Generating Litigation: N.D. Cal. Dismisses Some Copyright Claims in Andersen and Kadrey AI Cases
January 26, 2024
Authored and Edited by Maxime Jarquin; Daniel R. Mello, Jr.; Jenevieve J. Maerker
The groundswell of lawsuits filed against generative artificial intelligence (GenAI) companies based on various theories of copyright infringement shows no sign of abating. Readers following this issue will be aware, for example, of the claims filed by the New York Times Company against OpenAI and Microsoft alleging that the AI companies copied NYT articles and used them to train AI chatbots.
Meanwhile, some of the earlier cases attacking GenAI companies on copyright grounds have begun to yield court decisions that may serve as a roadmap for future litigants and courts to follow when grappling with the novel and thorny issues raised by this increasingly prevalent technology. In the highly publicized Andersen and Kadrey cases, the United States District Court for the Northern District of California has issued opinions dismissing some of the plaintiffs’ copyright claims, emphasizing in both cases the importance of substantial similarity in alleging that an output generated by an AI tool may constitute an infringing derivative work.
Three visual artists—Sarah Andersen, Kelly McKernan, and Karla Ortiz—brought copyright infringement claims on behalf of themselves and a putative class against three AI companies: Stability AI Ltd. (and related corporate entities), Midjourney, Inc., and DeviantArt, Inc. The artists alleged that Stability AI and the other companies use an AI image-generation software engine called Stable Diffusion, which was created by Stability AI. Stable Diffusion drives all three companies’ consumer-facing products—Stability AI’s DreamStudio, Midjourney Inc.’s Midjourney, and DeviantArt’s DreamUp—which enable users to request the creation of custom images based on text prompts.
The complaint accused the AI companies of copyright infringement, claiming that Stability AI used the artists’ works to train Stable Diffusion. Specifically, the artists alleged that Stability AI, using a third-party service, scraped (and thereby copied) over five billion images from the Internet and used them to train the Stable Diffusion software. The artists claimed that Stable Diffusion stores compressed copies of their works and, in response to user text prompts, produces images in the artists’ unique artistic styles. The artists also brought claims for vicarious infringement, violation of the Digital Millenium Copyright Act (DMCA), violation of the common law right of publicity, unfair competition, and declaratory judgment. Each defendant separately moved to dismiss all claims.
The court dismissed McKernan’s and Ortiz’s copyright claims with prejudice because neither artist had registered the copyrights in their images, a prerequisite to filing suit under the Copyright Act. The AI companies argued that the court should also dismiss Andersen’s direct infringement claim because she failed to specify which of her sixteen registered collections of works was used to train Stable Diffusion. Andersen relied on results she found by searching for her name on the “ihavebeentrained.com” website, which indicated that some of her registered works were used as training images for Stable Diffusion. The court rejected the companies’ argument, holding that Andersen’s reliance on the search result was sufficient at this stage to support her claim that her works had been copied, especially given her allegation that Stability AI had used five billion images scraped from the Internet to train Stable Diffusion.
Additionally, the court found that Andersen’s allegations that Stability AI downloaded or otherwise acquired copies of billions of copyright-protected images without permission, used those images to train Stable Diffusion, and stored those images in Stable Diffusion as compressed copies, were sufficient to support a claim of direct infringement against Stability AI, and the court therefore denied Stability AI’s motion to dismiss this claim.
However, the court granted DeviantArt’s motion to dismiss Andersen’s direct infringement claim against it. Andersen argued that DeviantArt directly infringed by: (1) distributing Stable Diffusion (which Andersen alleged contained compressed copies of her works) as part of its DreamUp product; (2) creating and distributing DreamUp, itself an infringing derivative work; and (3) generating and distributing images output by DreamUp in response to user prompts, which she claimed are infringing derivative works. Regarding the first two arguments, the court found that the complaint was unclear as to whether Stable Diffusion stored compressed copies of copyrighted works or rather used “statistical and mathematical methods” to capture and store concepts from the training images. As for Andersen’s third argument, the court was emphatic that the artists could not plausibly state a claim that the output images are infringing without showing that the output images are substantially similar to the training images. The court dismissed Andersen’s direct infringement claim against DeviantArt with leave to amend, allowing her to provide greater clarity as to how Stable Diffusion includes compressed copies of training images and to allege additional facts about how DeviantArt can be held directly liable, especially when the complaint also states that DeviantArt simply provides customers access to Stable Diffusion.
The court also granted Midjourney’s motion to dismiss Andersen’s direct infringement claims, with leave to amend. The court held that Andersen failed to allege any facts regarding what training data Midjourney used in its Midjourney product, if any. The court allowed Andersen to clarify her theory against Midjourney, specifically whether her claim is based on Midjourney’s use of Stable Diffusion, on Midjourney’s own independent use of training images for its own product, or both.
Andersen also alleged that the AI companies are vicariously liable for their users’ creation of infringing outputs using their respective GenAI platforms. The court dismissed her claim against Stability AI because the complaint did not offer plausible facts about how the compressed copies of the plaintiffs’ images might be stored in DreamStudio when users generate outputs. With regard to DeviantArt and Midjourney, the court noted that to hold the companies vicariously liable for the activities of their users in creating the output images, Andersen must plausibly allege direct infringement by the users. The court again stated that a plausible claim that the user-created output images are infringing would require an allegation of substantial similarity between output images and training images, which the complaint lacked. Therefore, the court dismissed Andersen’s vicarious infringement claims against DeviantArt and Midjourney with leave to amend.
Finally, the court found Andersen’s claims for violations of the DMCA, right of publicity, and unfair competition laws, in addition to a claim that DeviantArt had breached its own terms of service, to each be factually inadequate, and dismissed all with leave to amend.
In November 2023, the artists filed an amended complaint that, among other things, (1) provided more detailed allegations regarding how Stability AI allegedly copied, compressed, and stored the training images, and how Midjourney and DeviantArt allegedly used the same dataset to train their own platforms; (2) provided visual examples of the artists’ works used as inputs in Stable Diffusion and Midjourney, compared side-by-side with allegedly infringing outputs; and (3) added a new defendant, Runway AI, Inc., which the artists allege collaborated with Stability AI to train Stable Diffusion and incorporated Stable Diffusion into its own AI Magic Tools platform. At the time of publication, the AI companies had yet to respond to the amended complaint.
An example from the amended complaint of one of Gerald Brom’s images allegedly entered into Stable Diffusion (left) and the output generated by the platform (right).
A group of writers, including famous actress and comedian Sarah Silverman, filed a complaint for copyright infringement against Meta in the Northern District of California. Silverman and the others claimed that Meta copied their books to train its AI platform, LLaMa (Large Language Model Meta AI). The authors brought claims of direct and vicarious copyright infringement, violation of the DMCA, unfair competition, unjust enrichment, and negligence. Meta moved to dismiss nearly all of the authors’ claims, except for the claim that the unauthorized copying of the authors’ books for the purpose of training LLaMa constitutes infringement.
In connection with their claim that the text generated by LLaMa infringed their exclusive right to prepare derivative works, Silverman and her co-plaintiffs alleged 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.” Not so, said the court; to prevail on their claim that LLaMa’s outputs are infringing derivative works, the authors needed to allege and prove that the outputs “incorporate in some form a portion of the plaintiffs’ books” (internal quotations omitted). Moreover, citing to the Andersen decision, the court found that the plaintiffs would have to establish the existence of substantial similarity between the outputs and the allegedly infringed works.
The authors further claimed that the “LLaMa language models are themselves infringing derivative works” because the “models cannot function without the expressive information extracted” from the authors’ books. The court rejected this theory “as nonsensical” because a derivative work must be based on one or more preexisting works that are recast, transformed, or adapted into the derivative work. In the court’s view, the LLaMa models cannot be understood themselves as a recasting or adaptation of any of the authors’ works.
In connection with their claims of vicarious infringement, the authors claimed that “every output of the LLaMa language models is an infringing derivative work” and that because third-party users (rather than Meta) are the ones to initiate queries of LLaMa, “every output from the LLaMa language models constitutes an act of vicarious copyright infringement” by Meta. The court was not persuaded by this argument either—because, as noted above, the complaint failed to allege the substantial similarity necessary to make LLaMa’s outputs infringing derivative works. As the Andersen opinion had pointed out, a claim of vicarious infringement must be predicated on an allegation of underlying direct infringement, and the failure to plausibly allege a directly infringing output was therefore fatal to the authors’ claim.
Consistent with this analysis, the court likewise dismissed the remaining claims at issue in the motion to dismiss, leaving only the claim that Meta’s copying for purposes of training LLaMa constitutes direct copyright infringement.
In December 2023, the authors filed an amended complaint in which they provided more specific allegations regarding Meta’s use of the “Books 3” and “The Pile” datasets (which allegedly include the authors’ copyrighted books) to train LLaMa, and regarding Meta’s alleged knowledge that it was using copyrighted works. The authors removed their claim that LLaMa is itself an infringing derivative. In January 2024, Meta filed an answer to the amended complaint in which it denied the authors’ claims wholesale and raised numerous affirmative defenses. The court has held a case management conference and approved a discovery schedule for the litigation, culminating in a summary judgment hearing slated for March of 2025.
The Andersen and Kadrey motion to dismiss rulings provide guidance on what courts are likely to deem sufficient to state a viable claim of copyright infringement in the context of GenAI. First, if the allegedly infringing work is a GenAI output, courts are asking plaintiffs to identify with specificity the allegedly infringing output(s) and to plausibly plead that the output is substantially similar to a specific original work. Second, where multiple defendants are named in a complaint, connected only by use of the same AI technology (such as the defendants in Andersen, who all allegedly use the Stable Diffusion tool), the complaint must clearly and separately allege all claims and underlying facts against each defendant. Third, it is still unclear whether a direct copyright infringement claim can be substantiated solely on the scraping of works and the use of those works to train an AI tool (independent of any allegedly infringing output). This theory of liability remains to be assessed in future proceedings.
The cases are Andersen v. Stability AI Ltd, No. 3:23-cv-00201 (N.D. Cal. 2023) and Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal. 2023).
Copyright © 2024 Finnegan, Henderson, Farabow, Garrett & Dunner, LLP.
DISCLAIMER: Although we wish to hear from you, information exchanged in this blog cannot and does not create an attorney-client relationship. Please do not post any information that you consider to be personal or confidential. If you wish for Finnegan, Henderson, Farabow, Garrett & Dunner, LLP to consider representing you, in order to establish an attorney-client relationship you must first enter a written representation agreement with Finnegan. Contact us for additional information. One of our lawyers will be happy to discuss the possibility of representation with you. Additional disclaimer information.
Webinar
May 9, 2024
Webinar
Workshop
Life Sciences Workshop: Updates and Key Trends in Pharmaceutical and Biotechnology IP Law
May 2, 2024
Cambridge
At the PTAB Blog
USPTO Releases Notice of Proposed Rule Making Codifying Several Precedential Case Factors
April 25, 2024
Hybrid Seminar
Pandora’s Pixels: Can Privacy and Choice Survive in our Generative AI World?
April 11, 2024
Richmond
Due to international data regulations, we’ve updated our privacy policy. Click here to read our privacy policy in full.
We use cookies on this website to provide you with the best user experience. By accepting cookies, you agree to our use of cookies. Please note that if you opt not to accept or if you disable cookies, the “Your Finnegan” feature on this website will be disabled as well. For more information on how we use cookies, please see our Privacy Policy.
Finnegan is thrilled to announce the launch of our new blog, Ad Law Buzz, devoted solely to breaking news, developments, trends, and analysis in advertising law.