The Fence and the Commons – Libertarianous

When Benjamin Tucker counted the monopolies that kept labor from its product, in 1886, he named four: money, land, tariffs and patents. He put money first. The patent was a side door, a privilege for inventors and for the firms that bought them out. Today the side door is the main entrance. Software, data, designs, brands and trained models make up a growing share of what firms own, and nearly all of it rests on a legal right to stop other people copying.

Chapter 30 set the benchmark this part has used throughout. Cost is the limit of price. Competition, left free, pushes a price down toward what it took to make the thing and bring it to market, and whatever stays above that line needs explaining. Here the benchmark meets its strangest case. A song, a program, a drug formula or a model costs a great deal to make once and almost nothing to copy. If competition were free to work, the price of the next copy would fall toward almost nothing. For many such goods it hasn’t.

This chapter follows that puzzle in three moves. First the toll, the fences that hold the price of copies far above their cost, and who pays at the gate. Then the newest capital those fences build, the trained AI model, made mostly of everyone’s work and owned by a few. Last, what people build where knowledge is cheap and nobody is allowed to fence it.

The toll on copying

The first copy of a song, a program or a drug is costly. Every copy after it comes closer and closer to free. In the United States, copyright now runs for the author’s life plus seventy years, under the Copyright Term Extension Act of 1998, and a patent runs twenty years from filing. The first American copyright law, in 1790, gave fourteen years, renewable once. Cost the limit of price would cover the first copy. It would not hold a margin for a lifetime and seven decades more.

Even the name does work. Richard Stallman has long pointed out that “intellectual property” lumps together copyright, patents and trademarks, three laws with different origins and purposes, and by calling them property borrows the moral weight of a house or a coat. Nobody loses a coat when a song is copied. The original stays where it was. The word for what the fence collects tells the history too. A royalty was first a right belonging to the crown, like the royal mines of gold and silver, and patents began as royal grants, letters patent, until the Statute of Monopolies of 1624 limited the king’s power to hand them out. The toll still carries the king’s name. And like a crown grant, it goes to whoever holds the right, which is often not the maker. Labels and publishers own much of the music, employers own most patents, and many musicians earn little from the copies that pay their labels. The coat that no one loses is the heart of Roderick Long’s “The Libertarian Case Against Intellectual Property Rights,” first published in 1995. Patents and copyrights began as monopolies granted by the state. An idea, once shared, lives in other people’s minds and on their paper, and as Long put it, “you cannot own information without owning other people.” In Part I’s terms that is an agency claim. The holder gets a say over what others may do with their own hands, and over what they make from the work, since a derivative counts in law as partly the holder’s. One person’s acts are counted as another’s, the move Part I traced from the slave to the employee. And information, unlike land, isn’t scarce. Stephan Kinsella reached the same conclusion from the property side, as Chapter 35 showed. The case against the fence comes from this book’s side of libertarianism as much as from anywhere.

The strong case for the fence deserves to be put as its defenders would put it. The first copy is a real cost. A new drug can take a decade and a fortune in trials, and someone has to pay. Without a period of exclusive sale, the argument goes, nobody would pay it, because rivals could copy the result the day it appeared and sell it at the cost of the pills. Ayn Rand defended patents as the inventor’s property in the product of his mind, and the economists’ incentive theory says the same thing in drier terms.

The question this book asks is narrower. Does covering the first copy require a fence on strangers? There are other ways to cover it. A first mover has lead time and sells to the people who want the thing now. Prizes, commissions and subscriptions pay for the work up front. Open-source firms sell support, hosting and custom work around code anyone may copy. In 1998 John Kelsey and Bruce Schneier described a “street performer protocol,” in which a work is released once enough people have pledged to pay for it. And the record is less kind to the fence than its defenders suppose. Michele Boldrin and David Levine, in Against Intellectual Monopoly (2008), look at James Watt’s steam-engine patent, which Parliament extended to 1800. While it ran, the efficiency of steam engines barely moved, and Britain added about 750 horsepower of steam a year. In the thirty years after it lapsed, Britain added more than 4,000 a year, and fuel efficiency rose about fivefold between 1810 and 1835.

That leaves the question every maker will ask. When has someone been paid in full? No one has to set the number by decree. Take the fence away and others may sell copies at about what copies cost. The maker keeps a head start and whatever was paid up front, and earns the rest by being first, being trusted or making the next thing. Nor do the first buyers have to carry it all. Some pay more because they want the thing now, as with the hardcover before the paperback, and they choose to. Some pledge together before release, so the pledges are the sales estimate. Some schemes credit early backers as later sales come in, so the cost of the first copy ends up shared by everyone who bought. Without a fence, the guessing risk falls on the maker, not the buyer. The question stays open for a lifetime only because the fence lets the seller answer it alone.

Then there is who pays. A patent binds people who never dealt with the inventor. The rival who works out the same thing on his own is barred. The copier never agreed to anything. The patient who can’t afford the drug at the patented price is outside the deal altogether. It is the test Chapter 34 put to the mill upstream from the farms, a cost carried by people who were never asked. The toll even has its troll. A firm that makes nothing and buys patents only to demand payment from people who do make things is called a patent troll, after the troll under the bridge in the old Norwegian tale, who built nothing and charged everyone who crossed. The rest of the system differs from him in degree, not in kind. It, too, sits on a crossing someone else built and collects.

Data is the newer toll, and it isn’t property in the old sense at all. It is gathered as a by-product of people using a service, their searches, purchases, routes and messages, and pooled by the firm that runs the service. Nobody made it the way an author makes a book. The firm owns the pile, though not the people in it, and sells access to it, or sells what it predicts.

The people the data describes are outside that sale. They didn’t set its price, can’t see who bought it and can’t take their part back out. The toll that once ran on information people made now runs on information about them too.

It is also the fastest-growing toll. Jonathan Haskel and Stian Westlake showed in Capitalism Without Capital (2017) that investment in intangibles (software, data, designs, brands, research) overtook investment in machines and buildings in the United States by the mid-1990s, and across Europe and the United States together before the 2008 crash. Much of what makes an intangible worth owning is the right to keep others from copying it.

Carlo Vercellone gives the pattern a name. Rent, in his account of what he calls cognitive capitalism, is the other face of the common. Land was enclosed once and knowledge is enclosed now, and scarcity that nature doesn’t supply is supplied by law. Profit and rent blur, because the firm earns less from organizing work than from holding a gate. This book parts from Vercellone’s Marxian frame, which looks for the key in the law of value and its crisis. The observation stands without it. Where scarcity is made by law, the return on it is rent.

Near zero isn’t free

Jeremy Rifkin gave the trend its best-known name, but the libertarian left got there first: Warren’s cost principle in the 1850s, Tucker’s attack on the patent monopoly in the 1880s, and in Rifkin’s own decade Benkler and Carson. In The Zero Marginal Cost Society (2014) he argued that competition and technology drive the cost of one more unit toward zero, that goods become nearly free, and that capitalism, which lives on margins, undoes itself. He is right about the direction, but his version is the softer one. He expects capitalism to shrink to a junior partner beside a “collaborative commons,” and says little about the privileges that keep its margins up.

But competition doesn’t shrink cost to nothing. It shrinks the margin between cost and price. People still have to be paid for their work unless they give it, so neither of Rifkin’s words fits. Cost falls but doesn’t reach zero, and price falls toward cost, not to free. That is Josiah Warren’s rule, free of profit, yes, freed from paying more than the thing cost, not the end of production. Eric Raymond made the point from the other side in “Zero Marginal Thinking” (2014). Rifkin forgets fixed costs, he wrote, and someone still has to make the first copy and be paid for it.

For pure information the extra copy really does cost almost nothing, and the cost that remains is the first one. Warren’s rule answers there too. A copy should sell for about what it costs, plus a share of making the first, and that share shrinks with every buyer. A price that stays high long after the first copy has been paid for is the fence at work, not the cost.

Rifkin also skips the question that decides who gains. Kevin Carson, in The Homebrew Industrial Revolution (2010), argues that the job as the main way to make a living was partly an artifact of costly machinery only the rich could buy. Cheap tools lower the barrier to workers owning their own. Falling cost is a chance, not a result.

The chance hasn’t been taken on its own, and the reason is plain once the two stories are put side by side. Where copying is cheap, a margin survives only by law or by control of a bottleneck, a platform everyone has to pass through, a data set no one else has, a model too costly to train twice. The near-zero cost Rifkin celebrates and the toll in the last section are the same story told from two ends. The cheaper the copy, the more a margin depends on a fence. That is why the fight has moved to intellectual property.

AI as capital

A trained AI model is capital in the oldest sense, a produced means of production. Look at what it is made of. There are machines the firm owns and engineers the firm hires. And there is text, code and imagery written by millions of people, mostly gathered without asking. The first two are bought. The third isn’t all a commons. Much of it was published for anyone to read, but some was under copyright and used anyway, without permission and not always within fair use, pirated books among them. This book has no quarrel with the copying, for the reasons in the first section. The quarrel is with what comes after: all of it now sits fenced inside a model someone else owns.

Vercellone’s argument fits a model better than anything he had to hand in 2008. Value now comes from cooperation that happens outside the firm, and the firm captures it from outside. The knowledge of everyone is copied in, then fenced out.

The courts are where the fence shows. The firms argue that copying for training should be free, and so far the courts have given them much of that. In June 2025, in Bartz v. Anthropic, Judge William Alsup held that training on books the company had bought and scanned was fair use, but that downloading and keeping a library of pirated copies was not. Anthropic then settled with the authors for $1.5 billion, about $3,000 a book, approved in July 2026. Two days after Alsup’s ruling, in Kadrey v. Meta, Judge Vince Chhabria ruled for Meta, but narrowly, because the authors hadn’t made the right case about harm to their market. In Britain, in November 2025, Getty Images lost nearly all of its case against Stability AI; the court found that the model’s weights store no copies of Getty’s photographs. The suits the Authors Guild and the New York Times brought against OpenAI in 2023 are still pending. Where a tool competes head-on with what it copied, courts have been less kind. Copying Westlaw’s headnotes, the short summaries its editors write of each legal point in a court opinion, to train a rival legal-research tool was held not to be fair use in Thomson Reuters v. Ross in 2025, and the Third Circuit affirmed that ruling in 2026.

Now look the other way. OpenAI’s terms forbid using its outputs to build competing models, and in January 2025 it said the Chinese firm DeepSeek may have “inappropriately distilled” its models by training on what they produced. Copying in should be open, on this view, and copying out is theft.

Kinsella’s argument cuts here, and it cuts both ways. If ideas can’t be owned, training on everyone’s published work takes nothing from anyone, and the authors’ suits fail on principle. This book accepts that. But then the rule has to run both ways. Training on a model’s outputs, or copying its weights once they are out, takes nothing from anyone either. A firm that claims the first and forbids the second wants a fence it says no one else may have. That is special pleading, and it fails Ingersoll’s test. Whatever right a firm claims to copy in, it has to grant to everyone who would copy out.

The courts may not see it that way, and nothing guarantees they will be fair or even consistent; The Myth of the Rule of Law (Chapter 24) is why this book doesn’t expect them to. The test holds whatever the courts decide.

That doesn’t oblige anyone to publish. A firm can keep its weights to itself, the way a cook keeps a recipe, and sell access at whatever the market will bear. What it can’t do, on this view, is call in the law to stop others copying what got out, or learning from what it sells. Keep the secret, then, but no fence, and the margin lasts only as long as the secret does. The real cost of training is the strong objection, and it gets the same answer as the drug trial. Lead time, contracts and subscriptions can cover a first copy without binding strangers.

Chapter 35 found that Kirzner’s finders-keepers covers the find, not the fence built afterward. A model trained on a commons and then locked is the purest case of a margin that discovery can’t explain.

The deals of 2024 show where the value sat. In March, Microsoft hired most of Inflection AI’s staff, then paid Inflection about $620 million to license its models and about $30 million to waive claims over the hiring, money that went to repay investors, as Chapter 29 told. It wasn’t a one-off. In June, Amazon hired Adept’s founders and much of its team and took a license to its technology, and Adept’s investors were reported to get their money back. In August, Google rehired Character.AI’s founders and licensed its models in a deal reported at about $2.7 billion. Each time the value sat in people and what they knew. The payments went to whoever held the paper on them.

Will the margin hold? Training a frontier model is costly, which favors a few firms, as costly machinery once did. But the price of using a model of a given capability has fallen about tenfold a year, by one estimate, and open-weight models compete. That is Warren’s rule at work. The margin lasts only as long as the fences do.

Whether such systems could ever be more than capital belongs to Part I, The Unfinished Exclusions, and to Part IX. Here the question is only who owns them.

Where knowledge is cheap

Leonard Read’s “I, Pencil” (1958) is a free-market classic. No one person, the pencil says, knows how to make it. The cedar, the graphite, the lacquer, the brass and the rubber come from thousands of people who never meet. Read meant it as a lesson about markets. Turn it a quarter. The know-how that makes a pencil is a commons nobody fenced, and the pencil is cheap because of it.

Elinor Ostrom, whose pastures and fisheries Chapter 32 drew on, carried her design principles to knowledge herself. In Understanding Knowledge as a Commons (2007), with Charlotte Hess, she treated libraries, scientific findings and the internet as shared resources with their own rules. One difference matters most. A knowledge commons can’t be overgrazed, since my use of an idea doesn’t use yours up. Its two dangers are enclosure and neglect.

Yochai Benkler named what grows there “commons-based peer production” in The Wealth of Networks (2006). Linux and Wikipedia are built by people who aren’t paid by the piece, don’t take orders from a boss and don’t set prices for one another, and billions use what they make. The stories of how they cooperate belong to the Interlude. Their economics belongs here. Production on this scale, without wages, prices or managers, is something the usual theory of the firm said couldn’t happen.

Before the internet made sharing cheap, Karl Hess tried it with tools. A former Goldwater speechwriter turned left-libertarian, he spent the early 1970s in Washington’s Adams Morgan neighborhood helping neighbors raise vegetables on rooftops and fish in basements, and wrote up what he learned in Community Technology (1979): small, shared tools that people can understand, repair and run for themselves. What made that hard then was the cost of knowing how. Where knowledge is cheap, it is no longer the hard part.

Open source turned the fence against itself. The GNU General Public License uses copyright to keep code open. Anyone may copy, change and share it, but no one may close what they copied. It is a fence that keeps a commons open, and it fits Ostrom’s principles closely, with clear boundaries in the license, rules the users make and can change, and cheap ways to settle disputes.

The neglect problem is real, and it should be said plainly. In April 2014 a flaw called Heartbleed turned up in OpenSSL, code that secures a large share of the web’s traffic. The project had one full-time developer and took in about $2,000 a year in donations. Nadia Eghbal’s report for the Ford Foundation, Roads and Bridges (2016), showed how much of the digital world rests on such unpaid upkeep. Firms take freely from the commons and give back little. The outsider test runs both ways.

Where data and platforms sit, co-ops have begun to answer. Stocksy United, founded in 2013, is a stock-photo agency owned by its photographers, about 1,850 artist members today. MIDATA, founded in Switzerland in 2015, is a cooperative whose members own the health data they choose to share and decide together which research it serves. Trebor Scholz gave the idea a name in 2016, platform cooperativism. It is Ellerman’s case from Chapters 29 and 33, applied to the newest firms. Those who supply the work, or the data, own the place that sells it.

Chapter 35 ended on who may grant credit and on what terms. Mutual credit answers that the members grant it to each other. Switzerland’s WIR network, started by sixteen members in the Depression year of 1934, has let many thousands of small firms trade on credit they extend to one another, and its bank still runs. The lineage runs from William Greene’s Mutual Banking to Tucker’s attack on the money monopoly. A ledger is knowledge, and keeping one now costs almost nothing.

Run the part’s tests once more on this ground. Can the buyer see the cost? The cost of a copy is near zero and everyone knows it. What they can’t see is what the fence adds. Can they leave? Only where an open alternative exists, which is why the commons matters. Does the firm still win with its fences removed? For a toll on copies, no. Who carries the loss? The copier, the rival, the patient, the people in the data. Who is outside the deal, and were they asked? The millions whose work trained the models. They weren’t.

A knowledge commons is the newest ground floor, Braudel’s and Chapter 32’s, something to fall back on so that a market can be refused. Where nobody is allowed to fence knowledge, it gets cheap, and people build on it: code, encyclopedias, co-ops, credit. This part has asked what remains of capitalism once its privileges are taken away. On this ground the answer is clearest. Take away the fence, and what is left is the work and the people who did it. That is where this part meets Part I. Each of Tucker’s four monopolies, money, land, tariff and patent, closes off a way of working for oneself, through credit, ground, a market or a tool, and leaves renting oneself out as the way that remains. The fence around knowledge is the newest of them, and the rental it props up is the one The Unfinished Exclusions set out to end.

Question	Capitalist answer	The challenge	What gets built
What should a copy cost?	Whatever the rights holder sets, for life plus seventy years	A copy costs almost nothing; the margin is the fence	Lead time, commissions, pledges, subscriptions
Who owns a trained model?	The firm that trained it	It is made mostly of everyone’s work, and the fence runs one way	No fence either way; open weights where people choose
Who owns the data?	The firm that gathered it	The people it describes are outside the sale	Data co-ops
Who keeps the commons up?	No one’s job	Firms take freely and give little back	Copyleft, funded upkeep, platform co-ops
Who grants credit?	Licensed banks	The money monopoly (Chapter 35)	Mutual credit

Part IX turns from what people build to what anarchism itself still has to answer.

Sources

Benjamin R. Tucker, “State Socialism and Anarchism” (written 1886; Liberty, March 10, 1888). The four monopolies.
Richard Stallman, “Did You Say ‘Intellectual Property’? It’s a Seductive Mirage” (2004). gnu.org
Roderick T. Long, “The Libertarian Case Against Intellectual Property Rights” (Formulations, Autumn 1995). C4SS; and “Bye-Bye for IP” (1995 letter, reposted 2010).
Michele Boldrin and David K. Levine, Against Intellectual Monopoly (Cambridge University Press, 2008), ch. 1.
John Kelsey and Bruce Schneier, “The Street Performer Protocol” (USENIX Workshop on Electronic Commerce, 1998).
Jonathan Haskel and Stian Westlake, Capitalism Without Capital (Princeton University Press, 2017).
Carlo Vercellone, “The Crisis of the Law of Value and the Becoming-Rent of Profit,” in Andrea Fumagalli and Sandro Mezzadra, eds., Crisis in the Global Economy (Semiotext(e), 2010).
Jeremy Rifkin, The Zero Marginal Cost Society (Palgrave Macmillan, 2014); Eric S. Raymond, “Zero Marginal Thinking: Jeremy Rifkin Gets It All Wrong” (Armed and Dangerous, April 3, 2014).
Kevin A. Carson, The Homebrew Industrial Revolution (2010).
Bartz v. Anthropic (N.D. Cal., June 23, 2025; settlement approved July 2026); Kadrey v. Meta (N.D. Cal., June 25, 2025); Getty Images v. Stability AI [2025] EWHC 2863 (Ch); Thomson Reuters v. Ross Intelligence (D. Del., Feb. 11, 2025; 3d Cir. 2026); In re OpenAI Copyright Infringement Litigation, MDL 3143 (S.D.N.Y.).
OpenAI on DeepSeek and distillation (Axios, January 29, 2025).
Microsoft–Inflection (March 2024), Amazon–Adept (June 2024), Google–Character.AI (August 2024), as reported by Bloomberg, Semafor and the Wall Street Journal.
Guido Appenzeller, “Welcome to LLMflation” (a16z, November 2024).
Leonard E. Read, “I, Pencil” (The Freeman, December 1958).
Charlotte Hess and Elinor Ostrom, eds., Understanding Knowledge as a Commons (MIT Press, 2007).
Yochai Benkler, The Wealth of Networks (Yale University Press, 2006).
Karl Hess, Community Technology (Harper & Row, 1979); with David Morris, Neighborhood Power (Beacon Press, 1975).
Nadia Eghbal, Roads and Bridges: The Unseen Labor Behind Our Digital Infrastructure (Ford Foundation, 2016).
Trebor Scholz, Platform Cooperativism (Rosa Luxemburg Stiftung, 2016); Stocksy United; MIDATA.
WIR Bank (founded 1934); William B. Greene, Mutual Banking (1850).