Month: July 2025

Buy, Train, Delete, Repeat: A Generative AI Training Formula Emerges from the Anthropic Decision

We have what appears to be the first major court ruling on using copyrighted works to train generative AI, and the bottom line is: training AI is fair use because generative AI is “spectacularly transformative,” but copyrighted training materials must be purchased, and maintaining a permanent library of the copyrighted works is questionable.

At issue in this case was Anthropic’s training of its AI assistant, Claude. The company used millions of copyrighted works, some legitimately purchased, some downloaded from pirate websites.

Three authors whose books were used by Anthropic sued for copyright infringement, and Anthropic moved for summary judgment that the training is not infringement but instead a protected fair use.

The court’s opinion, by Judge William Alsup of the U.S. District Court for the Northern District of California, is a detailed and logical analysis of copyright fair use as applied to sourcing, digitizing, and storing copyrighted works as well as, of course, using the works to train generative AI. The court necessarily (and helpfully) breaks out and separately analyzes the different acquisitions and “uses” of the books as well as the different stages of copying the books during the whole process.

Anthropic acquired the millions of books it needed for training purpose in two ways: downloads from pirate libraries and legitimate purchases of mostly used books. The physical books were scanned or digitized into the company’s central library, and all the books stored there. Significantly, not all of the books were used for training and the central library was kept permanently as a general-purpose collection of books even after the training was completed.

During the course of the process, the books were copied in several different ways: first, the books were copied from the central library to create working copies for the training set; second, each work was cleaned to remove lower-value or repeating text, resulting in clean copies; third, each cleaned copy was translated into a “tokenized” copy for training; fourth, the tokenized books were repeatedly copied during the training.

As to the “uses” of the books, for its fair use analysis, the court found two. The first way in which Anthropic used the books was to build a huge central library of content. The second is the obvious one, i.e., to train the large language models (LLMs) for its AI assistant.

The court went through the four statutory fair use factors for each use, discussing how each factor applied to the training copies, the purchased copies, and the pirated copies for both uses.

The court’s central finding is that “the purpose and character of using copyrighted works to train LLMs to generate new text [is] quintessentially transformative.” This is because such training is not to “race ahead and replicate or supplant” the copyrighted works but instead to “turn a hard corner and create something different.”

The court also found that the rounds of copying of legitimately purchased books were reasonably necessary for the transformative training use, and the training did not and will not displace the market for copies of the authors’ copyrighted works. Thus, the copies used to train Anthropic’s LLMs are not infringing, and the training is a fair use.

As to the purchased copies Anthropic converted from print to digital, the court reasoned that such a format change did not add any new copies but rather eased storage and searchability without damaging the copyright owners’ interests. The format change also did not relate to a right reserved to authors of copyrighted works under the Copyright Act. Thus, the change from print to digital was a fair use.

But obtaining pirated copies of the books was not a fair use and constituted copyright infringement. The court said “the person who copies the []book from a pirate site has infringed already, full stop.” The court also held that the separate use of copying the pirated works and building a central library of them was not a fair use.

Finally, as to copies made from the central library copies but not used for training and copies retained even when no longer serving as sources for training copies, the court reserved judgment and denied Anthropic’s motion for summary judgment of fair use.

With this decision, the contours of copyright fair use law as applied to training generative AI are taking shape. The big takeaway is that developing training copies of copyrighted works and using them for training are protected fair uses.

However, to enjoy this protection, those training AI tools should adhere to a couple of ground rules: first, obtain the copyrighted works legitimately, i.e., buy the books; don’t copy pirated versions. Second, to the extent copies of the copyrighted works are not used for training or once training is complete, one might need to delete the works rather than maintain them permanently. That is, future trainings might require a new round of purchases.

For the generative AI training formula, then, a useful mantra might be: buy, train, delete, repeat.