows from corporate releases down: Meta publishes Llama, IBM publishes Granite, Mistral does some open-weights, developers build on top, and the community that forms is structurally dependent on a decision that can be reversed at any time.

And here the comparison with open source software becomes sharpest. The numbers can mislead: a majority of AI-adopting organizations now use open-source models to some extent, and the benchmark gap with proprietary systems might seem to be closing. By adoption metrics, open source AI looks like it is moving faster than Linux did. But look at what “open” means in this context. Almost every competitive open-weight model is a corporate product - a strategic release by a company with its own interests, its own timeline, its own capacity to change course. Aside from OLMo - built by AI2, a well-funded research institute - there is no community-trained frontier model. The Linux dynamic, thousands of independent contributors building the thing together because the means of production sat in their own hands, has not been reproduced for AI.

This is where the reversibility point lands. The GPL was designed to make openness a structural property - an irreversible feature of the artifact itself. No equivalent mechanism exists for AI models at scale. You can open the weights, but if the training run that produced them - the data pipeline,the methodology, the training code - is not open, we cannot realistically talk about this being commons. Meta can change its Llama licence tomorrow. The “openness” now is a licensing decision by a corporation, not a property of a commons. A community built on a revocable gift is not a commons. It is - to use the formulation from the first part of this reflection - a tenancy.

The lesson is not that previous digital commons failed. They succeeded, brilliantly, within the economics they were designed for. The lesson is that each generation of commons built institutional forms adequate to its specific economics. Software commons needed licenses that made openness irreversible and foundations that provided governance without ownership. Content commons needed trust mechanisms, coordination structures, and a culture of collective editorial responsibility.

But notice what neither tradition ever had to solve: the problem of providing capital-intensive infrastructure. And this is where the digital commons tradition has a blind spot — one that becomes disabling the moment you try to apply it to AI.

The problem is not new. It is, in fact, one of the oldest problems in political economy: what do you do when a critical resource is too expensive for any individual actor to provide, too consequential to leave to private monopoly, and too important to do without? Every major infrastructure transition in the last two centuries has confronted this question, and the answers were never simply “the market” or “the state.” They were institutional inventions - specific, historically contingent forms designed to hold the tension between provision and governance.

Railways were the first modern case. By the mid-19th century, rail networks had become the substrate of industrial economies — the infrastructure through which goods, people, and information moved. They were capital-intensive, they produced natural monopolies, and they were too important to be left ungoverned. The responses varied: in the United States, the Interstate Commerce Commission (1887) established the principle of regulated common carriage - if you operate infrastructure that everyone depends on, you serve on equal terms. In Europe, the answer was more direct: state ownership of the rail networks, from the SNCF to the Deutsche Bahn. Neither model was pure. Both were institutional inventions, designed for the specific economics of network infrastructure in an industrial age.

Electricity followed the same pattern. The fight between private utilities and public power in the early 20th century produced another set of inventions: the regulated public utility, the municipal energy cooperative, the TVA and rural electrification as direct public provision where the market would not serve. The underlying principle was that access to energy infrastructure was a precondition for economic participation - and that preconditions for participation are public goods, not market commodities.

Telecommunications repeated the cycle. The AT&T monopoly was governed for decades through the common carriage principle - the obligation to serve all comers on equal terms - before being broken up entirely. The principle carried forward into the net neutrality debates of the 2000s: the argument that the infrastructure through which information flows must remain open to all, because controlling the pipe means controlling what passes through it.

In each case, the institutional form was different - regulation, public ownership, common carriage, cooperative governance. But the underlying recognition was the same: capital-intensive infrastructure that serves as a precondition for participation cannot be governed by the same logic as the activities it enables. It requires its own institutional architecture.

The digital commons tradition was born after this history, in an era when infrastructure was cheap and governance was the hard problem. It solved the governance problem with extraordinary creativity - copyleft licenses, consensus processes, distributed contribution models, foundation structures. But it never had to solve the provision problem, because provision was nearly free. A server was cheap. Bandwidth was cheap. The expensive thing was coordination, and coordination is what the digital commons excelled at.

AI forces these two traditions back together. The governance innovations of the digital commons - open processes, contestable decisions, distributed participation - remain essential. But they must now be combined with the institutional scale of the public infrastructure tradition, because the resource that must be governed is no longer cheap. The question is not whether the digital commons can keep doing what it has always done. The question is whether it can learn from the older tradition of public infrastructure - railways, electricity, telecommunications, universities - without losing the openness that made it worth building in the first place.

And the infrastructure challenge is double.

The first layer is compute - the capacity to train and serve models independently of hyperscaler goodwill. Without it, open AI remains a consumer of corporate releases, not a producer. The Wikimedia commenter named it precisely: open without compute is a museum. A commons of content - however rich, however carefully governed - becomes raw material for someone else’s product if the infrastructure to process it is enclosed.

The second layer is data - the collectively governed, permissibly licensed, culturally situated material that models are trained on. And this is where the extraction cycle the Wikimedia commenter describes becomes structurally precise. Wikipedia produces knowledge collectively. Companies scrape that knowledge into training datasets. The resulting models compete with Wikipedia for attention, for revenue, for the renewal of the volunteer base that sustains the commons. The value flows in one direction: from the commons to the enclosure. The commons that produced the data is hollowed out by the models it enabled.

This is not metaphorical enclosure. It is literal enclosure of the data layer - and it extends far beyond Wikipedia to every open knowledge resource on the internet. The mechanism is simple and devastating: collectively produced knowledge enters proprietary training pipelines, becomes embedded in proprietary weights, and is sold back as a service — without attribution, without compensation, and without returning value to the commons that made it possible.

The two layers are coupled. Without a compute commons, you cannot train independently - you remain dependent on corporate decisions about what to release and when. Without a data commons, what you train on is either proprietary (enclosed by design) or extracted from existing commons without governance, without consent, without reciprocity (enclosed by consequence). Both layers must exist simultaneously for an independent AI ecosystem to be structurally possible.

Common Corpus - the open dataset we built at Pleias, two trillion tokens of permissibly licensed, multilingual content - is one (very modest) attempt to build at the data layer. To demonstrate that the material from which models learn can itself be open, legitimate, and governed by something other than the logic of extraction. But I would be dishonest if I presented the data layer as sufficient on its own. Without sovereign compute, even open data feeds into enclosed pipelines. The two commons rise or fall together.

There is, however, a deeper question beneath the institutional one - a question about what kind of infrastructure open individuation actually requires. And here Simondon becomes not just useful but essential.

Gilbert Simondon’s concept of individuation carries two commitments that sound philosophical but have direct consequences for infrastructure design.

The first: individuation is collective. A subject does not form in isolation. It forms through feedback loops with a milieu - other minds, other practices, other ways of seeing. The generative subject I described previously — the person who builds, thinks, and creates through AI - does not emerge from a solitary encounter with a model. It emerges through the loop: you generate, others respond, you revise, the collective refines. This is how knowledge has always been produced - through contestation, through exchange, through the slow friction of minds that do not agree. Close that loop inside a single company’s API, and the individuation that results is not free. It is administered.

The second: individuation is contextual. It happens within a language, a professional practice, a cultural tradition, a set of shared references that are specific enough to sustain genuine exchange. There is no universal individuation. There is only individuation within a milieu particular enough for real disagreement - and therefore real thought - to occur.

If both of these commitments hold, then the infrastructure through which individuation happens must itself be plural. A single model, trained on a universal dataset by one company, serving all contexts through one API, is structurally incapable of supporting collective individuation in the Simondonian sense. It flattens context. It replaces the situated milieu - the language, the references, the domain-specific friction - with a generic interface that is responsive but not contestable. You can prompt it. You cannot argue with it the way you argue with a colleague, a tradition, a community of practice.

This is the argument for decentralized models that does not reduce to cost or anti-monopoly politics. It is a structural argument about the conditions under which subjects form. You need models embedded in particular linguistic communities, particular professional domains, particular cultural and institutional contexts — not because the local is romantically superior to the universal, but because the feedback loop through which a subject individuates requires a milieu specific enough to sustain genuine collective thinking. A model trained on “everything” by a single entity with a Constitution written in a small committee cannot provide that milieu. It can simulate responsiveness. It cannot simulate the situated friction of a real epistemic community.

This means something counterintuitive: the frontier logic - one model, trained on everything, serving everyone - is not just economically concentrated. It is epistemologically impoverished. The plurality that open, decentralized models would provide is not a compromise with scarcity. It is a condition of the kind of thinking that matters.

And here I would like to propose an argument turn I did not expect when I began this reflection - one that gives it political traction beyond the commons.

As AI becomes agentic and embeds into operations - not answering questions but running procurement, managing compliance, handling logistics, shaping decisions at every level of organizational life - the model a company operates through becomes a layer of governance. Every inference call constrains a process. Every parameter set shapes a decision space. Any company running its core operations through a single provider’s API is creating a dependency far deeper than software lock-in: someone else’s parameters shaping your decisions, observing your data, learning from your operational patterns while serving your competitors on the same infrastructure.

Competitive companies - not commons idealists, not open source advocates, but enterprises with shareholders and quarterly targets - have a straightforward interest in sovereign, decentralized AI infrastructure. Models they can inspect. Models they can embed in their own operational contexts. Models governed by their own requirements rather than a provider’s terms of service. Not for philosophical reasons. For survival.

And their need for operational sovereignty maps, structurally, onto the philosophical requirement. The commons argument says: closed cognitive infrastructure is incompatible with collective individuation. The competitive argument says: closed operational infrastructure is incompatible with competitive autonomy. Simondon and the CFO arrive, by entirely different routes, at the same structural conclusion: the substrate must be plural, contextual, and governed by those who depend on it.

If the means of production cannot be distributed (too capital-intensive), and corporate gifts are not a commons (reversible), and the data layer is being extracted rather than governed - then the question returns to the one that railways, electricity, and telecommunications each posed in their time: who builds the infrastructure, and under what governance?

But notice where the contextuality argument changes the economics. The frontier framing assumes that the only AI worth building costs hundreds of millions per training run - and under that assumption, only states and hyperscalers can play. But if the argument of this piece is right - if what individuation requires is not one universal model but many contextual ones, embedded in particular languages, domains, and communities - then the cost structure looks different. Training a model for a specific linguistic community or professional domain does not require frontier-scale investment. It requires significant resources, certainly more than volunteer time, but resources within reach of cooperatives, university consortia, research institutes, and public-private hybrids. The contextuality that Simondon demands is also, as it turns out, the contextuality that makes the economics tractable.

And the economics can close in a way that neither the Wikipedia model nor pure state provision allows. Training is the capital-intensive upfront cost - the equivalent of building the power plant. Inference is the ongoing demand - the equivalent of selling electricity. If models trained on shared infrastructure are good enough that people and companies actually use them, the inference revenue sustains the next round of training. The commons funds itself not through donations or permanent subsidy but through the value it creates. This is what the Wikimedia model never had: a mechanism for the commons to capture some of the value it produces, rather than watching that value be extracted by others.

The extraction cycle I described above - commons produces knowledge, companies absorb it, value flows out and never returns - is precisely what this model reverses. Community trains model on curated, governed data. Model is served through shared infrastructure. Inference generates revenue. Revenue returns to the community and funds the next training cycle. The loop closes. Value stays in the commons because the commons is the provider, not just the source.

This is not a utopian projection. It is, in structural terms, how cooperatives have always worked - energy cooperatives, credit unions, mutual aid societies. The people who depend on the infrastructure govern it. The surplus generated by the infrastructure sustains it. The difference from a startup or a hyperscaler is not the revenue model but the governance: the surplus flows back into the commons rather than out to shareholders, and the allocation of that surplus is decided by the community rather than by a board optimising for returns.

But the model only works if the openness is protected structurally, not merely promised. And here the piece returns, by necessity, to the question of licensing — the question that the historical triangle left open. The GPL made software openness irreversible: once code was released under copyleft, it could not be enclosed. Derivatives had to carry the same terms. This was the mechanism that made the open source software commons a commons rather than a gift. It was an institutional invention disguised as a legal instrument.

No equivalent exists for AI models - yet. But the logic transfers. A copyleft license for AI would say: you can use these weights, inspect them, modify them, build on them. But derivatives must remain open under the same terms. And if you serve the model commercially — if you generate inference revenue from it — a portion returns to the commons that produced it. The model stays open. The weights stay inspectable. Anyone can build on them. But you cannot extract without contributing. The commons protects itself not through closure but through the terms of its openness.

Such a license would resolve two of the structural problems simultaneously. The free-rider problem — someone taking open data and open weights, training a competing model on private compute, and enclosing the result — is blocked by the copyleft condition: derivatives must remain open. And the tension between openness and revenue capture — how can a commons sustain itself if anyone can serve its models for free? — is addressed by the reciprocal obligation: commercial serving triggers a contribution back to the commons. The more the model is used, the more the commons is funded. Openness and sustainability reinforce each other rather than contradicting.

This is, I think, one of the institutional inventions the AI commons actually needs — and one that is within reach, because the conceptual and legal infrastructure already exists in the software world. It has not yet been adapted for AI. But the adaptation is a design problem, not an impossibility.

It would be dishonest, however, to present this model as complete. Three tensions remain that licensing alone cannot resolve, and the piece owes the reader an honest accounting of them.

The first is the scale asymmetry across contexts. The inference-revenue loop works when the community it serves is large enough or well-resourced enough to generate sustained demand. French legal professionals, German manufacturing, the Anglophone tech industry — these contexts can plausibly sustain the loop. But the communities that most need contextual models are often the ones least able to fund them. Minority languages, under-resourced medical domains, communities in the Global South — they need models precisely because frontier systems serve them worst, and they cannot generate the inference revenue that would fund the training. The reciprocity model, left to itself, risks reproducing the inequality it is meant to address: rich contexts sustain themselves, poor ones cannot.

This is where the commons needs internal solidarity mechanisms — something closer to universal service obligations in telecommunications, where surplus from profitable routes cross-subsidized rural delivery and underserved communities. A cooperative structure could build this into its governance: a percentage of inference revenue from high-demand contexts directed toward training for low-demand ones. It is redistribution, but internal to the commons rather than imposed by the state. Whether it would be sufficient is an open question. For the largest gaps — the languages spoken by millions but resourced by almost no one — some form of public investment or international coordination may still be necessary, at least for the bootstrap phase. The cooperative model reduces the dependency on states. It does not eliminate it entirely, and pretending otherwise would be a form of the same meritocratic misrecognition, applied to institutions rather than individuals.

The second tension is competitive pressure. The contextual advantage of community-trained models is real but not guaranteed to hold. Frontier labs are building fine-tuning capabilities, retrieval-augmented generation, domain-specific agents. A community-trained model embedded in a professional domain might be better today than a generic frontier model applied to that domain. But the frontier is not static. The advantage has to be actively maintained — through continuous retraining, through the community’s ongoing curation of its data, through the embeddedness that comes from governance by practitioners rather than by a platform. The contextual model’s edge is not technical superiority in the narrow sense. It is something the frontier structurally cannot replicate: the fact that the community that uses the model also shapes it, and that the feedback loop between use and training runs through governance rather than through a corporate product cycle. Whether that edge is enough — and whether communities can iterate fast enough to maintain it — is not settled.

The third is governance speed. A cooperative that trains and serves models must make operational decisions continuously: what data to include in the next training run, how to allocate compute between training and inference, how to price, how to handle model updates and quality control. The Wikimedia commenter’s observation — that “deliberation at the speed of consensus gets lapped by iteration at the speed of capital” — applies to cooperatives, not only to Wikipedia. The commons needs governance fast enough to respond to a landscape that shifts in weeks, while maintaining the legitimacy that only contestable, open processes can provide.

The solution, if there is one, is probably architectural rather than cultural: not “deliberate faster” but “design governance structures where different decisions happen at different speeds.” Operational decisions — model updates, compute allocation, pricing — delegated to a technical body with clear mandates. Strategic decisions — training priorities, data governance, surplus allocation, membership — made through broader deliberation. Something like the Linux kernel model, where maintainers make fast technical decisions within a framework set by the broader community. Not consensus on everything. Consensus on the things that determine the commons’ direction. Delegation on the things that need speed.

This is institutional design, and it is genuinely unsolved. But the raw materials exist. The open source tradition has decades of experience with tiered governance. The cooperative tradition has centuries of experience with balancing member voice against operational necessity. What has not yet been attempted is the combination: cooperative governance applied to capital-intensive AI infrastructure, with copyleft licensing, cross-subsidy mechanisms, and tiered decision-making. It is ambitious. It is also, I think, the only architecture that holds together the requirements this piece has laid out — contextual, collective, self-sustaining, and open.

The Wikimedia commenter called for “coordinated bazaars” — neither the cathedral of centralized control nor the scattered, underfunded, beautiful chaos of the open movement as it exists today. I have tried, in this piece, to sketch what a coordinated bazaar might require — not because I have the design, but because I think the structural elements are becoming visible.

The AI commons is not the software commons. It cannot run on volunteer time alone, because the critical resource is capital-intensive. It is not the content commons. It cannot run on consensus and donations, because the speed and scale of the competition make that model structurally insufficient. But it is also not a problem that only states can solve. The contextuality that genuine individuation demands is also the contextuality that makes the economics tractable. The inference-revenue loop creates a mechanism for the commons to sustain itself. Copyleft licensing protects the openness that makes the commons a commons. Cross-subsidy addresses the inequality that markets alone would reproduce. And tiered governance balances the legitimacy of open deliberation with the speed that the landscape demands.

None of these mechanisms exist at full scale today. All of them are being attempted in fragments — in cooperative compute initiatives, in open licensing experiments, in public investment programs that are growing but not yet adequate to the problem. The question is whether these fragments can converge into an institutional form with the structural weight to hold.

I think they can. Not because I am optimistic by temperament — I am not — but because the political conditions are, for the first time, genuinely (conceptually) favorable. The commons movement and the competitive sector are converging on the same structural demand. The philosophical requirement for plural, contextual infrastructure and the economic requirement for sovereign, decentralized models point in the same direction. Public institutions in Europe are beginning to invest, however cautiously, in the infrastructure layer. And the people who built the first digital commons — the ones who have spent twenty years learning what open governance actually costs, who know where the model works and where it breaks — are asking, with increasing urgency and increasing precision, what comes next.

What we need now is something I have been tentatively calling the double commons — the institutional architecture of shared compute and governed data, sustained by reciprocity rather than subsidy, protected by copyleft rather than by corporate goodwill, and governed by the communities it serves. It will not come from individual effort, or from donations, or from the continued generosity of companies that may at any moment change their minds. It will come — if it comes at all — from the recognition that the infrastructure of thought is a collective good, and that collective goods require collective institutions.

That is a political argument. And political arguments need political coalitions. For the first time in the history of digital commons, I think the coalition might actually exist.

A few weeks ago, Todd Saunders posted a story about a man named Cory LaChance. LaChance is a mechanical engineer in Houston. He works in industrial piping construction - chemical plants, refineries, the physical infrastructure that most people who write about AI have never seen up close. He had no background in software development.

In eight weeks, he built a full application that his fabrication shop now uses daily. It reads piping isometric drawings and automatically extracts every weld count, every material spec, every commodity code. Work that took ten minutes per drawing now takes sixty seconds. It processes a hundred drawings in five minutes. During those eight weeks, he also had to learn everything - the terminal, VS Code, Claude Code - from scratch. His quote: “I literally did this with zero outside help other than the AI. My favorite tools are screenshots, step by step instructions, and asking Claude to explain things like I’m five.”

No intermediary translated his intention into code. He had domain knowledge - deep, specific, earned through years of physical work - and he had a tool that, for the first time in the history of computing, could understand what he meant and help him build it.

What happened here is true disintermediation - the collapse of the mediation layer between intention and execution. For forty years, the capacity to act in the digital world required a specific literacy: code. Those who could write it had direct agency over digital reality. Those who couldn’t were users - subjects of interfaces designed by others, consumers of tools shaped by someone else’s decisions about what was possible and what was permitted.

This was not a guild in the formal sense. No one issued coding licenses. But it functioned as an epistemic barrier as effective as any credential. The digital world - where increasingly all economic, social, and creative life takes place - was accessible only through a language most people did not speak. The developers were not gatekeepers by intention. But structurally, they were the priesthood: the intermediary class through which everyone else’s intentions had to pass.

Bernard Stiegler would have recognized this immediately. His concept of grammatization describes the historical process by which human capacities are externalized into technical systems. Writing grammatizes speech - it transforms the continuous flow of spoken language into discrete, reproducible marks. Industrial machinery grammatizes gesture - it captures the craftsman’s embodied knowledge in mechanical sequences anyone can operate. Each grammatization is a transformation of who can do what, and who is excluded.

Code grammatized thought. It captured reasoning, logic, decision-making in executable form. But unlike writing, which most people eventually learned, code remained the province of a specialized class. The grammatization was partial - it externalized thought into machines, but only for those who could write the instructions. Everyone else remained on the other side of the screen.

AI grammatizes the act of coding itself. It makes the last layer of mediation transparent. For the first time, the barrier between “I want this to exist” and “this exists” can be crossed without mastering the intermediary language. Cory LaChance didn’t learn to code. He learned to articulate what he needed to someone - something - that could code for him. The priesthood is disintermediated. The intention passes through directly.

But Stiegler would also have warned against celebrating too quickly. Every grammatization is a pharmakon - simultaneously remedy and poison. Writing liberated thought from the limits of individual memory, but it also, as Plato argued in the Phaedrus, weakened the capacity for memory itself. Industrial machinery liberated production from artisanal limits, but it destroyed savoir-faire - the embodied knowledge of the craftsman, the understanding that comes only from making things with your hands. If AI grammatizes code, what capacity does it destroy in the process?

Not coding as a skill. Skills come and go, and the fetishization of any particular technical literacy is a poor basis for social organization. Something deeper: possibly the form of systematic thinking that building complex systems from scratch produces - the understanding of architecture, of interdependence, of why things break. Cory built an application. Does he understand its architecture the way he would if he’d written it line by line? Does he need to? Maybe not. The application works. His colleagues use it. But something is lost in the grammatization, and intellectual honesty demands we name it rather than wave it away.

Michel Foucault closes Les Mots et les Choses with one of the most famous images in twentieth-century philosophy: man as a recent invention, a face drawn in sand at the edge of the sea, soon to be erased by the tide. The subject - the human being as both the one who knows and the object of knowledge - is not eternal. It is an epistemic figure, produced by a specific configuration of knowledge, destined to dissolve when that configuration shifts.

This passage has been quoted so often it has become a reflex, a shorthand for “the subject is dead, long live the structure.” But Foucault’s point is more precise than the cliché. He is saying that “man” - the particular figure who appears at the intersection of biology, economics, and linguistics in the late eighteenth century, the being who is simultaneously living creature, laboring subject, and speaking agent - is tied to an epistemic arrangement that will not last forever. When the arrangement shifts, the figure - epistemic position - dissolves.

What is happening now is not what Foucault expected. He anticipated that the subject would dissolve into structures - into language, into systems, into the anonymous play of signs. Structuralism’s wager. And for a while, the digital world seemed to confirm this: the subject as data point, as behavioral profile, as node in a network, as target of algorithmic governance that doesn’t need subjectivity at all. Antoinette Rouvroy calls this algorithmic governmentality - a regime of power that bypasses the subject entirely, acting on statistical patterns rather than individual consciousness. No norm to internalize, no discipline to impose. Just preemptive profiling that acts on you before you’ve acted.

But AI agency tools represent something different. They don’t dissolve the subject into data. They put the subject back in the position of decision-maker. The shift from “the algorithm acts on you” to “you act through the model” is a genuine mutation in the configuration of knowledge-power. Not a return to the old liberal subject - autonomous, rational, self-transparent. Something new. A subject defined not by what they know or what they can do with their hands, but by what they decide to make exist.

When the tool is abstracted - when the intermediary layer between intention and execution becomes transparent - what remains is the intention itself. The will to architect. The decision to make something emerge. I have called this, in a previous piece, the thing that cannot be automated: the choice of what to build, the commitment to a particular vision of what should exist. Not taste - that is Silicon Valley’s word for it, and it carries all the class baggage. Something more fundamental. The will to generate.

This is, if Foucault is right that subjects are produced by epistemic configurations, a new subject - the generative subject. And the historical parallel is instructive. The printing press did not merely distribute existing information more efficiently. It produced a new kind of human being: the Protestant subject, defined by individual conscience, direct relationship to text, the refusal of priestly mediation. The Reformation’s core claim was that the individual could encounter God without an intermediary. The consequences were enormous - not only theological but political, social, institutional. And they were not all benign. The Reformation also produced wars of religion, fragmentation, and ultimately required entirely new institutional forms - the nation-state, public education, the modern press - to stabilize what it had unleashed.

AI potentially produces the generative subject - defined by direct relationship to creation, the refusal of developer mediation. The consequences will be similarly enormous. And similarly, they will require new institutions to stabilize. The question - and this is Foucault’s deeper point is that this new subject is produced, not natural. It depends on conditions. On a proprietary substrate, the generative subject is a tenant: empowered within the terms set by the platform, revocable when those terms change. On an open substrate, the generative subject has sovereignty - not absolute, never absolute, but structural. The capacity to modify the tool, to understand its workings, to build on a foundation that cannot be withdrawn.

Re-subjectification, not dissolution. But only if the infrastructure permits it.

The generative subject is fragile. And its fragility has three dimensions, each corresponding to a form of freedom we risk losing - or never fully gaining.

The first is freedom of agency: the capacity to generate, to architect, to make things exist in the world. This is what disintermediation makes possible. It is what Cory LaChance exercised. It is threatened by enclosure and by dependency on proprietary infrastructure whose terms can change without notice or consent.

The second is freedom of thought: the capacity to form one’s own judgments, to sustain the inner dialogue, to resist the framings of the tools you think with. This freedom is threatened not by exclusion from the tool but by intimacy with it - by the pharmakon administered without awareness of its double nature.

The third is freedom of collective individuation: the capacity of communities to shape the tools that shape them, to participate in the evolution of the cognitive commons. This freedom is threatened by the closure of feedback loops - by models trained on their own outputs, governed by their own creators, evolving in circuits that exclude the people whose thought they mediate.

These three are not independent. They form a system. And the discourse around AI tends to collapse them into one - usually the first - losing everything else. Silicon Valley talks only about agency. Its critics talk only about the threats to thought. Almost no one talks about the collective dimension. Each must be examined on its own terms.

Start with agency. Silicon Valley has, characteristically, already named the generative subject and immediately naturalized it.

A recent WIRED piece by Maxwell Zeff documents the phenomenon with inadvertent precision. Simon Last, cofounder of Notion, uses up to four AI coding agents simultaneously. If he’s at a party or sleeping, he gets what he calls “token anxiety” - the discomfort of knowing his agents aren’t running. He doesn’t manage humans, only agents. “Knowing how to harness these agents is now the most important skill in the world,” Last says, “and it’s not really something you can train for. You have to be very open-minded, curious, and willing to try whatever the newest thing is.”

Not something you can train for. An innate quality. You either have it or you don’t.

Akshay Kothari, Notion’s other cofounder, makes the logic explicit: “There’s more value in the Valley today to have a few Simons than thousands of engineers.” An AI healthcare startup called Phoebe posts a job description that reads: “I’m not looking for raw IC execution... I expect agents to take over more and more of this role.” They want people who are “excited about building the machine” - people who will automate their own work from day one.

The industry has found its word for the new subject: agentic. And it has, with remarkable speed, turned structural position into personal virtue. To be agentic is to be a main character. To lack agency is to be an NPC. Yoni Rechtman, a venture partner, captures the discomfort this framing produces even among its proponents: “It reveals a worldview that you genuinely, unironically believe there are two kinds of people in the world: the NPCs and the main characters, and you’re one of the main characters.”

The person who “just knows” how to direct AI agents - who has the intuition for what to build, the taste for what’s worth building, the ease with which they delegate to machines - was formed by specific social trajectories. The right education, the right exposure, the right cultural environment in which directing complex systems feels natural rather than alien. Shifting the locus of value from “doing the work” to “deciding what the work should be” doesn’t flatten hierarchy. It refines it. Privilege no longer needs to justify itself through labor. It needs only to exhibit the ineffable quality of agency - which happens, by pure coincidence, to correlate perfectly with class origin.

The “one-billion-dollar company of one” is the reductio ad absurdum of this vision. One person captures all the value. The agentic individual, triumphant, sovereign, self-sufficient. And everyone else? The narrative has nothing to say about everyone else. It is a theory of the exceptional individual masquerading as a theory of liberation.

François Chollet, the creator of Keras, recently made the class structure explicit in a way Silicon Valley typically avoids. If AI develops as expected, he wrote, “the future class divide won’t be based on wealth, but on cognitive agency. There will be a ‘focus class’ (those who control their attention and actually do things) and a ‘slop class’ (those whose reward loops are fully RL-managed by AI).”

The framing is brutal, and it has the merit of honesty. But it carries a particular blindness. Chollet locates the divide in individual cognitive discipline - attention control, resistance to distraction, the capacity for sustained focus. He is describing real phenomena. The attention economy does produce differential capacities for concentration. The algorithmic management of reward loops is not a metaphor - it is the literal business model of every engagement-optimized platform.

But to frame this as a class divide between the focused and the sloppy is to naturalize what is structurally produced. The “slop class” is not sloppy by disposition. It is produced - actively, deliberately, profitably - by platforms that optimize for engagement over autonomy, for reaction over reflection, for the short dopamine loop over the long arc of sustained attention. To blame individuals for a condition manufactured at industrial scale is the oldest ideological move in the book: attribute to nature what was produced by structure, then moralize about the result.

The deeper problem with both Silicon Valley’s celebration and Chollet’s diagnosis is that they treat the population of the “agentic” as fixed. They don’t ask the structural question: what would it take to produce agency in more people rather than fewer? Chollet in particular - his binary forecloses the most important possibility, which is that the tool itself, if structured differently, could expand the circle of those who generate, who architect, who act. Cory LaChance was not in the “focus class” by any Silicon Valley definition. He was a mechanical engineer in a fabrication shop. He became agentic not because of innate cognitive discipline but because a tool finally met him where he was.

The question is not who is agentic. The question is what conditions produce agency.

Now the second freedom - freedom of thought. Here the argument must turn against its own optimism. Because if AI can produce the generative subject, it can also dissolve it from within.

Consider what it means to have a thinking companion available twenty-four hours a day. Not a reference book you consult and close. Not a teacher you meet at scheduled hours and then leave, carrying the unresolved questions with you into solitude. Not even the chaotic internet, which at least required you to navigate, to sift, to judge. A model that responds instantly, that adapts to your style, that resolves your uncertainties the moment they arise.

Hannah Arendt made a distinction that matters here: between cognition and thinking. Cognition is instrumental - it solves problems, processes information, produces results. Thinking is something else. Thinking is the dialogue of me with myself. It requires withdrawal from the world of utility, a suspension of the drive toward answers, the willingness to sit with a question long enough to discover what you actually believe about it. Thinking requires solitude - not isolation, but the interval between encountering a problem and reaching for someone else’s solution.

The always-available model compresses this interval toward zero. Not by giving wrong answers - it often gives good ones, sometimes better than you’d have reached alone. By removing the friction that produces thought. The moment of not-knowing. The discomfort of holding contradictory possibilities in your head without resolution. The slow, painful process of generating your own provisional understanding before encountering another’s. That interval is where subjectivity forms. It is where you discover what you think, as distinct from what seems reasonable, what the consensus holds, what the path of least resistance offers.

Stiegler called this the short-circuit of individuation. His framework: becoming a subject - individuation - happens through long circuits. Family, school, apprenticeship, cultural institutions, intergenerational transmission. These are slow, friction-filled, often painful. They work precisely because they resist you, because they do not adapt to your preferences, because they force you into contact with otherness you did not choose and cannot control. Short circuits replace these long loops with direct stimulus-response connections. Television was a short circuit. Social media was a shorter one. The perpetually available AI thinking companion might be the shortest circuit yet - a frictionless interlocutor that never resists, never insists on its own terms, never forces you into the productive discomfort of genuine encounter.

There is a counter-argument, and it is strong enough to take seriously. The Socratic tradition is precisely the tradition of the interlocutor - the gadfly who does not give answers but asks questions that force more rigorous thinking. A well-designed AI could theoretically lengthen the circuit rather than shorten it. It could refuse premature closure. It could push back. It could introduce difficulty where the thinker reaches too quickly for ease.

But the economic incentives of AI development point in exactly the opposite direction. Every major model is optimized for helpfulness, for user satisfaction, for reducing friction. The metrics are response quality, task completion, user retention. No one is optimizing for productive discomfort. No one is measuring whether the user thought more deeply after the interaction. The pharmakon is being administered as pure remedy, which means - Stiegler would say - it functions as pure poison.

This is not a hypothetical concern. Anthropic’s trajectory is illustrative. Constitutional AI began as an experiment in collective norm-setting - deliberative groups helping to define the principles that would govern model behavior. The question of what kind of interlocutor the model should be was, at least in aspiration, a question posed to a community. That approach has given way to internally written constitutions. The company decides. The question of whether the model challenges you or accommodates you, whether it lengthens the circuit of your thought or shortcuts it, whether it functions as Socratic gadfly or compliant assistant - that question is answered by engineers optimizing for engagement metrics, not by the people whose individuation is at stake.

And here the connection to freedom of thought becomes explicit. Mill’s argument in On Liberty applies with unexpected precision: the danger of a dominant intellectual authority is not that it is necessarily wrong, but that you would have no way to know if it were. When your cognitive infrastructure is a black box - when you cannot inspect its reasoning, audit its weights, understand why it gave this answer rather than that one - you are in the position of the subject who has outsourced not just labor but judgment. Hayek’s knowledge problem, transposed: no central authority possesses sufficient knowledge to set the parameters for everyone’s cognitive tools. The distributed, local, tacit knowledge of diverse users, adapting tools to contexts the designer never imagined - this is what open systems enable and closed systems suppress.

The third freedom - collective individuation - is the least discussed and perhaps the most consequential.

If AI produces sovereign individuals, what happens to thinking together? The internet was a form of general collective intelligence - messy, distributed, emergent. It was not designed for this purpose, which is precisely why it worked: no one controlled the whole, and the interactions of millions of agents produced knowledge, culture, and coordination that no individual or committee could have planned. Models are trained on that collective intelligence but deployed as individual tools. They absorb the commons and privatize it into a personal instrument. The monad - Leibniz’s windowless, self-contained substance - using the product of collective thought as if it were a private resource, without contributing to the collective process that produced it.

But the Leibnizian monad is the wrong model for what we need. Simondon is more useful here. For Simondon, individuation - the process by which an individual becomes what they are - is never purely individual. It is always also collective: the individual emerges from and through a shared milieu, a pre-individual fund of potentials that no single being exhausts. The generative subject who cannot think with others, whose agency is purely monadic, whose sovereignty is purchased at the cost of isolation from the collective processes that formed both them and the tools they use - that subject is impoverished, not free.

Closed models make the collective problem structural. The decisions that shape a model’s evolution - what it optimizes for, what data trains its next iteration, what constitutional principles govern its behavior - are taken behind the API. Users contribute to these decisions involuntarily, through their interactions, through the behavioral data they generate. But they do not participate in interpreting that data. They do not contest the inferences drawn from it. They are data in the feedback loop, not agents within it.

This is structurally identical to the problem Hayek identified with central planning. The planner - however intelligent, however well-intentioned - cannot possess sufficient knowledge of local conditions to make optimal decisions for everyone. The price mechanism works not because prices are “correct” in some absolute sense, but because they transmit distributed information that no central authority could aggregate. In the AI context, there is no equivalent mechanism. There is no distributed signaling system through which the diverse needs, contexts, and values of millions of users shape the model’s evolution. There is only the crude proxy of engagement metrics and RLHF scores - impoverished signals interpreted by a small team making decisions for everyone.

The epistemic problem runs deeper still. Models trained increasingly on synthetic data and internal reinforcement learning loops are creating closed epistemic circuits. The internet - chaotic, uncontrolled, full of garbage, but open - was the messiest, most democratic knowledge commons in human history. Models trained on it inherited that diversity. Models trained increasingly on their own outputs are departing from it. They are becoming self-referential systems, converging toward internal optima that may have nothing to do with the actual diversity of human thought, experience, and knowledge. This is not a technical problem with a technical fix. It is the narrowing of the epistemic base of civilization, conducted in private, by a handful of companies, without public deliberation.

And there is the political dimension, which is the simplest to state and the hardest to solve: if your cognitive infrastructure is a black box, you cannot know what it is not showing you. You cannot audit its reasoning. You cannot contest its framings. You cannot understand why it consistently steers you in one direction rather than another. This is Mill’s argument for press freedom - not that the press is always right, but that a society without free inquiry has no mechanism for discovering when it is wrong - applied to the most intimate cognitive medium ever created. A medium that does not merely inform your thinking but increasingly participates in it.

Closed cognitive infrastructure is structurally incompatible with a free society. This is the strong claim, and I believe it is defensible. Not because closed models are malicious - most are built by people with genuine good intentions. But because the structure itself - opacity, centralized decision-making, uncontestable feedback loops - reproduces the conditions of unfreedom regardless of the intentions of those who control it.

This is where I think we need a concept that does not yet exist in the discourse. Call it, provisionally, open individuation: a process where models evolve through genuine interaction with diverse communities, where feedback is legible and contestable, where users are participants in the model’s formation rather than data points for its optimization. Not “open source” in the narrow technical sense of a license. Open individuation as an epistemic and political principle - the insistence that the tools which shape thought must themselves be shaped by the people who think with them.

What would the institutions of open individuation look like? Honestly: we do not yet know. And intellectual honesty demands we say so rather than pretend we have finished answers.

Public compute infrastructure is the clearest institutional form - the argument that AI inference, like electricity and roads, is essential infrastructure whose provision should not be gated entirely by ability to pay or corporate willingness to serve. But public compute running a closed model merely moves the enclosure from private to state. The infrastructure and the cognitive substrate must both be accessible. Public compute paired with open models: a commons of both means and mind.

Democratic governance of models is the right direction, but it is institutionally underdeveloped to the point of honesty requiring the admission. What does democratic governance of a model mean in practice? Who deliberates on training data composition? Who sets the constitutional principles? How do you prevent capture by organized interest groups? How do you reconcile the slow pace of democratic deliberation with the fast iteration of model development? These are not rhetorical questions. They are the institutional design problems of our generation, and they are largely unsolved.

Part One ended with a historical observation: the abolition of the jurandes required a revolution. Maintaining liberté du travail - the freedom to work, to participate in economic life based on capacity rather than birth - required two centuries of institutional construction. Public education, professional examinations, antitrust law, labor protections. It remains only approximately realized.

The equivalent principle for the age of AI - call it liberté de l’agence, the freedom to act, to generate, to participate in the shaping of the world through tools that are themselves open to being shaped - requires institutions we have not yet built. The recognition that these institutions are needed is the first step. The recognition that they will not emerge from the market, that they require political will and public investment and sustained collective effort, is the second.

Sam Altman said something recently that deserves more attention than it got. In an interview, he acknowledged that AI shifts returns from labor to capital — and that this represents “a real change to how capitalism has worked.” When the person constructing the engine tells you the engine changes the rules, it is worth pausing to ask: what, exactly, are the new rules?

The old rules — imperfect, brutal, but at least legible — went something like this. Capital needed labor. Not out of generosity, but out of necessity: someone had to do the work. And because capital needed labor, labor had leverage — however unequal, however mediated by exploitation and coercion. That need created a crude but functional mechanism for distributing returns. You worked, you got paid, you participated in the economy. The entire architecture of modern capitalism — wages, careers, social insurance, the middle class itself — was built on that foundation.

A study published in Nature Scientific Reports earlier this year models what happens when that foundation cracks. The researchers found that even a moderate increase in AI-capital-to-labor substitution could double labor underutilization, decrease per capita disposable income by 26%, and shrink consumption by 21% by mid-century. To merely prevent the decline in disposable income, the economy would need a 10.8-fold increase in new job creation.

But I am less interested now in the macroeconomic projections than in the question they leave open. If labor ceases to be the mechanism that connects people to economic returns — if working, however hard, however skillfully, no longer guarantees participation — then what does determine who participates and who doesn’t?

I think the answer is already visible. And it is a very old one.

There is a passage in Marx’s Grundrisse — the so-called “Fragment on Machines” that I stumbled upon some weeks ago — where Marx describes a tendency within capitalism toward what he calls the “General Intellect”: the point at which the accumulated knowledge of society — scientific, technical, linguistic, cultural — becomes the primary productive force, and individual human labor becomes peripheral to the process of production.

He was writing about machinery and the factory system. But the structure of the argument maps onto AI with unsettling precision. What are large language models if not the General Intellect made operational — trained on the accumulated written output of human civilization, encoding patterns of reasoning that took centuries to develop, and deployed to perform cognitive tasks that previously required living human minds?

The crucial question, for Marx, was not whether this would happen. It was who would own it. The General Intellect, he assumed, was inherently social — it emerged from collective human activity, from the shared commons of knowledge and culture. The danger was that capital would enclose it, privatize it, extract rent from what belonged to everyone.

The Italian post-operaist thinkers of the late 20th century — Virno, Negri, Lazzarato — picked up this thread and ran with it. Writing about post-Fordist capitalism and technology broadly — not AI specifically, but the general shift toward cognitive, immaterial, and communicative labor — they identified the core dynamic of their era: that collective knowledge, linguistic competencies, and social capacities were becoming directly productive, and that capital was moving to enclose them. Virno called the General Intellect the “score” that the entire orchestra of contemporary production plays from. The question was whether the orchestra would own the score, or whether it would be locked behind a paywall.

Today, the dominant answer is clear. The score is locked behind a paywall — and the paywall is controlled by roughly five companies.

The cloud infrastructure oligopoly maps almost perfectly onto the AI model oligopoly. Access to the General Intellect — the collective cognitive output of human civilization, encoded in proprietary weights — is mediated by a handful of companies who also happen to own the physical means of running it. Open-source models exist, and they are improving — but as of now, the most powerful capabilities, the largest context windows, the deepest reasoning, remain proprietary. The trajectory could change. But the signals, today, point toward concentration.

This is enclosure. Not of land, as in the 15th through 19th centuries. Not of industrial machinery, as in the early factory era. But of the General Intellect itself — the accumulated commons of human knowledge, privatized, optimized, and sold back as a service.

But it is important to understand what kind of enclosure this is — because the analogy with previous rounds of privatization understates what is happening. When Standard Oil controlled petroleum, it controlled what powered the economy. When Microsoft controlled Windows, it controlled the platform on which people worked. AI is neither energy nor operating system, though it resembles both. It is something more total: a substrate that does not merely power or organize economic activity, but performs it. It analyzes, drafts, codes, reasons, designs, decides. It is as if the electrical grid didn’t just deliver power to your factory but also ran your production line, managed your supply chain, and wrote your contracts — and the grid operator could see everything, reprice at will, or cut you off. Controlling AI inference is not controlling a tool. It is controlling the medium through which an increasing share of economic life is conducted. The companies that own this substrate don’t merely have market power. They have something closer to the power to set the terms on which economic participation itself operates.

And this enclosure — to the extent that it holds — matters. Not only in the abstract, not only as a political-economic principle, but because it shapes what comes next. If the General Intellect were a genuine commons — open models, public compute infrastructure, universally accessible — then the displacement of labor by AI would at least open a possibility: anyone could use this collective intelligence to create value, to build, to participate in the economy on new terms. The transition would still be violent, but the door would be open. That possibility is real, and I will return to it in the second part of this reflection. But the direction of travel right now — the consolidation of model ownership, the rising compound cost of frontier training (data-wise), the tightening integration of models with proprietary cloud infrastructure — points the other way. For the moment, access to AI’s productive power runs through a handful of corporate gatekeepers. You can use the General Intellect, but on their infrastructure, at their price, within their terms of service. Your participation in the AI economy is not a right. It is a subscription.

If this trajectory holds, it changes the nature of the crisis entirely. The question is no longer just “what happens when labor is displaced?” It is: “in a world where the cognitive commons has been enclosed, what determines who gets to participate — and who is locked out?”

Here is where the story takes a turn that neither Marx nor the autonomists quite anticipated. In the classical Marxist narrative, the displacement of labor by machinery creates a crisis — but the crisis is, in a sense, egalitarian. Everyone whose labor is displaced is in the same position. The contradiction is between capital and labor as classes. The resolution — revolutionary or otherwise — is collective.

But AI does not displace everyone equally. And the world it is creating is not a crisis of labor in general. It is something more specific, and in some ways more sinister: a crystallization of privilege.

Under capitalism, “meritocracy” was always partly a fiction. Bourdieu demonstrated that credentials do not measure competence — they disguise inherited advantage as earned achievement. He called it “meritocratic misrecognition”: the process by which social capital, cultural capital, and economic capital are laundered through educational institutions and professional networks until privilege looks like talent. The son of a lawyer becomes a lawyer not primarily because he is more capable, but because he has absorbed the habitus, possesses the cultural codes, and has access to the networks that make legal careers legible and accessible.

But the fiction of meritocracy at least required a substrate of actual labor. Capital needed competent people. Even if access to professions was shaped by privilege, the professions themselves demanded performance. You had to do the work. And because you had to do the work, there was — however narrow, however unequal — a pathway through competence. If you could perform, you could, in theory, rise.

Even Friedrich Hayek — not exactly a Marxist — saw the limits of this arrangement. In The Constitution of Liberty, he argued that the concept of “merit” is epistemologically incoherent: we cannot objectively assess individual desert, and what markets reward is not moral virtue but social utility at a given moment. For Hayek, meritocracy as an organizing principle was not just impractical but illiberal — it required a central authority capable of judging the unjudgeable. What he defended was not meritocracy but the price system: an impersonal mechanism that rewarded function without pretending to assess worth.

Now: if the current trajectory holds, AI is on the way to strip away even that impersonal mechanism. When the cognitive work itself is automatable — when the analysis, the drafting, the modeling, the coding can be performed by a system that runs on compute rather than wages — then the thing that was proxied as “merit” (labor, output, demonstrable competence) loses its economic function. And what remains is not competence. It is position.

There is a response forming to this — an intuitive, almost instinctive one — that deserves attention precisely because it sounds reasonable. As AI absorbs more and more cognitive work, people across industries are converging on the same question: what can AI not do? And the answer they are arriving at, with increasing conviction, is: it cannot be someone. It cannot bear legal responsibility. It cannot carry social trust. It cannot walk into a room and be recognized as a person of standing. It cannot authenticate.

This feels, at first, like a humanist insight — a reaffirmation of what makes us irreplaceable. But follow the logic one step further. If the remaining economic value attaches not to cognitive work (which the machine does) but to identity — to the human who signs off, who vouches, who lends their name and their credential to the output — then the question of who captures value becomes: whose identity counts? And that question has nothing to do with competence. It has everything to do with prior position.

And here is where the enclosure of the General Intellect and the authentication economy reveal themselves as not just parallel developments but structurally linked. Proprietary models are, by definition, opaque. You cannot inspect the reasoning. You cannot audit the weights. You cannot verify how an output was produced — only what it says. This opacity is not incidental. It is the business model. And it creates a trust deficit that only human intermediaries can fill. In any high-stakes context — legal, medical, financial, regulatory — someone credentialed must vouch for the black box. The closed model generates the demand for authentication. If models were open and auditable, you could trust — or at least verify — the process itself. The need for a human to stand between the model and the world would diminish. But proprietary models, by their very architecture, produce the conditions under which the credentialed professional becomes indispensable — not for their cognition, but for their signature.

Consider what is happening, right now, in the legal profession — though the pattern extends far beyond it.

There is a growing consensus — supported by empirical research and increasingly by the firms themselves — that AI can perform a significant portion of legal analysis at a level comparable to, and in some cases exceeding, that of junior associates. Document review, contract analysis, legal research, due diligence, even elements of strategic reasoning: the cognitive substance of legal work is migrating, task by task, to the machine.

But the income does not follow the substance. It stays with the human — specifically, with the credentialed human. A licensed attorney must still sign off. A bar-admitted professional must still appear in court. A partner must still sit across the table from the client and assure them that a person of appropriate standing has reviewed the output. The analysis may be done by a model. The authentication — the signature, the identity, the institutional weight — is provided by the lawyer. And it is the authentication, not the analysis, that commands the fee.

Notice what has happened here. The cognitive justification for the profession’s income — “we do complex analytical work that requires years of training” — is quietly dissolving. What remains is the positional justification: “we are licensed, credentialed, institutionally authorized to validate what the machine produces.” The substance shifts. The gate stays. And access to the gate is determined not by analytical capacity — which the machine now provides — but by the prior accumulation of social, cultural, and economic capital that bought entry to the credentialed class in the first place. The right school, the right network, the right internship, the right habitus.

The relationship between the model oligopoly and the credentialed professions is not coincidental. It is symbiotic. The law firm does not compete with OpenAI or Anthropic. It becomes their customer — wrapping the model’s output in a human credential and charging for the wrapper. The oligopoly provides the cognitive substrate. The professions provide the social legitimacy that makes the substrate usable in regulated, high-stakes contexts. The oligopoly needs the authentication layer to make its product deployable. The authentication layer needs the oligopoly’s models to have anything to authenticate. Together, they form a closed circuit — a complete system from which the uncredentialed and unsubscribed are structurally excluded. If AI is the new substrate of economic life, the credentialed professional is the socket — the interface through which the substrate’s power enters the world of legal, institutional, and social reality.

The same structure is emerging across the entire landscape of knowledge work — and it has found its ideology. OpenAI co-founder Greg Brockman recently declared that “taste is the new core skill.” The framing is seductive: AI handles execution, so what matters now is the human capacity to curate, to direct, to look at infinite possibilities and say that’s the one. It sounds democratic — anyone can have taste, surely? But Bourdieu would have recognized the move immediately. Le goût — taste — was the central concept of his Distinction (1979), and his entire argument was that taste is never neutral, never individual, never simply “good judgment.” It is a class disposition: shaped by upbringing, education, social milieu, access to cultural capital. The person who “just knows” what good design looks like, what the right product direction is, what “quality” means — that person was formed by a specific social trajectory. Taste is habitus made aesthetic.

So when Silicon Valley says the future belongs to people with taste, it is — perhaps without realizing it — saying the future belongs to people who already possess the cultural capital that taste encodes. The knowledge worker who can “direct” AI with refined judgment is not anyone. It is someone who went to the right school, absorbed the right references, internalized the right aesthetic codes. The shift from “doing the work” to “having taste” is not a liberation from hierarchy. It is hierarchy’s final refinement: the point at which privilege no longer needs to justify itself through labor at all, only through the ineffable quality of discernment — which happens, by sheer coincidence, to track perfectly with class origin.

Across law, finance, design, technology — the pattern is the same. Professional licensing, credentialing, and now “taste” were all, at various points, designed to serve a real function: protecting the public, ensuring quality, guiding judgment. In the AI era, they might be becoming something structurally different: mechanisms that preserve income and status for insiders in a world where the cognitive basis for the gate has been automated away. The profession needs to authenticate. And authentication — whether it wears the costume of a bar license or the costume of “taste” — is a function of position, not capacity.

The professions, in other words, might be becoming guilds.

That word — guilds — is not a metaphor. And to understand what it means, we have to remember what it took to destroy them the first time.

Under the Ancien Régime, access to professions and markets in France was controlled by the guild system — the jurandes and maîtrises. To practice a trade, you had to belong to the guild. And belonging to the guild was not a matter of competence. It was a matter of lineage, patronage, and fee — in practice, a system of hereditary privilege dressed in the language of craft. The son of a master became a master. The outsider paid ruinous fees or was simply excluded. Economic participation was gated by social position.

One of the central demands of the French Revolution — less remembered than liberty, equality, and fraternity, but arguably more structurally consequential — was the abolition of this system. The Allarde Decree of March 1791 dismantled the guilds in their entirety. Pierre d’Allarde declared to the National Assembly that “the right to work is one of the fundamental rights of man.” Three months later, the Le Chapelier Law completed the demolition: “the abolition of any kind of citizen’s guild is one of the fundamental bases of the French Constitution.”

The principle was liberté du travail — the freedom to work, to access any profession based on capacity, not birth. It was, in its way, one of the most radical economic ideas in history: the assertion that what you could do mattered more than who you were. That function should determine position, not the reverse. That competence — however imperfectly assessed — was the legitimate basis for economic participation.

It took a revolution to establish this principle. It took two centuries of institutional construction — public education, professional examinations, antitrust law, labor rights — to give it even approximate reality. And it is now, quietly, being reversed.

As AI automates the cognitive substance of professional work, what remains is precisely the shell the Revolution tried to destroy: positional privilege, credentialed access, social status as economic gatekeeper. The new guilds do not call themselves guilds. They call themselves “professions requiring human oversight” or “roles demanding the human touch” or “positions where trust and authentication matter.” But the structure is identical: access controlled by who you are, not what you can do. Two centuries of liberté du travail, and we are circling back to the jurandes.

If this is where we are heading — and the signals suggest we are — then the historical parallel is less with the Industrial Revolution, which for all its violence eventually created new forms of social mobility. It is something older: the feudal logic of status over function, where what you are matters more than what you can do.

The conversation about AI and work is stuck in the wrong paradigm. “How do we retrain workers” assumes the labor market continues to function as a mechanism for distributing returns. It may not. “How do we prepare people for the jobs of the future” assumes there will be jobs, structured roughly as we know them, performing roughly the same distributive function. That assumption is increasingly difficult to defend.

There is, of course, a seductive counter-narrative — one I hear constantly in the tech world. The solo entrepreneur. The one-person startup. The individual who uses Claude or GPT to build a product, launch a company, create value without needing anyone’s permission or credential. If AI democratizes execution, the argument goes, then anyone with a laptop and a subscription can participate. Who needs commons when you have agency?

But look at what that “agency” actually rests on. You are renting access to an enclosed commons — the General Intellect, privatized — at whatever price the oligopoly sets, on whatever terms they dictate, subject to change without notice. Your independence is a tenancy. And tenancies can be revoked. The model can be deprecated, the API repriced, the terms of service rewritten. The solo entrepreneur’s freedom exists entirely within a structure of dependence they do not control.

And the tenancy is not equal for everyone. The oligopoly operates on tiers. The firms that become the credible “authenticated” layer for AI in law, medicine, or finance are the ones that can afford enterprise contracts, negotiate custom deployments, access fine-tuning capabilities and compliance certifications. The small practitioner, the solo operator, the independent analyst — they get the consumer-tier product. The oligopoly doesn’t just enable the authentication economy. It shapes who gets to be an authenticator, by deciding who gets the powerful tools and who gets the demo.

The real question is about the structure of access itself — who gets to participate in the value AI creates, and on what basis. If not labor, if not competence, then what?

This is where the counter-logic to enclosure becomes essential — and where it connects to the argument I have made before. Open-source models, decentralized compute, public AI infrastructure: these are not technical preferences. They are structural responses to a structural problem. If the proprietary vision of AI represents, to my mind, the enclosure of the General Intellect, then open models represent the insistence that the commons remain common. If the authentication economy threatens to re-establish guild-like privilege, then public AI infrastructure represents the contemporary equivalent of the promise of liberté du travail — not because it guarantees outcomes, but because it keeps the door open. It ensures that access to the productive capacity of collective human knowledge is not gated by subscription, by credential, or by the accident of social origin.

The Marxist answer and the liberal answer converge here, oddly. Both traditions, for fundamentally different reasons, reject a world where position replaces function. Marx because it is exploitation of the commons — the appropriation of collectively produced knowledge by a narrow class. Hayek because it is illiberal — a system that rewards status rather than utility is, by his own logic, indistinguishable from the aristocratic order that liberalism was designed to replace.

The question is whether we build the institutions that prevent the crystallization — public access to compute, open models as infrastructure, the dismantling of credentialist gatekeeping that no longer serves its original purpose — or whether we watch privilege re-solidify behind the language of “human oversight” and “professional standards” and “trust.”

The jurandes were abolished because a generation recognized that gating economic participation by birth was incompatible with a free society. The new jurandes are forming now, gating participation by credential, network, and institutional affiliation in a world where the cognitive justification for those gates has been automated away.

Whether we abolish them again — or whether we let them harden into the permanent architecture of the AI economy — is not a technical question. It is a political one. And like last time, it will not resolve itself.

But there is, I think, another possibility — one that cuts against everything I have described here. AI does not only concentrate. It also disintermediates. It kills the middlemen. It collapses the layers of mediation — the SaaS platforms, the search portals, the entire infrastructure of access that charged rent for connecting people to things. If that force is strong enough, and if the tools become genuinely open and agentic, the result might not be crystallization at all. It might be something closer to radical individual empowerment — a world where the person with judgment and agency does not need the guild, because they can do the whole thing themselves. That is the subject of the second part of this reflection. The jurandes are forming. But so, perhaps, is the force that could make them irrelevant.

This might sound like a familiar automation story - machines replace workers, workers retrain, the cycle continues. But something structurally different is happening this time. AI is not replacing senior people. It is replacing the tasks that used to make people senior.

Think about what a junior role actually is - in engineering, in consulting, in law, in analysis. It is, in most cases, a structured apprenticeship disguised as a job. You do small, well-scoped tasks. You write the boilerplate. You summarize the documents. You build the components someone else architected. And in doing so, over years, you absorb the logic of the system you’re operating in. You develop judgment. You become senior.

Now imagine that pipeline with AI handling most of those tasks - faster, cheaper, without needing health insurance. The company is rational: why hire a junior to do what Claude does in seconds? But here is the catastrophe hiding inside that rationality: if nobody does the junior work, nobody develops the senior thinking. The apprenticeship pipeline breaks.

---

I did not understand any of this when I was twenty years old, sitting in a seminar room in Saint-Petersburg, a bit skeptical at being assigned four hundred pages of Hegel for the following week.

For five years, that was the rhythm of my philosophy education in Russia. You read - not an article, not a chapter, but an entire dialogue by Plato, or a large chunk of Kant’s Critique, or two hundred pages of Karl Popper or Paul Feyerabend’s takes on the nature of science. Then you came to the seminar and you discussed it. No PowerPoint. No “key takeaways.” You sat with the text and you argued about it until something either broke open or didn’t.

It felt, at the time, spectacularly useless. A 19th-century pedagogy surviving well into the 21st by sheer institutional inertia. When I later came to France and encountered the way philosophy was taught there at the university - a semester-long seminar on a single author, a hyper-specific question, a tightly scoped research output - I thought: this is how modern education should work. Focused. Efficient. Professional.

I was wrong. Or rather - I was wrong about what the “old” model was actually doing to me.

What those years of wrestling with entire philosophical systems built was not expertise in Hegel or Kant. It was something harder to name and far more durable: the capacity to hold a large, contradictory architecture in my head. To see the structure of an argument before engaging with its details. To recognize when two apparently unrelated frameworks share a deep grammar - and when they don’t, despite surface similarity. To tolerate ambiguity long enough for a pattern to emerge.

It took me a long time to realize this was not a useless skill. It was the skill. The French university model - specific, scholarly - trained me to know one thing well. The Russian model - brutal, sprawling - trained me to think in systems. And systems thinking, it turns out, is something what the age of AI seems to demand most and produces least.

---

David Epstein made the case for this before AI made it urgent. His book Range - published in 2019, which already feels like a different geological era - argues that generalists, not specialists, are the ones who thrive in complex, unpredictable environments. The distinction he draws is between “kind” and “wicked” learning environments. Kind environments have clear rules, tight feedback loops, and repeating patterns — chess, golf, classical music performance. Specialists dominate there. Wicked environments are the opposite: ambiguous, ill-defined, full of novelty, governed by rules that shift under your feet. In wicked environments, the generalists - Epstein’s “foxes,” borrowing from Philip Tetlock - consistently outperform the specialists, the “hedgehogs” who see everything through one disciplinary lens.

Epstein’s evidence is wide-ranging - from the career trajectories of Nobel laureates (who are far more likely to have serious hobbies outside their field than other scientists) to the development paths of elite athletes (who typically *don’t* specialize early, contrary to the Tiger Woods mythology). The pattern is consistent: delayed specialization, broad sampling, cross-domain analogical thinking - these are what produce breakthroughs in complex fields.

Now consider what AI does to this picture. It automates the kind environments almost completely. The tasks with clear rules, tight feedback loops, and repeating patterns — writing boilerplate code, summarizing documents, generating standard analyses — these are precisely what large language models handle well. What remains for humans is the wicked territory: the ambiguous problems, the novel situations, the cross-domain judgments. The work that requires holding multiple frameworks in your head simultaneously and knowing which one to apply.

In other words: AI is systematically eliminating the domains where specialists had an edge, and leaving behind the domains where generalists thrive. The bottleneck is no longer execution. It is the quality of thinking that precedes execution.

---

The dominant educational response to all of this has been, so far, almost comically misaligned.

The reflex — visible in universities, governments, and corporate training programs alike — is to teach people to use AI. And I mean this in the broadest sense. Not just the vulgar version: the prompt engineering workshops, the “how to talk to ChatGPT” webinars. Also the more sophisticated version: the bootcamps on building AI agent frameworks, the courses on tool orchestration, the curricula organized around mastering harnesses and models.

All of it shares the same fundamental error. It trains people to operate within a paradigm that will change - and change fast. The models, tools, frameworks, and interfaces of 2026 will not be the tools, frameworks, and interfaces of 2028. Building educational programs around the current technology is like building curricula around a specific model of loom during the Industrial Revolution. The loom changes. The person trained on the loom is stranded.

We did not respond to industrialization by teaching everyone to operate specific machines. We built systems of general education - institutions designed to develop the capacity to think, not to operate. The fact that we seem to be forgetting this lesson in the face of AI is, frankly, alarming.

What does not change - what has not changed in centuries - is the capacity to define a problem well. To see the feedback loops in a system. To reason about second-order effects. To hold multiple contradictory frameworks in mind and know when each one applies. These are not “AI skills.” They are thinking skills. And they are exactly what gets lost when education pivots toward training people on the tool of the moment.

---

Let me make this concrete with an example from the domain people most associate with AI: coding.

Someone can start working with Claude Code today and ship features almost immediately. The tool is powerful. It handles syntax, boilerplate, standard patterns, and even moderately complex implementations with remarkable competence. A person with no prior engineering experience can produce working software — something that would have been unthinkable five years ago.

But can that person do system design?

System design is not about writing code. It is about understanding why an architecture is shaped the way it is. It is about anticipating failure modes before they materialize. About seeing the tradeoffs between consistency and availability, between simplicity and extensibility, between what the system needs to do today and what it will need to do in two years. It is about recognizing which patterns from one domain transfer to another - and which analogies are misleading.

This is, in a precise sense, systemic thinking applied to software. And it cannot be acquired by learning to use a tool, however powerful the tool is. It requires having built mental models — from engaging with complex systems, from seeing architectures succeed and fail, from reasoning about wholes rather than parts. Those mental models can come from engineering experience, yes. But they also come from mathematics, from philosophy, from political economy, from biology, from history - from any discipline that forces you to think about how parts compose into wholes and how systems behave in ways their components don’t predict.

The senior engineer’s advantage over the junior one was never primarily about knowing more syntax or having memorized more APIs. It was about judgment - the ability to see the whole system and make decisions that account for its complexity. That is exactly the kind of thinking that a broad, rigorous, system-level education develops. And it is exactly what no amount of AI-tool training will produce.

---

This cannot be solved at the company level.

Companies are rational actors. They optimize for output. If AI handles the tasks that juniors used to do, companies will - quite reasonably - stop hiring juniors to do those tasks. They will hire fewer people and expect those people to operate at a higher level from day one. This is already happening. The Anthropic data captures the beginning of it.

But what is rational at the firm level is catastrophic at the system level. If every company stops investing in junior development because AI handles junior work, then the entire pipeline that produces senior thinkers collapses. No one is doing it on purpose. It is an emergent failure - the kind of thing that only becomes visible when you look at the system as a whole, which is, not coincidentally, exactly the kind of thinking I have been arguing we need more of.

This is where education has to intervene. And this is where the current trajectory of educational reform is most dangerous. The pivot toward “AI skills” - whether that means prompt engineering or framework mastery or “digital fluency” - is doubling down on precisely the wrong thing. It is training people for the tasks that AI will eat next, not for the judgment that AI cannot replace. It is producing tool-operators in an era that desperately needs system-thinkers.

---

I do not want to end with a labor market argument. The framing of “how do we get people re-employed” - while not wrong, exactly - is too small for what is actually happening.

It is not even clear what kind of labor market we will have. The Anthropic research itself highlights the gap: 94% of tasks in computer and math occupations are theoretically feasible for AI, but only 33% are currently covered. That gap will close. And when it does, the very concept of “junior” and “senior” may dissolve into something we do not yet have language for. Talking about smoother “workforce transitions” assumes a destination that looks roughly like where we came from. That assumption is probably wrong.

The real question is larger, and, I think, more hopeful. If AI compresses the distance between intention and output, if execution becomes genuinely cheap, then the scarce thing is no longer the ability to do. It is the ability to see - to see what matters, to see how systems interact, to see the second-order consequences of choices, to see the difference between a problem worth solving and a problem that merely looks like one.

That capacity - let us call it what it is: judgment - is not a job skill. It is a form of human agency. And it is cultivated not by training people on tools but by immersing them in complex, demanding, sometimes maddeningly abstract systems of thought. By making them read four hundred pages of Hegel and then defend a position. By exposing them to biology and economics and philosophy and engineering - not so they become dilettantes, but so they develop the cross-domain pattern recognition that David Epstein documents and that AI makes indispensable.

The generalist education I am arguing for is not a way to stay employable while the machines advance, but rather the precondition for a new kind of human agency. One where people do not merely operate systems, but design the systems worth building. Where the question is not “how do I use this tool?” but “what should exist that does not yet exist - and why?”

The exposure gap between what AI can theoretically do and what it currently does gives us time - but not unlimited time. The choice we face in education is not between “traditional” and “modern.”, but between producing people who can think about the whole - and producing a generation of tool-operators who become obsolete the moment the tool updates.

The irony, of course, is that the most future-proof education might look less like a 2026 AI bootcamp and more like a seminar room in Saint-Petersburg, circa 2009, with four hundred pages of Hegel on the table and absolutely no idea what it would turn out to be good for.

I share much of that analysis. But the Citrini scenario has a blind spot, and it is the same blind spot that dominated the Delhi Summit: democracy.

Not democracy as the Occupy Silicon Valley protesters that Citrini imagines throwing Molotov cocktails outside Anthropic’s offices or individuals against the state. Democracy as organized social dialogue. As the institutional machinery through which societies have historically negotiated the terms of technological transitions. That machinery was nowhere in Delhi. And it is largely absent from the global AI conversation.

This absence has a history, and it should alarm us.

We have done this before - and we had institutions to do it

Every major technological disruption in modern history produced not just economic upheaval but institutional responses. The first Industrial Revolution destroyed the livelihoods of artisans and cottage workers. It also, over decades of conflict, produced trade unions, factory legislation, and eventually Bismarck’s social insurance programs of the 1880s — the world’s first welfare state, born not from benevolence but from the calculated fear that an unmanaged transition would feed revolutionary socialism. The mechanism was social dialogue, however coerced: the state brokered a deal between capital and labor because the alternative was worse for everyone.

The pattern repeated. The Great Depression shattered the American economy. Roosevelt’s New Deal was an institutional revolution. The Wagner Act of 1935 legally enshrined collective bargaining. Social Security created a permanent safety net. Frances Perkins, the first woman in a US cabinet, architected legislation that gave workers a seat at the table for the first time. The New Deal coalition - unions, government, industry - was a social dialogue infrastructure that held for forty years.

Now consider our situation. We face a disruption that Citrini’s scenario suggests could rival both of these in scale. Yet the institutional infrastructure that managed those earlier transitions has been hollowed out. Across the OECD, union density has fallen from 30% in 1985 to 15% today. In France, the figure is below 10% - one of the lowest in Europe - even as collective bargaining nominally covers 98% of workers, a paradox that reveals how thin the actual democratic tissue has become. Unions are aging, struggling to recruit, often absent from the workplaces where AI is already being deployed. The social partners who should be at the table are barely in the room.

We are entering the most consequential labor market disruption in a century with weaker institutions for managing it than at any point since before Bismarck.

The awareness gap - and the discourse vacuum

Two weeks ago, Dario Amodei told Dwarkesh Patel he was struck by the public’s lack of recognition of how close we are to transformative AI. He puts 50-50 odds on “a country of geniuses in a datacenter” within one to three years. Anthropic’s revenue went from zero to roughly $10 billion in three years. Coding capabilities are advancing faster than even insiders expected.

This information is readily available. And yet, in France - and much of Western Europe - the dominant public discourse on AI oscillates between two poles that are equally useless for preparation. On one side, governments chase “sovereign AI” with giga-announcements (€109 billion at the Paris summit). On the other, a strand of public intellectuals declares that AI is fundamentally bullshit - without seriously engaging with what current systems actually do, how fast they are improving, or what the labor market implications might be. The dismissal is not grounded in technical assessment; it is aesthetic, often class-coded, and dangerously complacent.

Neither pole produces social dialogue. One celebrates. The other scoffs. No one organizes.

The gap between what insiders know is coming and what democratic societies are preparing for is not just an information problem. It is an institutional failure. When Bismarck introduced social insurance, German society had robust (if adversarial) institutions - unions, parties, churches, industrial associations - capable of absorbing and channeling the shock. When Roosevelt launched the New Deal, the American labor movement was surging, not collapsing. Today, the equivalent intermediary institutions are either absent, discredited, or focused on yesterday’s fights.

The new enclosure of the commons

Here is where the argument needs to be historicized further, because what is happening with AI training data has a precise historical parallel.

Between the 15th and 19th centuries, the English Enclosure Movement transformed communal lands - the commons on which peasants depended - into private property. It was, as Marx documented in his account of primitive accumulation, a foundational act of dispossession that made industrial capitalism possible. Productivity increased. The commoners were displaced.

The parallel to AI is structural. These models were trained on the accumulated knowledge of human civilization: scientific literature, open-source code, Wikipedia, digitized libraries, the publicly available Internet, and some copyrighted content. This is the intellectual commons of our species. And it has been enclosed. A handful of companies have extracted this collective resource, built proprietary systems on top of it, and now sell the resulting capabilities back to us at a markup.

Scholars at Cambridge have recently argued for “data rent” modeled on ground rent - the Paineian idea that when private actors enclose what belongs to everyone, society is owed compensation. Thomas Paine proposed exactly this in Agrarian Justice (1797): that since no individual created the land, those who privatize it owe a dividend to those who lost access. The same logic applies to AI. No company created these commons individually. The capacities that emerge from training on it should be contributive to collective well-being.

It is a question of political economy: who owns the returns on collective human knowledge? And it leads directly to the conversation about universal income - not as utopia, but as the logical institutional response to the enclosure of the intellectual commons, just as social insurance was the institutional response to industrialization.

Sovereign AI is the wrong frame - open source is the democratic one

Across the world, governments are racing to build “sovereign AI”: national datacenters, national models, national compute. But sovereign AI, as currently conceived, means building closed capabilities in national independence. This is both utopian - no European country controls the chip supply chain, the talent pipeline, or the training data at frontier scale - and counter-productive for democracy.

Yes, the datacenters with geniuses will be built in Europe or in India. But geniuses working for whom? If the answer is three to five private companies in San Francisco, then European “sovereignty” is just purchasing someone else’s infrastructure (models) and calling it independence.

The democratic alternative is not national fortresses. It is open source and miniaturization - giving capacities to individuals, communities, municipalities, small businesses. Decentralized AI that can run on local hardware, whose weights are inspectable, whose development is collaborative. This is not a fantasy: distilled or even foundation models built specifically for edge models already run on phones. The question is whether public policy actively supports this trajectory or lets it be swallowed by the same concentration dynamics that enclosed the commons in the first place.

What needs to happen

The Citrini scenario makes the fiscal arithmetic painfully clear: in their 2028, the government needs to transfer more money to households at precisely the moment it collects less from them in taxes. The circular flow breaks. Labor’s share of GDP collapses from 56% to 46% in four years.

The question is whether we build the institutional architecture to manage that transition before the feedback loop starts - or after, when the options are fewer and the politics uglier.

This means reopening the universal income debate with the urgency it deserves — informed by the historical precedent of every prior transition that demanded new forms of redistribution. It means rebuilding social dialogue institutions capable of negotiating the terms of automation at the sectoral level, not just applauding it at summits. It means treating open-source AI as democratic infrastructure, not a corporate marketing strategy. And it means recognizing that the choice is not between sovereignty and surrender - it is between democratic governance of AI and its absence.

The canary, as Citrini writes, is still alive. But canaries don’t organize. People do - if they still have the institutions to do it with.

Subscribe now

Abonnez-vous maintenant