Robot Art Riles Artists

Visual artists have survived a motion to dismiss their class claims brought against generative artificial intelligence (AI) companies related to the companies’ use of the artists’ visual works without consent. The plaintiffs claimed that the defendants’ text-to-image AI products were trained in part on their copyrighted works. ABA Litigation Section leaders agree that the case sets up a showdown between copyright interests and the “democratization” of art through AI.

Scraping the Internet Leads to Copyright Scrap

In Andersen v. Stability AI, Ltd., pending in the U.S. District Court for the Northern District of California, the plaintiffs filed a putative class action on behalf of other artists, alleging that their copyrighted works were among five billion images “scraped,” i.e., copied, from the internet into datasets by Large Scale Artificial Intelligence Open Network (LAION), a not-for-profit organization. The defendants included five leading companies in the text-to-image, generative AI business: Stability AI Ltd. and Stability AI, Inc. (Stability); Midjourney, Inc.; DeviantArt, Inc.; and Runway AI, Inc. The defendants developed or based their products on an AI software program named Stable Diffusion, a model that generates artistic images based on textual input.

Stability had hired LAION to create the scraped datasets. The LAION datasets were then used by Stability to train Stable Diffusion. The other defendants used some version of Stable Diffusion in their own generative AI products. The copyright claims included direct infringement by using the images for training the AI models and inducement of infringement by distributing Stable Diffusion for free.

The plaintiffs also brought claims under the Digital Millennium Copyright Act (DMCA), for removing and altering the copyright management information (CMI) of training images and several other claims.

An initial round of motions to dismiss was largely successful with only one direct infringement claim surviving. The plaintiffs had alleged in their initial complaint that Stable Diffusion contained “compressed copies” of the LAION training images. The district court ruled in dismissing the initial complaint with leave to amend that the plaintiffs needed to define “compressed copies” and provide greater clarity on how their copyrights were violated by the defendants. The amended complaint was met with more motions to dismiss.

Copyright Claims Survive a Technological Thicket

On the second go-around, the district court denied motions to dismiss direct and induced copyright claims, citing the need for discovery to help navigate a tangle of technological issues. The amended complaint was found to contain sufficient specificity in describing how the plaintiffs’ copyrights had been directly infringed. The court pointed to amended allegations that the LAION datasets consisted only of URLs and not the training images themselves. A URL is simply a web address that identifies the location of a file on the internet. Thus, to access the copyrighted images for training purposes, Stability had to use certain technologies to extract the actual images from the LAION dataset. The amended complaint also added specificity on the defendants’ use of the process of CLIP-guided diffusion to allow their generative AI models to better match text prompts to the copyrighted images.

The amended complaint added a claim for inducement of copyright infringement through the defendants’ distribution of Stable Diffusion for free, thus enabling third parties to access the plaintiffs’ protected images. In motions to dismiss, the defendants argued that the inducement allegations lacked the required specificity on active steps taken by the defendants to encourage direct, third-party infringement. Here the plaintiffs successfully pointed to a statement by the Stability CEO that “[w]e took 100,000 gigabytes of images and compressed it to a two-gigabyte file that can recreate any of those [images] and iterations of those.”

Another added allegation that the court found compelling was that simply using the plaintiffs’ names as prompts in at least some of the models generated outputs that copy some elements of the copyrighted works.

The court rejected the defendants’ argument that AI technology is essentially like the VCR, which had been found in other cases to be non-infringing. Unlike the VCR, which did not come pre-loaded, here the plaintiffs alleged that AI models are created using copyrighted works and by operation these models necessarily infringe and facilitate infringement by users. The court summarized the essential fact issue to be whether the plaintiffs’ works can be found in some form within the AI models. Whether that form is the image itself or some algorithmic or mathematical representation was found to be irrelevant to the court. The court left for the evidentiary phase of the case the issue, whether substantial similarity between the protected image and the AI output is required.

In a similar vein, the court rejected the defendants’ argument that the fair use defense applied as a matter of law because the copyrighted works were transformative and that they represented a miniscule portion of the massive datasets used to train the models. The court further noted that the fair use defense could be tested by motions for summary judgment.

DMCA Claims Dismissed with Prejudice

The district court dismissed with prejudice claims made under the DMCA dealing with providing or distributing false CMI or intentionally removing CMI. On the claim under subsection (a) regarding false CMI, the court agreed that the generic license that accompanies Stable Diffusion makes no claims to the plaintiffs’ copyrighted works. Dismissal was also based on the failure to allege plausible facts demonstrating the required “double scienter” that the defendants knowingly provided false CMI with the intent to induce infringement.

The court dismissed the subsection (b) claim of removing CMI from the plaintiffs’ works in the training process based on the admittedly “unsettled” requirement that the plaintiffs allege that some output from Stable Diffusion be identical to some plaintiffs’ work. The court acknowledged but ultimately disagreed with case law adopting a less stringent substantial similarity test.

Reconciling Property Rights and Generative AI

Litigation Section leaders agree with the court’s denial of the motions to dismiss the copyright claims. “The Andersen case is one of several important cases in the emerging generative AI environment,” points out Andrew W. Coffman, Tupelo, MS, Co-Chair of the Copyrights Subcommittee of the Section’s Intellectual Property Litigation Committee. “The facts in this complex environment need to be developed through discovery, which will be complicated and contentious,” he says, though he ultimately expects the case will be decided by summary judgment.

“The factual issues involving the training and operation of the AI models and the copyright claims and defenses need to be fleshed out through discovery and are better resolved by summary judgment or at trial,” agrees Matthew D. Kohel, Baltimore, MD, Co-Chair of the AI Subcommittee of the Intellectual Property Litigation Committee. “There is a compelling argument that generative AI output is transformative and thus qualifies as fair use, not requiring the artists’ permission,” he adds, suggesting that the fair use defense will play an important role in this and future cases. “That must be weighed against a potential finding that the use of content to train large language models is commercial in nature,” he cautions.

“Ultimately the law will be determined through cases like this making their way through the circuits and possibly the Supreme Court,” Coffman notes. “There is significant support for both sides of the equation, so I do not expect to see consensus for a legislative solution.” Kohel agrees that congressional action in the field is unlikely but believes this case and others like it may be headed for some form of licensing scheme as part of a settlement.

The stakes are high. “For Stability AI and the other defendants, the copyright claims here represent an existential threat,” Coffman points out. “These generative AI models require access to enormous amounts of data for training, and the time and expense of ascertaining copyright rights may be prohibitive.”