# How One Judge's Split Ruling on Anthropic Became AI's Copyright Rulebook > Source: > Published: 2026-08-23 17:39:27+00:00 *Judge William Alsup said training AI on books is fair use. He also said stealing those books from pirate sites is not, and that split cost Anthropic $1.5 billion. Now it's the playbook publishers are using against Google.* Here's the plain answer to a question every AI company's lawyers are now billing hours over: yes, you can train a model on copyrighted books. No, you can't download them from a pirate site to do it. That's the rule a federal court in San Francisco laid down in the case against Anthropic, and TechCrunch's August 23 explainer lays out why it now shapes every AI copyright fight in the country, including a fresh one against Google. Judge Alsup ruled last year that feeding books into an AI model is fair use, comparing it to how a writer studies the authors who came before them. That part of the ruling was a win for Anthropic and, by extension, for every lab building large language models. But Alsup drew a hard line at how Anthropic got the books. The company had downloaded millions of them from shadow libraries, pirate sites that host copyrighted text with no license and no payment to anyone. Alsup found that building a permanent library out of pirated copies wasn't fair use no matter what happened to the books afterward. That distinction turned into real money fast. A federal judge gave final approval to Anthropic's $1.5 billion settlement with the authors in July, after an earlier version of the deal got rejected for being rushed and thin on detail. Authors are set to receive roughly $3,000 per affected work, and Anthropic has to destroy its pirated copies. According to Forbes, the fine lands on a company projecting close to $200 billion in annual revenue by 2028, which tells you something about how these companies are pricing legal risk against growth. Frankly, that's the part every AI startup without Anthropic's balance sheet should sit with. A $1.5 billion settlement is a rounding error for a company chasing $200 billion in revenue. It's existential for almost everyone else building on the same training methods. [Claude AI Suffers Widespread Outage Across All Its Models on August 18](https://startupfortune.com/claude-ai-suffers-widespread-outage-across-all-its-models-on-august-18/) Claude went down across every major model on August 18, with Downdetector logging more than 4,000 user reports within an hour and Anthropic's status page offering no root cause. The outage is the latest in a run of at least seven incidents over the prior six days, hitting Claude Chat, the mobile app and Claude Code. - [claude AI outage August 18 all models down](https://startupfortune.com/claude-ai-suffers-widespread-outage-across-all-its-models-on-august-18/) - [anthropic status page confirms widespread Claude errors today](https://startupfortune.com/claude-ai-suffers-widespread-outage-across-all-its-models-on-august-18/) Publishers didn't wait long to borrow the argument. Hachette, Cengage, and Elsevier, along with novelist Scott Turow and the writers' group S.C.R.I.B.E., sued Google in July, alleging the company trained its Gemini models on books it never had rights to use that way. According to TechCrunch, the lawsuit claims Google used copies of books originally submitted for Google Books, a program that only ever promised snippet-level search, and books uploaded to the Google Play store, then folded them into Gemini's training data without new permission. The suit also alleges Google stripped or altered copyright information to obscure where the material came from. That last allegation matters more than it might look. Removing copyright management information is its own violation under federal law, separate from the fair use question entirely, and it's exactly the kind of conduct that turns a fair-use argument into a piracy argument, the same split that sank Anthropic's defense. An internal Google memo reportedly flagged potential penalties in the range of $10 billion to $100 billion, a spread that shows how unsettled the damages math still is even after Alsup's ruling gave everyone a framework. The Google case is not a rerun of Anthropic's. It's a stress test of the same legal seam: was the data legitimately acquired, or was it lifted from somewhere it was never supposed to leave. Every major lab, OpenAI, Meta, Google, Anthropic, trained its models on some mix of licensed content, scraped web text, and books whose provenance nobody wants examined too closely. Alsup's ruling didn't close that exposure. It just told everyone exactly where the line sits. What it means practically is that AI companies now have a clear incentive to license data upfront rather than gamble on a settlement later, and settlements are getting priced like the cost of doing business rather than an existential threat, at least for the labs that can afford them. For the ones that can't, a shadow-library shortcut that looked like a shortcut for training data now looks like a bet on a $1.5 billion tab, or worse. That's not a hypothetical. It's the exact number a court already wrote down. **Also read:** [Flock Safety's CEO Calls For Compromise As Cameras Draw Fire From Both Sides](https://startupfortune.com/flock-safetys-ceo-calls-for-compromise-as-cameras-draw-fire-from-both-sides/) • [Anthropic Cuts Claude Opus Prices in Half as Enterprises Balk at the Bill](https://startupfortune.com/anthropic-cuts-claude-opus-prices-in-half-as-enterprises-balk-at-the-bill/) • [ShipHero CEO Aaron Rubin Says Claude Mythos Beat His Pentest Firm for $10K](https://startupfortune.com/shiphero-ceo-aaron-rubin-says-claude-mythos-beat-his-pentest-firm-for-10k/)