Who Pays For AI?

When three authors sued Anthropic in August 2024, nobody expected it to snowball into the largest copyright settlement in American history. But fast forward to July 2026, and a federal court approved a staggering $1.5 billion payout.

Bartz v. Anthropic didn’t just break records, it drew a massive line in the sand for the booming, Wild West era of AI development. It finally gave us a clear answer to a burning question: Can AI companies lawfully train their models on copyrighted books? The court’s answer: Yes, but you can’t steal them first.

How the Fight Started

The beef was straightforward. Authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson discovered Anthropic was training its “Claude” AI using pirated copies of their books scraped from shadow libraries like LibGen.

Anthropic didn’t deny using the books. Instead, they leaned hard on the “fair use” defense. They argued that feeding text into an AI to teach it language is a highly transformative process that doesn’t actually compete with the original books in the marketplace.

The Plot Twist: A Split Decision

The defining moment arrived in June 2025, when Judge William Alsup issued a ruling that effectively split the baby:

• The Win for AI: The judge ruled that training AI on legally acquired books is, in fact, “quintessentially transformative” and protected by fair use.
• The Big “But”: However, downloading millions of books from piracy websites to build that training library is not fair use.

In short, the judge ruled that a transformative end product doesn’t give you a free pass to shoplift your raw materials. This distinction was devastating for Anthropic, exposing them to massive liability for how they sourced their data.

The Historic Settlement

Facing a class-action trial and the terrifying prospect of paying statutory damages on 500,000 books which could have bankrupted the company, Anthropic blinked.

They agreed to a $1.5 billion settlement. Here is what made the deal so unprecedented:

• Massive Payouts: Rights holders netted roughly $3,000 per work, about four times the standard statutory minimum and leaps and bounds beyond typical copyright recoveries.
• Scorched Earth: Anthropic was forced to permanently destroy the pirated datasets.
• A Narrow Truce: The settlement only forgave Anthropic for past data scraping. Authors kept their right to sue over future conduct or AI-generated outputs.

By the time the dust settled and Judge Araceli Martínez-Olguín gave the final sign-off in July 2026, over 91% of eligible works had been claimed. She also drastically slashed the attorneys’ fees from a requested $300 million down to a more reasonable $101.5 million.

Why This Changes Everything

Bartz v. Anthropic is a giant, billion-dollar warning sign for Silicon Valley: data provenance matters.

You cannot launder stolen data through a sophisticated algorithm. While AI companies now have legal backing to train models on copyrighted works, they have to acquire those works legitimately. For creators, the case proves that class-action lawsuits have serious teeth against Big Tech. With the door left wide open for future lawsuits over AI-generated outputs, the copyright wars are far from over but the rules of engagement are finally starting to make sense.