Authors stumble in claim of Meta AI infringement

Authors who claim Meta Platforms’ artificial intelligence program infringed on their copyrighted works stumbled over an obstacle Thursday that is also plaguing plaintiffs in similar cases: showing an example that a company’s large language model is producing outputs that are substantially similar to their protected works.

The authors are represented by the Joseph Saveri Law Firm LLP and sole practitioner Matthew Butterick in Los Angeles. They also represent clients in three other AI-related cases regarding GitHub’s Co-pilot, ChatGPT and Stable Diffusion. Kadrey et al. v. Meta Platforms Inc., 23-cv-03417, (N.D. Cal., filed July 7, 2023).

Their artist clients suffered a loss against Stability AI two weeks ago when almost all of their claims were dismissed with leave to amend because their complaint was defective in numerous respects and plaintiffs’ attorneys had admitted none of Stable Diffusion’s outputs will likely be a close match for any of the artists’ copyrighted pieces. Andersen et al. v. Stability AI Ltd. et al., 23-cv-00201, (N.D. Cal., filed Jan. 13, 2023).

There have been at least 10 lawsuits filed nationwide against generative AI and this case is at least the third in the Northern District Court of California to have a hearing for a motion to dismiss.

Much of the hearing involved U.S. District Judge Vince Chhabria asking Joseph R. Saveri to clarify exactly what plaintiffs were alleging. The judge said he understood the authors’ counsel’s core theory that Meta copied the writers’ books to train the LLaMA program, but was struggling to grasp their vicarious liability claims.

Meta did not contest the plaintiffs’ core theory, so Chhabria allowed discovery for that claim to proceed and said he would issue an order shortly dismissing the authors’ other claims with leave to amend.

Bobby Ghajar of Cooley LLP spoke on behalf of Meta. He argued plaintiffs should not have leave to amend because they have not even begun to allege Meta’s LLaMA produced an output that was substantially similar to their protected works.

“They avoided alleging that because they can’t,” Ghajar contested. “I don’t believe that they should be given leave to amend to plead that. If they could have, they would have.”

Chhabria took issue with the complaint because he interpreted plaintiffs’ counsel’s argument to be that any time anyone puts a query into LLaMA and gets a result, that is copyright infringement, which was too general a claim in his eyes. The judge said the plaintiffs need to put forward a narrower argument that involves claims based on outputs that are substantially similar to the authors’ works.

“I am abandoning this [argument] as it’s alleged right now,” Saveri said, admitting that the claim was not specific enough. He pledged that he and his co-counsel would amend the complaint to address that problem.

However, Saveri said, “We think that under 9th Circuit law, that substantial similarity is not required. If the offending work is a direct copy, we do not have to show substantial similarity.”

Chhabria disagreed, saying, “If you come back and you allege that an output that is not substantially similar to the protected work is copyright infringement, that’s also going to be dismissed.” The judge continued that an output by definition is not a direct copy and is at best derivative.

The attorney also said that plaintiffs have an alternative theory for direct copyright infringement. He claimed that the process by which LLaMA takes copyrighted work and converts it into an output is also infringement.

Ghajar responded that a language model being able to tell a user about an author’s book or style is not infringement, but something similar to an article or a review of the writer. He also echoed Meta’s argument in filings that AI training is similar to a child learning language by hearing everyday speech or bedtime stories and copyright law does not protect linguistic information that may have been extracted from plaintiffs’ books during training.