A class action against generative artificial intelligence (AI) companies cleared initial hurdles this week in the U.S. District Court for the Northern District of California, which permitted claims of direct and contributory copyright and trademark infringement to proceed, while dismissing other claims. On August 12, 2024, in Andersen, et al. v. Stability AI Ltd., et al. (Case No. 23-cv-00201-WHO), Judge William H. Orrick assessed the legal sufficiency of an amended complaint brought by individual artists in a putative class action against a generative AI developers Stability AI, Midjourney, DeviantArt and Runway AI.
According to the plaintiffs, these defendants scraped the Internet for artworks and billions of other images and aggregated those images into datasets, which were used to train AI models to generate outputs (including images and video) based on text inputs. By doing so, the plaintiffs allege, these generative AI companies impermissibly copied and utilized works protected under the Copyright Act. Judge Orrick’s newest order was his second assessment of the plaintiffs’ claims; on motions to dismiss the original complaint, the court largely dismissed the originally pled claims with leave to amend.
As to Stability AI, the plaintiffs alleged in their amended complaint that the company created and distributed AI models based on copyrighted works, and that those models were impermissible copies and/or derivative works. The plaintiffs further allege that Stability AI induced infringement by distributing the models to end users, who downloaded and used them in violation of copyright law. Judge Orrick found the plaintiffs’ amended allegations sufficed to state copyright infringement claims, noting that the software “was built to a significant extent on copyrighted works” and “necessarily invokes” copies or elements of those works. Judge Orrick held that “[t]he plausible inferences at this juncture are that Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate that infringement by design.”
As to Runway AI—which specializes in using AI to generate video outputs—the amended complaint asserts four theories of infringement. First, that Runway AI used protected works to train its AI software. Second, that the AI software itself is an infringing copy and/or derivative work due to its ingestion and transformation of protected images. Third, that distributing the AI software was equivalent to distributing the underlying protected works themselves, and therefore amounted to infringing upon the plaintiffs’ distribution rights. At this stage, Runway AI did not challenge these three theories of direct infringement. The plaintiffs’ fourth theory is that Runway AI induced copyright infringement by end users because it distributed its software, which it developed and trained, knowing that the software would function based on protected works. Judge Orrick also allowed this claim to proceed.
For its part, Midjourney argued that the copyright claims failed as a matter of law because the plaintiffs failed to identify each protected work allegedly infringed. However, because of the unique facts of the case, including the size of the dataset, Judge Orrick reasoned, the complaint did not have to contain such level of detail to survive a motion to dismiss. Accordingly, the court found that the existing allegations were plausible and sufficiently supported the plaintiffs’ infringement theories. The plaintiffs also brought a claim for trademark infringement against Midjourney, alleging that Midjourney used the artists’ names in connection with the software, thereby misleading consumers into believing that the plaintiffs endorsed the software. Judge Orrick found that these allegations were sufficient to plausibly state a claim under the Lanham Act. In the same vein, he held that the plaintiffs sufficiently stated a claim for trade dress violation by alleging that Midjourney’s software was trained to recreate the “inherently distinctive [] look and feel” of the plaintiffs’ artwork.
Finally, as to DeviantArt, the plaintiffs allege that it has implemented software provided by other defendants into its own product, which uses images generated by the software without directly using the underlying protected works. DeviantArt argued that the ratio of protected works or elements thereof that its product is capable of generating is marginal and, therefore, insufficient to support a claim for copyright infringement, or at least that such ratio is insignificant as to constitute fair use as a matter of law. Judge Orrick, however, held that these arguments relied upon evidentiary support and, therefore, were not appropriate for adjudication at the pleading stage.
At bottom, Judge Orrick largely held that the plaintiffs (which included newly added plaintiffs) had sufficiently alleged facts that—taken as true—permitted the copyright and trademark claims to proceed and that they would be evaluated based on evidence at the summary judgment stage (in his order, he did grant the defendants’ motion to dismiss DMCA, unjust enrichment, breach of contract, and breach of the implied covenant of good faith and fair dealing claims).
It is worth noting that, as to the copyright claims, the Court noted that the nature of the claims against generative AI companies means that reliance on “run of the mill” substantial similarity copyright cases are “unhelpful in this case where the copyrighted works themselves are alleged to have not only been used to train the AI models but also invoked in their operation.” He then noted that questions of fair use are for summary judgment—which is where we expect the real battle will be joined.