What the machine reasons belongs to no one.

On 20 July 2026, Judge Araceli Martínez-Olguín granted final approval to the US$1.5 billion settlement between Anthropic and the authors of the books the company had downloaded from pirate libraries, the largest copyright recovery on record ^[1][12]. Thirteen months earlier, Judge William Alsup had resolved the other half of the case in the opposite direction: buying printed books, cutting off their covers, scanning them and discarding the paper is fair use, and training a model on them is “exceedingly transformative” ^[2]. On 23 February 2026, in the interval between the two rulings, Anthropic alleged that three Chinese labs had extracted Claude's capabilities through 16 million exchanges across some 24,000 fraudulent accounts ^[3][11].

Read together, the three pieces describe not a contradiction but a single criterion applied twice, and the consequence is the news no one has stated out loud. Neither the court nor the company holds that anyone owns what comes out of a compression: the judge freed the model from the authors' claim, and February's accusation does not assert in a single line that Claude's answers are Anthropic's property, because there is no title to assert. The larger phenomenon behind these facts is that intellectual property has withdrawn from the discussion about intelligence without anyone signing the withdrawal, and what is left being fought over is far smaller: contracts, accounts and permissions.

What the law blessed: buy, cut, scan

In February 2024 Anthropic hired Tom Turvey, who had run the book deals at Google, with a brief the record states without ornament: “all the books in the world” ^[4]. The operation had an internal name, “Project Panama,” and the scanning was carried out by a digitization contractor, Datamation Imaging Services ^[4]. The company bought millions of used copies from Better World Books and World of Books, and its contractors cut off the bindings, trimmed the sheets, scanned every page and threw the paper away ^[4].

Alsup validated that practice on 23 June 2025 along two separate lines ^[2]. The first is the first-sale doctrine: whoever buys a copy may convert it to another format for their own use, and because the original is destroyed no additional copy is created and nothing is distributed. The second is the more consequential, because the judge held that training a model on those books is exceedingly transformative — that is, that what comes out of the process is so far from what went in that the book's author retains no claim over it. The only thing Alsup declared unlawful was the library of more than seven million books Anthropic had downloaded from pirate sites, and the settlement came out of that. The Authors Guild received the ruling as a half victory and objected precisely to the doctrine holding it up, that of transformative training ^[13].

The appetite for paper did not end with the case. Fortune reported on 31 July 2026 that Pieter de Vries, an antiquarian bookseller in Haarlem, took for a scam the email in which a company called 2077AI asked him for more than three thousand titles ^[5]. Precision is worth the trouble here, because the point stands without exaggerating it: Anthropic states that none of its acquisition programs “acquires or destroys rare or antiquarian books” ^[5], and ISBNdb, which in 2024 announced bulk book sales for “training needs,” maintains that the offer was an exploratory concept that never launched ^[5]. What is documented here is an industry's appetite, not a charge against one company.

The settlement that left almost everyone out

The approved settlement pays around US$3,000 per work in installments through September 2027, requires the pirated files to be destroyed, and covers a little more than 480,000 books, of which close to 91% have already been claimed ^[1]. It does not, on the other hand, release any claim over what the models generate ^[1].

The fine print is what matters for anyone reading this from the south. Only works registered in time with the United States Copyright Office enter the class, a formality the Berne Convention makes unnecessary in order to be born protected and which for that reason almost no one completes outside that country; foreign editions without an ISBN or without registration were expressly excluded ^[6]. The result is brutal: the class amounts to about one-fourteenth of the more than seven million downloaded ^[6]. A Venezuelan or Argentine author whose book was in that library is inside the model and off the list of creditors, and not because the right was denied to them but because they never filled out a form their own law does not require of them.

The distillation accusation, read closely

The 23 February 2026 complaint names DeepSeek, Moonshot AI and MiniMax, and describes an industrial operation: some 24,000 fraudulent accounts in coordinated architectures, commercial proxy services to get around the fact that Anthropic offers no commercial access to Claude in China, and more than sixteen million exchanges aimed at training weaker models on the answers of a stronger one ^[3]. The company adds a safety argument that is not decorative: a distilled model inherits the capability without inheriting the safeguards that prevent its use for biological weapons or cyberattacks ^[3].

What is decisive, however, is what the text does not say. Every verb in the announcement points at access — terms-of-service violation, regional restrictions, fraudulent accounts, unauthorized scale — and at no point does it claim that the reasoning traces are Anthropic's property ^[3]. Dario Amodei had shown the same prudence in January 2025, when he wrote about DeepSeek and expressly declared that he was taking no position on the reports of distillation from Western models ^[7]. The adjective holding up the whole accusation is “illicit,” which suggests theft without going so far as to name an owner.

Why they don't say it: the law had already answered

That omission is not rhetorical timidity but calculation, because in the United States copyright protection requires human authorship and a purely automatic output has none. On 2 March 2026 the Supreme Court declined to review the Thaler case, leaving settled the criterion that a work without a human author cannot be registered ^[8][14]. A model's reasoning traces are therefore no one's intellectual property: it is not that Anthropic declines to claim them, it is that there is no title to claim.

What does exist is a contract. Anthropic's commercial terms prohibit accessing the service to build a competing product, including training rival models ^[9], and that clause is enforced in the only way an access clause can be enforced, which is by closing the door: the company revoked OpenAI's access in August 2025 and blocked xAI in January 2026 ^[10]. The xAI case measures how far that force reaches, because the company — American, not Chinese — went on using Claude outputs through personal accounts and through an intermediary until mid-May 2026 ^[10]. The moat, then, is not a title of ownership but a revocable permission, and its strength depends on the other party having accepted it and being identifiable.

The same criterion, applied twice

It is tempting to read the two positions as a contradiction, and the temptation is worth resisting because they are not one. Anthropic does not hold that the reasoning traces belong to it, and its claim is narrow and legitimate: someone pretended to be thousands of different customers in order to evade a block the company has every right to impose. Charging hypocrisy would be arguing with a position no one is defending, and missing, along the way, what actually happened.

What happened is that one and the same criterion was applied on both floors of the building and survived. One floor down, compression transformed the corpus so thoroughly that the book's author retains no claim over what came out; one floor up, what comes out of the model has no owner either, this time because it lacks the human author registration requires. The doctrine is coherent, then, and its consequence is uncomfortable for everyone alike: if the criterion holds in both directions, reasoning itself is no one's property, neither the person who wrote the books nor the one who paid for the training. What can be owned is far less glamorous, because it is the conditions of access, and a condition of access binds only whoever accepted it.

Those are the facts. What they say about what can be owned is the business of the section that follows.

The phrase circling this discussion is that intelligence tends to be free, and it is worth taking seriously exactly as far as the evidence carries it. “Too cheap to meter” was the promise Lewis Strauss made in 1954 about atomic electricity, and it failed, because the cost never lay in the fuel but in the plant that had to be built and wired to every consumer. The analogy holds that far and then breaks, and the break is the good news: electricity required every user to hang off the plant, whereas a model, once trained, travels. Building the frontier costs hundreds of billions and using it costs less every year, so capability has stopped requiring ownership of the infrastructure that produced it ^[15].

From there follows what the region would do well to stop doing. Much of the Latin American conversation about artificial intelligence is organized around access — a seat at the table, a quota of chips, a sovereign model — and that is precisely the constraint that has been loosening on its own. What Anthropic denounces is not that someone learned from a model but that fake accounts evaded a commercial block ^[3], and downloading weights published under a license that authorizes it looks nothing like that: two of the three accused labs, Moonshot and MiniMax, are exactly the ones publishing those weights. Spending political capital to obtain that permission amounts to fighting for an authorization that every release of weights makes less necessary.

The counterpoint deserves to be granted without irony, because the safety half is not a pretext and the company's claim is narrow and reasonable. A distilled model inherits the capability without inheriting the safeguards, sixteen million exchanges through twenty-four thousand fake accounts are fraud by any standard, and anyone has the right to choose whom they contract with ^[3]. It is also worth not overstating the outcome, since the door still exists for the heaviest loads and the iron is still rationed; what is asserted here is a direction and not an arrival, because the access constraint gives way year by year and the constraint of one's own capability does not.

That leaves the question of what actually is scarce, which is where the discussion ought to be. Every input that becomes free moves the value to the complements that remain expensive, and in this story Latin America stands exactly where those complements are, the ones no one can supply it from outside: the data of its own reality, the institutions capable of buying and deploying, and the trust to distribute in its markets. The settlement demonstrates it better than any argument, because the region ended up off the list of creditors not through dispossession but for not having filled out a form ^[6], and that figure — present as material, absent as a party — repeats at every table where there is no registered corpus, no evaluation of its own and no public buyer able to sign.

THE PLAY

1. Stop paying for the permission and use what is already published. The budget and the political capital that today go into negotiating access to the frontier yield more when they fund deployment on published weights, whose license grants what no API terms of service will ever grant. The useful question is not which model we are allowed into, but which problem of ours is solved by the one we can already download.
2. Register the corpus, because it is the complement whose price does not fall. The mechanism that left the region out of the largest copyright settlement in history is bureaucratic and therefore repairable: a collective registration program with publishers, ministries and university presses turns an invisible corpus into a claimable one before the next settlement, and into a licensable asset for the day reasoning is worth nothing at all.
3. Legislate public procurement, not authorship. Whoever drafts artificial intelligence rules in the region has a surplus of debate about who owns what a machine writes and a shortage of the one that decides the region's industrial future: how the state buys, evaluates and deploys on models it can host itself. A ministry that knows how to procure intelligence is worth more than any paragraph on the ownership of a token.