For the second time in a week, a federal district court judge in California has sent the plaintiffs in an AI copyright infringement lawsuit back to the drawing, dismissing most of the charges brought against Facebook-parent Meta over its Llama generative AI tool while granting them leave to amend and refile their complaint.
The ruling partially granting Meta’s motion to dismiss came down on Friday, one week after another federal judge dismissed most of the copyright infringement charges brought against StabilityAI and other AI image generators while granting leave for plaintiffs to amend and refile their complaint.
The latest case was filed in July, by comedian Sarah Silverman and two other authors, against Meta and ChatGPT-developer OpenAI. Meta and OpenAI each filed motions to dismiss; Friday’s ruling came in response to Meta’s petition. The authors alleged Llama (and ChatGPT) were trained on illegally-acquired datasets that included their books, which they claim were acquired from “shadow library” websites like Bibliotik, Library Genesis, Z-Library, and others.
Their complaint also charges that, as Meta’s training data included their expressive works, “every output of the LLaMA language models is an infringing derivative work, made without Plaintiffs’ permission and in violation of their exclusive rights under the Copyright Act,” and accuses Meta of removing copyright-management information.
The latter charges left U.S. District Judge Vince Chhabria nonplussed. “I understand your core theory,” Chhabria told attorneys for the authors. “Your remaining theories of liability I don’t understand even a little bit.”
Even more baffling to the judge, was the plaintiffs’ claim that, “Because the LLaMA language models cannot function without the expressive information extracted from Plaintiffs’ Infringed Works and retained inside the LLaMA language models, these LLaMA language models are themselves infringing derivative works.”
“That makes my head explode when I try to understand that,” Chhabria said.
The preliminary dismissals of most of the charges in lawsuits against Meta and StabilityAI reflect the difficulty authors, artists and rights owners, and their lawyers, are having formulating and articulating a coherent legal argument as to how, when and where they believe generative AI models are infringing their copyrights.
The problem goes deeper than mere legal drafting, however, and it has been evident since the earliest days of the debate over copyright and AI. As I wrote in April, following the first of the U.S. Copyright Office’s four listening session on the issue, we lack the critical, let alone shared vocabulary even to fairly debate what we’re trying to debate.
The same lack of a common vocabulary runs through the public comments filed with the USCO as part of its inquiry.
Those what have made their living or built their business on copyright naturally tend to view technology they believe threatens their business or livelihood as an infringement of copyright. Those business is based on technology development view anything that might impede that development, including pesky copyright laws, as intolerable.
The problem is evident even when viewed through the other end of the telescope from the authors suing Meta. In his lawsuit against the Copyright Office over its refusal to register a work he maintained was “autonomously created” by a generative AI, technology developer Stephen Thaler was reduced to arguing (implausibly and unsuccessfully) that courts and the Copyright Office have been misreading copyright law and the constitution for the past 139 years.
In principle, courts and the law should be able to mediate between the two views. That will continue to be difficult, however, so long as the two sides cling to their reductionist framings of the debate.
It will continue to be difficult for rights owners to describe the training and use of generative AI models by reference to the exclusive rights granted to authors by §106 of the Copyright Act. Generative AI simply stands outside the technological and economic paradigm for which the scope and definitions of those rights were developed, and no amount of lawyering is likely to force it back in.
At the same time, the argument from AI developers that the use their models make of copyrighted training inputs is no different from, or is analogous to what a human does by reading a book or viewing a painting, is at best a mere legal convenience and at worst a bad-faith red herring.
A generative AI model extracts commercial and economic value from the aggregate of its training inputs, irrespective of any expressive content. Right now, that exchange of value is a one-way street. What we lack — and what copyright law, for now, seemingly cannot supply — is a means to put an equitable price on that value as would happen in any other market transaction.
We’re gonna need a new playbook.