I have just looked at the judgment in Bartz v Anthropic, a case decided in California, but which is likely to have a broader reach by its example. Three authors—a novelist and two non-fiction writers—accused the AI company Anthropic of “stealing” their books to train a large language model named Claude. What the California court ruled was this: training AI on books may well constitute “transformative” use under American copyright law, and thus may not, in itself, be illegal. However, he refused to dismiss the case outright, allowing a trial to proceed on the question of how the books were obtained and stored. The case has been hailed as a copyright milestone. It possibly is that, but is also a symptom of a deeper decay in legal reasoning and our understanding of what machines, and laws, are for.

Let us first deal with the word “stealing.”

Anthropic did not reprint the authors’ books. It did not sell them under its own name. It did not post them online for free. It fed them into an algorithm that adjusted statistical weights across billions of nodes to predict language. This is not theft. It is not publication. And—despite the hyperventilations of rights-holders and their lobbyists—it is not infringement in any meaningful sense. It is the digital equivalent of reading, something no regime has yet dared outlaw outright. But if the plaintiffs and their backers in the publishing world succeed, reading may soon be an activity subject to rent.

What Anthropic did is no different in substance from what any literate man does when he reads a book and allows it to shape the way he speaks. The model, like the man, absorbs patterns. The man is praised for being learned. The machine is sued for being trained.

The idea that reading a book—even by machine—constitutes infringement is as grotesque as accusing a violinist of “stealing” Brahms by practising his Violin Concerto. But grotesque ideas live long when they serve those who collect fees.

This is not about authors. It is about rent. It is about securing control over every fragment of language—every sentence, every metaphor, every rhythm—and charging for their reuse. These lawsuits are not defensive measures. They are tollbooth operations.

The authors involved are largely irrelevant. Their names are valuable for headlines. Their grievances are sentimental. Their purpose ends with the writ. The real accusation is not that Claude outputs their prose, but that it dares learn from it without first asking permission.

Such a claim only makes sense if one believes that ideas can be owned, that exposure is a form of reproduction, and that memory itself must be licensed. This is not copyright. This is pre-emptive censorship masquerading as law.

In England, our copyright regime is slower, but not saner. The Copyright, Designs and Patents Act 1988 does not recognise America’s fantasy of “transformative use,” though it occasionally blinks at the idea under “fair dealing.” Our permitted exceptions—quotation, criticism, parody, private research—are tightly constrained and often unavailable for commercial AI work.

The British Government, to its brief credit, once proposed an expansion of the data mining exception to allow commercial use. The shrieks were immediate. Publishers, unions, commentators—all the usual parasites—proclaimed cultural collapse. The Government folded. We remain in a regime where having a child read aloud is permitted, but having a machine read silently may incur civil liability.

The California ruling, then, is a bureaucratic recognition of the inevitable. It accepted that training on lawfully acquired texts may be fair use—even when those texts are digitised and destroyed in the process. A small victory for reality. But there is a catch.

The court also ruled that storing seven million “pirated” books—obtained via open-source libraries like LibGen—is actionable. This, the judge held, constitutes willful infringement.

Here, the court is formally consistent but substantively wrong. Anthropic’s “central library” was not a publication. It was not indexed or quoted. It was a temporary structure for internal use. To treat this as infringement is to declare that the acquisition of knowledge, not its application, is the crime.

The ruling is clear: what matters is not what the model produces, but how the training data was sourced. You may learn, but only from approved sources. You may read, but only from publishers who receive tribute. The principle is Orwellian: knowledge, unless licensed, is theft.

A large language model does not memorise books. It does not store them like an archive or reproduce them like a press. It digests. It internalises. It forgets. What remains is not quotation but structure.

The legal theory behind the lawsuit confuses training with reproduction, memory with mimicry. It treats learning as a regulated activity. That may be consistent with current law. But it is inconsistent with any plausible notion of freedom.

To be clear: copyright infringement is not theft. It is not conversion. It involves no dispossession. It is a statutory cause of action—rooted not in property, but in government-granted privilege. And it must be understood as such.

The defenders of the regime invoke the supposed social value of copyright. They speak of incentives. But these are post-hoc rationalisations. Copyright began as censorship. It continues as rent. It is a method of limiting access, not encouraging creativity.

What is needed is not reform. It is abolition. Reducing the copyright term to five years would be an improvement. Eliminating criminal penalties for copying would be progress. But these are palliatives. The disease remains.

If reading must be licensed, we are not in a free society. If learning by machine is a legal risk, we are not in a rational one. The proper response to this crisis is to dismantle the structure of ownership over language.

Let all AI training on publicly available materials be recognised as non-infringing. Let the burden of proof lie with those who claim reproduction. Let the courts, if they must exist, serve truth, not monopoly.

Meanwhile, the corporate custodians of AI—OpenAI, Anthropic, Google—will strike licensing deals. They will negotiate truce with the same publishing houses that once condemned them. They will sanitise their models, tune them for safety, and sell them as certified non-offenders.

This is not liberation. It is cartelisation. It is a new index, administered by digital monks. And it will be praised as ethical.

If these lawsuits succeed—if the courts accept that reading is infringement—we will enter an age of licensed literacy. Every dataset pre-cleared. Every phrase pre-authorised. Every model a mouthpiece for its paymasters.

This is not authors versus AI. It is monopolists versus memory. It is censors versus civilisation.

If we had a real Parliament, it would repeal the Copyright Act and start again. If we had judges with integrity, they would hold that language is not property.

But we have neither.

We have instead a managerial tyranny, policed by failed novelists and run by moral accountants.

They deserve each other. But we deserve better. The future is not theirs to license.