Judge tosses most claims in Sarah Silverman's AI lawsuit against Meta

If an AI model trains on your book, does the entire software become an illegal copy of your work? A federal judge just weighed in, and the answer is a firm no.

U.S. District Judge Vince Chhabria dismissed the vast majority of author and comedian Sarah Silverman’s copyright lawsuit against Meta over the use of her books to train its LLaMA language model. It marks the second major ruling where a court pushed back against creators targeting generative AI companies.

Calling out a "nonsensical" argument

Silverman and her fellow plaintiffs had put forward a sweeping legal theory: because Meta copied their books without permission to train LLaMA, the AI model itself is an infringing derivative work, and every single response generated by Meta's tools constitutes infringement.

Judge Chhabria pulled no punches, calling that reasoning "nonsensical" in his order. He wrote that there is simply no way to view the LLaMA models themselves as a recasting, transformation, or adaptation of the authors' books.

He also tossed out the claim that every AI output infringes on their work, noting the plaintiffs failed to show that any actual output copied protected elements from the books. To claim derivative infringement, plaintiffs must prove the outputs incorporate portions of their work under standard copyright similarity rules. The judge gave Silverman's legal team a chance to amend and replead these claims.

The core battle remains alive

Meta did not emerge with a clean slate, however. Importantly, the company did not even move to dismiss the foundational claim in the case: that copying copyrighted books to build a training dataset constitutes direct copyright infringement.

AI developers argue that training on publicly accessible material qualifies as legal fair use. Creators contend tech giants are building commercial empires off unpaid human labor.

This mirrors a parallel lawsuit against Stability AI and Midjourney handled by the same legal team, where Judge William Orrick similarly questioned whether artists can claim infringement without producing identical outputs.

Why it matters

This ruling takes the nuclear option off the table for generative AI companies, at least for now.

If the court had accepted the theory that models and all downstream answers are inherently infringing derivative works, open-source models like LLaMA and every developer building on them would face catastrophic liability. Judges are signaling that traditional copyright standards still apply: without proof of substantial similarity in the actual output, you cannot claim the AI copied you.

For creators and writers, this is a harsh reality check. Because modern large language models synthesize patterns rather than spitting back verbatim passages, proving infringement under the substantial similarity test will be an uphill battle.

We think the court took a legally disciplined approach by rejecting the plaintiffs' overreach. However, the real showdown is still ahead. The survival of the core training-data copying claim means courts will eventually have to rule on whether web-scale training is fair use, and that decision will ultimately reshape the economics of AI.