A Judge Gets an AI Crash Course in Anthropic’s Copyright Battle

Judge William H. Alsup turned his San Francisco courtroom into a lecture hall for two hours Thursday. The lesson: a crash course on how the large language models powering popular AI chatbots work.

Attorneys for the AI startup Anthropic PBC and a group of novelists suing the company for copying their books without permission lectured the judge on the key technology involved in building and training Anthropic’s chatbot Claude.

The parties, who got equal time to present power point slides, largely found agreement on the basic facts underlying the case: an AI company downloads large swaths of text which then are converted into numerical “tokens,” the AI model begins to recognize statistical relationships between words, and the final chatbot product predicts the best response to a user’s query.

Alsup has built a reputation for diving into the science behind the many Silicon Valley cases he hears at the US District Court for the Northern District of California, employing educational tutorials where attorneys and experts can speak freely without worrying that what they say can be used against them later in the case.

The judge famously taught himself the programming language Java to better understand a copyright battle over computer code between Google LLC and Oracle. He’s employed similar educational tutorials in a lawsuit against oil producers over climate change, and a patent infringement case involving self-driving car technology.

“I know a fair amount about computers, I know a fair amount from having all these cases over the years,” Alsup said. “The actual idea of AI and how it works and what it’s capable of doing, I’m not up to speed.”

Anthropic, backed by investments from Amazon.com Inc., was hit with the proposed class action from authors last summer. The suit contends that the AI company downloaded hundreds of thousands of pirated ebooks in violation of copyright law in order to train Claude.

The case is one of more than two dozen copyright cases from authors, media organizations, musicians, and artists against major AI companies including OpenAI Inc., Meta Platforms Inc. and Stability AI Ltd. The AI firms are largely arguing that their use of copyrighted material falls under the fair use exception to copyright infringement.

‘How does it predict plot?’

At times, the judge appeared impressed by Claude’s cogent responses. During the authors’ presentation, Justin Nelson of Susman Godfrey LLP showed the judge Claude’s three-paragraph response to a question about why books are particularly valuable for training AI models.

The model said that full-length books, unlike blog posts or Wikipedia articles, are highly edited and curated, and provide greater storytelling context for the AI model to train on.

“Claude produced that on it’s own?” Alsup said. “That’s pretty cool.”

Alsup said he understood the idea that an AI model is trained to predict the most likely next word in a sentence. For example, the most likely last word in the phrase: “One small step for man, one giant leap for...” is “mankind,” whereas “penguin” is less likely.

But the judge said he didn’t understand how a chatbot could craft a larger piece of text like a novel only using prediction.

“How does it predict plot?” he asked while examining Claude’s response to the prompt: write a mystery novel set in the Sierra Nevada mountain range of California. “A good writer would have tension. In one chapter the characters are fighting over something or disagreeing over something.”

“This is pretty good from what I can tell,” Alsup said of the first few paragraphs of the AI-generated mystery novel.

Nelson said the model is capable of creating tension and plots because it has seen hundreds of thousands of books that contain similar plot elements. “It tries to create tension, it knows that that is what mystery novels are supposed to do,” he said.

“You’re not convincing me that it’s not understanding” because plot structure seems to be on a “higher level,” Alsup said. But “maybe they do have the whole universe broken out into many different plots, and they follow one path through the woods.”

During Anthropic’s presentation, Douglas Winthrop of Arnold & Porter LLP said that one area where the parties do appear to disagree is in the certainty with which AI models provide answers.

He said models have a range of possible answers to a query, and there is not one correct answer, pointing to an example of Claude generating two different responses to the same prompt asking for haiku. Nelson’s presentation gave the impression that AI models are “deterministic,” which isn’t right, Winthrop said.

Lieff Cabraser Heimann & Bernstein LLP and Cowan Debaets Abrahams & Sheppard LLP also represent the authors. Latham & Watkins LLP also represents Anthropic.

The case is Bartz v. Anthropic PBC, N.D. Cal., No. 3:24-cv-05417, 1/30/25.