Fair Use for AI Training: A Historic Precedent That Provides Certainty (and Defies Expectations)

View on LinkedIn

Fair Use for AI Training: A Historic Precedent That Provides Certainty (and Defies Expectations)

Fair use for training: the clarity the industry needed, and an obvious disappointment for those expecting multimillion-dollar settlements.

Without any negative caveats, this is a historic precedent that will set an international benchmark.

The world had been waiting for this judicial order to establish the ground rules for the AI industry. While Anthropic celebrates a victory on behalf of developers and technologists, it does not walk away unscathed. And that too is a positive outcome: even if it ends up costing millions, the message is clear. The ends do not justify the means. Copyright still serves its purpose.

On June 23, 2025, Federal Judge William Alsup of the Northern District of California issued a ruling in Bartz v. Anthropic PBC (No. C 24-05417 WHA). He held that training language models on copyright-protected works constitutes fair use under Section 107 of the Copyright Act, provided those works were lawfully obtained. He cautioned, however: building a general-purpose library of pirated books for permanent use is neither fair nor legal.

The key question: how did you obtain the work you used to train your AI?

I welcome this ruling as marking the beginning of the end of a vacuum that enabled byzantine debates. Many of those who have historically been aggressive in collecting royalties — having consolidated positions of power over decades — are now finding that courts no longer automatically rule in their favor.

This outcome allows AI development to move forward without obstruction, but within clear limits. That is worth celebrating.

What is permitted:

Train AI on any material lawfully obtained Digitize physical books purchased for internal libraries "Memorize" complete content during training

• Train AI on any material lawfully obtained
• Digitize physical books purchased for internal libraries
• "Memorize" complete content during training

What must not be done:

Download millions of books from pirate sites such as LibGen, Books3, or PiLiMi Use piracy as a shortcut to avoid licensing costs

• Download millions of books from pirate sites such as LibGen, Books3, or PiLiMi
• Use piracy as a shortcut to avoid licensing costs

The legal analysis:

Judge Alsup made clear that the use was "transformative, spectacularly so," comparable to how a person reads in order to learn to write.

📘 Case facts:

In 2021 and 2022, Anthropic downloaded more than 7 million pirated books from sites such as LibGen and PiLiMi to build an internal "general library." In parallel, the company purchased millions of physical books, destroyed them, and scanned them to create digital copies for training purposes. Among the affected works are books by Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — authors who sued Anthropic for copyright infringement. The trial formally began in August 2024, and the fair use ruling was issued on June 23, 2025.

• In 2021 and 2022, Anthropic downloaded more than 7 million pirated books from sites such as LibGen and PiLiMi to build an internal "general library."
• In parallel, the company purchased millions of physical books, destroyed them, and scanned them to create digital copies for training purposes.
• Among the affected works are books by Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — authors who sued Anthropic for copyright infringement.
• The trial formally began in August 2024, and the fair use ruling was issued on June 23, 2025.

📘 Outcome:

Training on works lawfully acquired or digitized from owned physical copies does constitute fair use. Maintaining pirated copies as a permanent library does not constitute fair use, even if those copies are not used for training.

• Training on works lawfully acquired or digitized from owned physical copies does constitute fair use.
• Maintaining pirated copies as a permanent library does not constitute fair use, even if those copies are not used for training.

The judge was categorical: "There is no exception in the Copyright Act for artificial intelligence companies."

With statutory damages that could reach $150,000 per willfully infringed work, the total cost of Anthropic's "shortcuts" could run into hundreds of millions of dollars. Even so, that is a lesser price than a ruling that would have declared the training of AI models on copyright-protected material — even when lawfully acquired — to be illegal.