Anthropic’s Record Settlement Marks the Moment AI Companies Lost Their Free Lunch
Anthropic just paid $1.5 billion to settle a copyright case, in a case that might change everything about how AI gets built
Many of you are aware that, in addition to this Substack, where I post (too infrequently), I am also the executive editor for Pete Pachal’s MediaCopilot.ai, which is currently undergoing a relaunch. I also handle the newsfeed for his substack. Clearly everyone should subscribe to it if you’re interested in AI and media.
I don’t often opine on the news from there, but the preliminary approval of Anthropic’s $1.5 billion settlement to authors is a big moment in AI development.
The settlement, approved on Thursday by Judge William Alsup in San Francisco federal court, covers roughly 500,000 books that Anthropic allegedly downloaded from pirate sites, including Library Genesis and Pirate Library Mirror. At $3,000 per book, it’s the largest copyright recovery in U.S. history.
The case started when authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson sued Anthropic for using their books to train Claude without permission. Anthropic had two sets of books: ones they bought and scanned legally, and millions they grabbed from pirate sites.
Alsup’s June ruling split the difference. He called Anthropic’s use of legally purchased books “among the most transformative we will see in our lifetimes” and ruled it fair use. But pirated works? “Inherently, irredeemably infringing,” he wrote. No fair use defense there.
“This settlement sends a powerful message to AI companies and creators alike that taking copyrighted works from these pirate websites is wrong,” said Justin A. Nelson, representing the authors.
Faced with potential damages reaching tens of billions, Anthropic settled. The company agreed to pay at least $1.5 billion, destroy all pirated libraries within 30 days, and provide written certification of the destruction.
The settlement only covers past conduct through August 25, 2025. It doesn’t protect Anthropic from future claims or from lawsuits about Claude generating copyrighted text. Authors can still sue if the chatbot spits out their work.
“Anthropic is hardly a special case when it comes to infringement,” said Maria A. Pallante, president and CEO of The Association of American Publishers. “Every other major AI developer has trained their models on the backs of authors and publishers, and many have sourced those works from the most notorious infringing sites in the world.”
The ripple effects are already visible. During depositions, Anthropic founder Ben Mann testified1 he’d downloaded the same Library Genesis dataset while working at OpenAI in 2019. Meta employees called it “a data set we know to be pirated” in court documents. Both companies face similar lawsuits.
The $3,000 per-work figure sets a new benchmark. It’s four times the minimum statutory damages of $750 and 15 times what Anthropic might have paid if they’d convinced a court they were innocent infringers.
So, for AI companies, the math just changed. Training data isn’t free anymore. They’ll need to license content, verify data sources, and budget for copyright compliance. Enterprise customers will demand stronger indemnification. Proactive licensing suddenly looks cheap compared to litigation.
The settlement doesn’t kill AI development. Alsup’s ruling protects “transformative use” of legally acquired content. But the days of downloading whatever’s available online and calling it research are over.
AI companies now face the same choice the music industry faced after Napster: Pay for content or prepare for lawsuits. Anthropic’s $1.5 billion check just showed everyone what choosing wrong costs.
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal. June 23, 2025), 3, ll. 3–5.