Robin Hood, the legendary antihero, is beloved for stealing from the rich and giving to the poor. But what if he stole from the rich and then gave to his own bank account, with the explicit intent of writing checks to the poor at some undefined future date . . . is this an equally redeemable act?
This distinction between immediate transformative use and indefinite storage lies at the heart of a recent federal court decision that may inadvertently encourage AI companies to adopt “pirate and delete” strategies when acquiring training data. In Bartz et al. v. Anthropic PBC, Judge William Alsup ruled that while using copyrighted books to train artificial intelligence constitutes fair use due to its transformative nature, maintaining a permanent library of pirated content does not qualify for such protection. The court’s reasoning creates a troubling framework where the speed of data processing, rather than the legitimacy of acquisition, may determine legal liability — effectively rewarding efficient digital theft over legitimate licensing arrangements.
Anthropic PBC, the company behind the Claude AI assistant (a generative AI platform similar to ChatGPT), has been quite successful — generating over a billion dollars annually from its AI services. To build an AI that can understand and generate human-like text, Anthropic needed to expose its system to enormous volumes of well-written content, with books serving as prime examples of sophisticated language use. The explicit goal, said Anthropic’s cofounder, Ben Mann, was to amass a central library of “all the books in the world” to retain “forever.” In early 2021, Anthropic downloaded Books3, a collection of 196,640 books from questionable sources. Not satisfied with this modest haul, they proceeded to download “at least five million copies” from Library Genesis (LibGen) in June 2021 followed by “at least two million copies” from Pirate Library Mirror (PiLiMi) in July 2022 knowing that the source material was pirated.
The plaintiffs — authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — discovered their copyrighted works had been swept up in Anthropic’s digital dragnet without permission. Their complaint centers on how their books ended up in Claude’s training diet, though notably, they do not allege that Claude ever regurgitated their work to users (thanks to guardrails put in place by Anthropic to prevent such use).
Reading to Learn vs. Reading to Copy
The court stated that Anthropic’s use of works to train its LLMs was akin to training a human to read and write. It argued that authors cannot prevent anyone from using their works for “training or learning as such.” The court likened the process to how people read, re-read, admire, memorize, and internalize books over centuries, drawing upon them to write new things in new ways.
“Everyone reads texts, too, then writes new texts,” the court noted. “They may need to pay for getting their hands on a text in the first instance, but to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways, would be unthinkable.”
The authors also argued that training LLMs would lead to an “explosion of works competing with their works.” The court dismissed this concern by stating that this complaint is “no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works.” The court clarified that the Copyright Act aims to advance original works, not to shield authors from competition. This raises a provocative question: if an AI produces a work that is impactful, meaningful, and widely sought after by readers, is that creation inherently less valuable to society simply because it emerged from silicon rather than synapses? After all, American markets have long operated on the principle that competition drives innovation — and competition from AI-generated content might ultimately benefit authors by pushing them toward more innovative, distinctly human forms of expression that no algorithm can replicate.
The Road to Infringement Is Paved With Good Intentions
With emphatic superlatives bordering on hyperbolic, the court held that Anthropic’s use of both purchased and pirated text was “spectacularly” and “quintessentially” transformative. And then, in the same opinion, expressed with equal conviction that the act of pirating initial copies, especially when lawfully purchased copies were available, is “inherently, irredeemably infringing.” While training AI with these data may be fair use, the creation of a “permanent, general-purpose library” filled with unlicensed works was not. Such a use, the court reasoned, “plainly displaced demand for Authors’ books — copy for copy.”
What is the difference between these two use cases? The court highlighted that Anthropic retained pirated copies even after deciding it would not use them to train LLMs (at all or ever again). This demonstrated that the pirated library served a purpose beyond just training, functioning as a “permanent, general-purpose resource” or “hard resource” for other or future uses. This primary “library building” use was deemed infringing, irrespective of any secondary, transformative use for LLM training that might occur later.
However, not all of Anthropic’s massive library was built illegally. Anthropic also spent millions of dollars to purchase millions of print books, often in used condition, for its “research library.” These acquired books included copies of all the authors’ works at issue in the case. Service providers for Anthropic would strip the bindings from the purchased print books, cut their pages to size, and scan them into digital form. Crucially, the original paper copies were then discarded, meaning that one copy replaced another rather than creating new, additional copies. The result was a PDF copy containing images of the scanned pages with machine-readable text. Walking through the fair use factors, the court held that this portion of their library did qualify for protection — the primary purpose of the use was easy storage and searchability (making it transformative), the works were lawfully acquired, a reasonable amount was used (format conversion required copying the entire text), and the acquisition with no intent to redistribute had a neutral market effect.
What About Lost Licensing Fees?
The authors contended that Anthropic’s action of scanning purchased print books and discarding the originals effectively displaced the purchase of new digital copies that Anthropic would otherwise have made directly from the authors or publishers. They believed that they “might have wished to charge Anthropic more for digital than for print copies” and that they “could have succeeded if Anthropic had been barred from the format change.” Furthermore, they suggested that converting print to digital copies could expose them to the usurpation of the opportunity to sell rightful copies because digital copies might be more readily transmitted or redistributed illicitly than print copies.
The court stated that any “losses [from displaced purchases of new digital copies] did not relate to something the Copyright Act reserves for Authors to exploit.” Furthermore, the Copyright Act does not grant copyright owners the right “to divide markets or a concomitant right to charge different purchasers different prices for the same book, [merely] say to increase or to maximize gain.” Anthropic had lawfully purchased the print copies, which gave it the entitlement to “dispose[]” each copy as it saw fit under the First Sale Doctrine. The format change was further considered “transformative“ because it eased storage and enabled searchability, effectively replacing a purchased print copy with a more convenient digital one for internal use. Crucially, this process “added no new copies, created no new works, or redistributed existing copies” outside the company.
Move Fast and Steal Things
While the court did its best to discourage and denounce piracy, the ruling may inadvertently create a legal loophole. The court made clear that the training process itself was fair use regardless of how the books were originally acquired — the transformative nature of AI training was what mattered, not the provenance of the training materials. The problem for Anthropic was their decision to keep a permanent “central library” of pirated works for potential future use.
This distinction suggests a perverse legal strategy: download pirated works, immediately use them for AI training, then delete the originals. Under this court’s reasoning, such a “pirate and delete” approach might qualify for fair use protection because:
- the training use itself is “spectacularly transformative;”
- no original copies are retained beyond what’s necessary for the immediate training purpose;
- there’s no permanent library being built for general purposes; and
- the AI retains only “compressed” statistical patterns, not reproducible copies.
In essence, Anthropic’s mistake wasn’t using pirated books for training — it was being packrats about it. Their “store everything forever” philosophy turned what might have been defensible fair use into clear-cut infringement.
This creates a morally hazardous precedent where speed becomes a shield against copyright liability. The message seems to be: if you’re going to pirate books for AI training, don’t dawdle or build libraries — get in, train your model, and delete the evidence.
It’s worth noting that this interpretation would still carry significant risks. Courts in other jurisdictions might disagree, the initial downloading would still constitute infringement (even if brief) and willful infringement could still result in enhanced damages. But the legal logic of this decision does seem to create an uncomfortable safe harbor for quick-acting pirates.
