Two recent orders in AI copyright cases in the United States District Court for the Northern District of California offer early guidance to the following two questions: 

1. Can the use of copyrighted works in training an AI model lead to infringement?
2. Can the output of an AI model infringe?

On October 30, District Judge William Orrick released an order dismissing two of the plaintiffs’ three claims in Anderson v. Stability AI . On November 20, District Judge Vince Chhabria released an order in a Kadrey v. Meta Platforms dismissing the plaintiffs’ claims with leave to amend. Together, these judicial decisions offer early insights into how the law may view AI models and datasets accused of incorporating copyrighted content.

Can the use of copyrighted works in training an AI model lead to infringement?

The order in Kadrey sharply rejects this idea. Judge Chhabria wrote: 

The plaintiffs allege that the “LLaMA language models are themselves infringing derivative works” because the “models cannot function without the expressive information extracted” from the plaintiffs’ books. This is nonsensical. A derivative work is “a work based upon one or more preexisting works” in any “form in which a work may be recast, transformed, or  adapted.” 17 U.S.C. § 101. There is no way to understand the LLaMA models themselves as a recasting or adaptation of any of the plaintiffs’ books. (Kadrey Order, p. 1)

It is clear the court takes issue with the idea that an AI model is similar in “form” to a book. This type of analysis throws cold water on any infringement theories that AI models themselves will infringe copyright. 

Judge Orrick in Anderson took a less critical approach and, instead, asked the plaintiffs for more facts to support their theories that a trained diffusion model (a type of AI model) could infringe because it either 1) directly infringes through the distribution of an AI model containing “compressed copies of the training images,” or 2) the model itself is an infringing derivative work.  Judge Orrick expressed concern over this theory and wrote:

Plaintiffs will be required to amend to clarify their theory with respect to compressed copies of Training Images and to state facts in support of how Stable Diffusion – a program that is open source, at least in part – operates with respect to the Training Images. If plaintiffs contend Stable Diffusion contains “compressed copies” of the Training Images, they need to define “compressed copies” and explain plausible facts in support. And if plaintiffs’ compressed copies theory is based on a contention that Stable Diffusion contains mathematical or statistical methods that can be carried out through algorithms or instructions in order to reconstruct the Training Images in whole or in part to create the new Output Images, they need to clarify that and provide plausible facts in support … if plaintiffs can plausibly plead that defendants’ AI products allow users to create new works by expressly referencing Anderson’s works by name, the inferences about how and how much of Anderson’s protected content remains in Stable Diffusion or is used by the AI end-products might be stronger. (Andersen Order, pp. 9, 10)

It appears from the order in Anderson that plaintiffs will need to plead facts showing that an AI model recreates the protected training images at some point in its operation for there to be a valid claim of copyright infringement based on the model. 

Based on these two orders, it appears unlikely that a model trained on copyrighted images can itself be the basis for a copyright infringement claim. The operation of AI models tends to involve referencing abstract representations of training data and not the full replication of these data; accordingly, a valid claim that comports with these two orders will be hard to plead. 

Can the output of an AI model infringe?

Both orders took this question head-on and had the same conclusion: the plaintiff must show some similarity between the output of the AI model and the copyrighted work. The Kadrey order took aim at the theory that the output of an AI model could infringe without having to show similarities with a copyrighted work and cited Andersen in the process:

The plaintiffs are wrong to say that, because their books were duplicated in full as part of the LLaMA training process, they do not need to allege any similarity between LLaMA outputs and their books to maintain a claim based on derivative infringement. To prevail on a theory that LLaMA’s outputs constitute derivative infringement, the plaintiffs would indeed need to allege and ultimately prove that the outputs “incorporate in some form a portion of” the plaintiffs’ books. Litchfield v. Spielberg, 736 F.2d 1352, 1357 (9th Cir. 1984); see also Andersen v. Stability AI Ltd., No. 23-CV-00201-WHO, 2023 WL 7132064, at *7-8 (N.D. Cal. Oct. 30, 2023) (“[T]he alleged infringer’s derivative work must still bear some similarity to the original work or contain the protected elements of the original work.”) (Kadrey Order, p. 2)

Based on the orders in Kadrey and Andersen, a plaintiff will need to show some similarity between a copyrighted work and the output of an AI model to survive a motion to dismiss. 

In conclusion, the recent orders in Kadrey and Andersen provide a clearer path for entities leveraging AI training sets that may include copyrighted works. These orders underscore early indications of a judicial trend requiring more substantive and specific allegations for copyright claims to proceed, particularly emphasizing the need for demonstrable similarity between AI outputs and copyrighted works. While legal complexities remain, these decisions set a bar for plaintiffs that will offer some reassurance to innovators and developers in the field of AI.

The Andersen order can be found here: https://www.courtlistener.com/docket/66732129/118/andersen-v-stability-ai-ltd/

The Kadrey order can be found here: https://www.courtlistener.com/docket/67569326/56/kadrey-v-meta-platforms-inc/