Nov 22, 2023 15:08:00

The claims of the authors who sued Meta for infringing the copyright of AI 'LLaMa' are almost dismissed

Three authors, including well-known comedian and actress Sarah Silverman, sued Meta for copyright infringement, and the plaintiffs claim that using their copyrighted works to train AI constitutes copyright infringement. All claims were dismissed except for.

Court Grants Meta's Motion to Dismiss Claims in AI Lawsuit

https://www.thefashionlaw.com/court-grants-metas-motion-to-dismiss-bulk-of-claims-in-authors-ai-lawsuit/

Sarah Silverman Hits Stumbling Block in AI Lawsuit Against Meta – The Hollywood Reporter
https://www.hollywoodreporter.com/business/business-news/sarah-silverman-lawsuit-ai-meta-1235669403/

LLM litigation · Joseph Saveri Law Firm & Matthew Butterick
https://llmlitigation.com/

Authors Richard Kadley, Sarah Silverman, and Christopher Golden say that it is a problem that Meta's large-scale language model LLaMA was trained based on copyrighted works, and that the products of such models are sued Meta alleging copyright infringement. The plaintiff also sued OpenAI on the same day with the same content.

In response to these complaints, Meta requested that all claims be dismissed, except for the claim that ``unauthorized reproduction of books for the purpose of LLaMA training constitutes copyright infringement.'' On November 20, 2023, Meta's request was granted and several plaintiffs' claims were dismissed.

◆1: LLaMA itself is a derivative work that infringes copyright
A derivative work is a work created by reconstructing, transforming, or adapting one or more existing works (original works).

Regarding LLaMA, the plaintiff points out that the function that allows the output content to be adjusted to approximate the textual representation of the dataset cannot function without the representation of the copyrighted work in the first place. It is an imitation of the expression, and the model itself should be considered a derivative work that infringes copyright.''

However, U.S. District Judge Vince Chhabria, who was in charge of this lawsuit, completely denied this, stating that LLaMA itself cannot be understood as a reconstruction or adaptation of the plaintiff's book. Judge Chhabria dismissed it as ``nonsense'', saying, ``It's like putting LLaMA and Mr. Silverman's book side by side and claiming that they're similar.''

◆2: All output contents of LLaMA are derivative works that infringe on copyright.
The plaintiff claims that all outputs produced by LLaMA, which was trained on the copyrighted work, by altering the expressions of the copyrighted work are derivative works and infringe on copyright. Furthermore, the complaint alleges that third-party users are causing collateral copyright infringement by using LLaMA.

However, in order to sue for copyright infringement, it is necessary to show similarities between the original work and the derivative work.

Regarding this point, the plaintiff stated, ``Since the model was a complete copy of the original work during training, there is no need to show similarities,'' but this argument was deemed to be false, and ultimately It was decided that similarity must be proven.

◆3: LLaMA hides author information

Section 1202 of the Digital Millennium Copyright Act (DMCA) provides that ``no person shall knowingly, with intent to induce, enable, facilitate, or conceal an infringement, provide false copyright management information; shall not be distributed or imported for distribution.' In other words, it is illegal to distribute a book by falsely identifying the author.

The plaintiff alleged that LLaMA distributed the copyrighted work with the author's information deleted, which violated the above provisions, but the plaintiff claimed that LLaMA distributed the plaintiff's copyrighted work without the author's information. This claim was dismissed, as there was no evidence to support the claim, let alone the fact that the plaintiff had distributed the copyrighted work in the first place. In order for the plaintiff to pursue this claim, it must first prove the claim in the previous paragraph that ``all content output from LLaMA is a derivative work that infringes copyright.''

◆4: Meta violates the Unfair Competition Prevention Act
The plaintiff claims that Meta's use of the copyrighted material for LLaMA training without the plaintiff's permission violates the California Business and Professions Act. In this case, based on

the doctrine of preemption, which states that a higher legal authority (in this case federal law) takes precedence over a lower legal authority (California state law), copyright infringement must first be proven. It was not approved.

◆5: Meta is gaining unfair advantage
According to the complaint, Meta used a so-called pirated site to train LLaMA, causing monetary damages to the plaintiffs. This case was also dismissed on the basis of preemption.

◆6: There is a mistake in Meta's information management
The plaintiffs allege that Meta's use of the copyrighted material for training and failure to retain the copyrighted material used for training is illegal under California Civil Code Section 1714, but as above, based on preemption, It was rejected.

A federal district court granted plaintiffs permission to amend their complaint and dismissed all of the claims. Although the plaintiff needs to review the evidence, Meta does not seek to dismiss the central point of the plaintiff's claim, which is that 'unauthorized copying of books intended for LLaMA training constitutes copyright infringement.' Therefore, it is expected that this point will continue to be a point of contention in future court cases.