AI and Books: Anthropic Wins on Fair Use but Takes a Big Risk for Piracy | Fine-tune llm on your own data | How to train llm on your own data reddit | | Turtles AI

Anthropic Gets Green Light on Purchased Book Training (Fair Use), but Will Face Separate Trial for Use of Over 7M Pirated Copies. Judge Alsup Sets Significant Precedent.

Key Points:

• Use of purchased and destroyed books for digitization is fair use. 
• Claude’s training deemed “quintessentially transformative” by courts. 
• Large-scale piracy (Books3, LibGen, PiLiMi) not justified and subject to prosecution. 
• Liability Implications: Even if Anthropic later buys pirated copies, it is not immune from harm.

On June 24, 2025, Judge William Alsup of the Northern California District Court defined the method adopted by Anthropic to train Claude – the purchase, dismemberment and definitive digitization of millions of books – a “fair use” under the US fair use, as it is transformative and not aimed at reproducing or redistributing the original works in their entirety; the format conversion was considered an operation of ease and research, without harming the market of authors.

Alsup also ruled that making the purchased books a multi-format “central library” was legitimate, but he set a clear limit: the use of pirated material – over 7 million works obtained from Books3, LibGen and PiLiMi – does not fall within the protection of fair use. For these pirated copies, kept “for convenience and wallet” even after the training began, a trial in December will be necessary to assess any damages, potentially up to $150,000 per work.

The court also highlighted how Anthropic had initially attempted a safer route – through TomTurvey, former head of GoogleBooks – by negotiating licenses, but then opted for a cheaper and riskier approach, buying used copies and downloading pirated content.

This decision represents an important turning point in the jurisprudence on copyright and AI: the first American ruling to define training on purchased texts as protected by fair use, but it also establishes a clear boundary against the use of illegal sources. This is the first ruling to explicitly characterize transformative use for AI purposes and initiates a process aimed at quantifying responsibilities for pirated material.

The verdict opens up concrete scenarios for other companies – including OpenAI, Meta and Microsoft – that use large textual corpora for their models, offering a legal path (purchase + digitization + destruction) and highlighting the risks associated with indiscriminate piracy.

The Bartz, Graeber and Johnson vs. Anthropic case therefore marks a turning point in the balance between scientific progress, creativity and copyright protection, and only the piracy issue remains pending with the trial already scheduled.