Judge Swipes Left on Silverman's Suit

OpenAI and the "Nonsensical" Copyright Arguments

Not everyone is feeling thankful this Thanksgiving in the United States. The comedian, actress and author Sarah Silverman and three other plaintiffs just had most of their lawsuit against OpenAI and others dismissed. Following through on his previous in court comments, District Court Judge Vince Chhabria of the Northern District of California granted Defendant Meta’s motion to dismiss. You can download the short dismissal order here.

The court dismissed all Plaintiffs’ claims except “the one alleging that the unauthorized copying of the plaintiffs’ books for purposes of training LLaMA constitutes copyright infringement.” Id. at 1. The LLMs at issue, referred to as LLaMA have been released as open source to the public despite costing millions to train and produce. In fact, for those of you interested, you can download a free app to your computer called LMStudio and experiment with LLAMA and other open source models along with paid ones that OpenAI offers. Facebook must be thinking, no good deed goes unpunished. But, at least for now, the district court judge has granted them a huge win.

It is indeed unusual to see written in a judge’s order that a now dismissed claim by a party is “nonsensical.” But, that’s what happened here. The court reiterated that the plaintiffs were alleging, in part, that the LLaMA language models are derivative and infringing works because they “cannot function without” taking information from the plaintiffs’ copyrighted works. Id. But the court held that there is “no way” to understand the LLM at issue as somehow being a “recasting or adaptation” of plaintiffs’ works. Id.

The second theory the court dispatched was the notion users of tools like ChatGPT and LLaMA are themselves infringing Plaintiffs’ works. By extension, Meta and others creating the LLMs enabling this infringement are vicariously liable. Sort of like Sony creating the betamax machine enabling rampant movie copying back in the day, but claiming, “hey we are not responsible for people using our machine to infringe copyright.” It seems here the court’s primary focus was that the Plaintiffs offered no evidence that the output of the models at issue actually contained infringing content.

The Plaintiffs claimed they were not legally required to show similarity between the model output for a given user’s query and their copyrighted works. The court rejected that citing case law for the proposition that any alleged infringement must contain some similarity to the copyrighted work. The court held that the Plaintiffs “would need to prove that the outputs [produced by the model in response to user queries] (or portions of the outputs) are similar enough to the plaintiffs’ books to be infringing derivative works.” Id. at 3.

DMCA Not Helpful

Plaintiffs’ also claimed that the Digital Millennium Copyright Act §1202(b) applied because the models were in essence distributing Plaintiff’s books. The court rejected that as well because the Plaintiff’s failed to present facts supporting this claim. A related DMCA claim was also rejected because Plaintiffs could not even “plausibly allege” that the LLaMA was an infringing derivative work. Id.

Several state claims were dismissed by preemption. The court held that those claims wade into territory exclusive to the Copyright Act and by so doing could only be maintained if not dismissible under the Copyright Act - which the court reasoned they were.

The court gave the Plaintiffs’ leave to amend their complaint except for one of their claims, negligence, which the court dismissed with prejudice.

The Other Cases

As we all now, despite other jurisdictions not viewing this decision as controlling, it will be persuasive. Most other courts will want to align with what these decisions are doing. The only hope for Plaintiffs in all of these cases at this point is that some well reasoned decisions arrive that are contrary to this one. In that way, there is the potential for some disputes in Federal Circuit Courts which the U.S. Supreme Court would likely entertain. The problem for Defendants, however, will be enduring some potential losses along the way to those Circuit disputing decisions. My expectation is that even where some District Courts differ from this decision or others around the edges of these issues, the Appellate Courts are likely to coalesce around a determination of no infringement. From a practical standpoint, by the time any one decision is considered at the Supreme Court, the number of open source (as in freely available and usable) models that will be in use will make it a practical impossibility to shut it all down. As of this writing, the number of open source LLMs exceeds 100. There is even a leaderboard in the style of sports leaderboards evaluating those models. The pace of model development, refinement and ease of use is too fast for the law to catch up to this one - and remember this is just the United States. Individuals and small companies worldwide are training and testing models at a geometric rate at this point.

The End Game

LLMs are merely next-word-predicting machines. They have no intelligence per se. They have no reasoning ability (yet). While being trained on a mix of public domain and copyrighted material, courts are aligning in their interpretations that their output is not a creative work in the sense the Copyright Act was intended to regulate. Instead, the output of LLMs is the text version of the answer to a complex, but well understood, mathematical problem involving vectors. We have previous posts explaining in non-technical speak about what vectors are in the context of LLMs and their operation.

The option of finding LLMs are infringing copyright would effectively end their creation and use. Regardless of the sympathy that many feel for the work of artists and how the vectors LLMs created from those works helped build those models, suspending LLM development permanently seems an unlikely move for any court.