Like NFTs, AI is imposing a paradigm shift in the art world. We wrote previously about an individual who is suing the U.S. Copyright Office for its refusal to register AI-authored artwork and asking the court to grant the registration. Recently, a German photographer declined to accept a prize won at a photo contest by admitting that the photograph was generated by using AI. He said he submitted the AI-authored work in order to start a debate: “We, the photo world, need an open discussion…A discussion about what we want to consider photography and what not. Is the umbrella of photograph large enough to invite AI images to enter—or would this be a mistake?” The photographer’s personal views on it is that “AI is not photography”; hence, his refusal to accept the award.
The release of products such as Stable Diffusion, an open-source tool that uses deep learning to create images from text prompts, is described as being ground-breaking and a game-changer in the world of technology art. This article will not dive into the technical aspects of how the deep learning occurs, but suffice it to say that the technology requires the tool to be “trained” by accessing data (5 billion pairs of images and text) taken from datasets scraped from the Internet. The end result is AI-generated art that matches the text: type in “dragon with nine heads” and you should get some output image that resembles that.
Not surprisingly, a group of people who are unhappy with these technological developments are artists whose artworks have been included as part of the training data set. In Sarah Andersen et al. v. Stability AI Ltd., et al., three artists filed a class action lawsuit in federal court in California against Stability AI, the founder of Stable Diffusion; Midjourney, another company offering an AI image product, whose product is allegedly based on Stable Diffusion; and DeviantArt, an online art gallery and community (to which the plaintiffs belong) that also offers DreamUp, an AI image product that relies on Stable Diffusion to produce images.
The core allegations in the artists’ complaint are as follows. In order to train Stable Diffusion, Stability downloaded billions of copyrighted images without the authors’ permission to create Stable Diffusion, and caused those images to be stored at and incorporated into Stable Diffusion as compressed copies. The new images generated when prompted by a user are entirely based on the training images, and accordingly are alleged to be derivative works of the training images. These images compete with the original works in the marketplace—whereas before, a user had to pay to license an original image from the artist; now with Stable Diffusion and other AI image software, users can input an artist’s name in the program and generate artworks “in the style of” that artist without compensating the artist. Plaintiffs allege that among the billions of training data are plaintiffs’ and the other proposed class members’ copyrighted artworks, including those that were placed on DeviantArt, which were scraped and used without their (or DeviantArt’s) consent. Plaintiffs assert claims for direct and vicarious copyright infringement, violation of the Digital Millennium Copyright Act, violation of statutory and common law right of publicity, unfair competition, breach of contract (against DeviantArt only) and declaratory relief.
All of the defendants have filed a motion to dismiss the complaint, and DeviantArt has filed an anti-SLAPP motion to strike the right of publicity claims on the basis that the claims chill free speech, on which the other two defendants join in. The defendants raise a number of arguments as to why the complaint is subject to dismissal.
One obvious potential hurdle for plaintiffs may be that a couple of the artists fail to allege that they own copyright registrations, which is a prerequisite to suing. Aside from this, the defendants argue that the complaint makes general allegations of infringement of plaintiffs’ works, but does not identify any specific work that is alleged to have been infringed. Nor does it identify any specific output image that is alleged to be infringing. Further, the complaint alleges, “[i]n general, none of the Stable Diffusion output images provided in response to a particular Text Prompt is likely to be a close match for any specific image in the training data.” The defendants seize on these allegations (or lack thereof) to argue that plaintiffs are unable to make out a claim for direct copyright infringement, not only because no specific work is described, but because any showing of infringement requires a showing of substantial similarity. Defendants also reject plaintiffs’ claim that the output images constitute derivative works of plaintiffs’ works simply because they are based on them, as the law requires substantial similarity for there to be infringement of an owner’s exclusive right to prepare a derivative work. As to Midjourney and DeviantArt, they argue that the complaint literally does not allege their involvement in the direct infringement because Stability AI is the only one alleged to have scraped the data for training purposes.
Plaintiffs also bring a claim for vicarious liability. But a predicate for such liability is the existence of direct infringement, which defendants argue is not sufficiently alleged. Although plaintiffs allege that a few impostors have created fake images, defendants claim plaintiffs have not identified what they are, and that again, have pleaded that the output images are not a close match to the original training image. Defendants further argue that plaintiffs insufficiently plead a causal link between the infringement and a financial benefit to the defendants.
Another noteworthy issue which is encompassed in the defendants’ motion to dismiss as well as DeviantArt’s motion to strike (on which the other defendants join) concerns the defendants’ arguments that the right of publicity claims are preempted by the U.S. Copyright Act. Preemption involves the legal principle that a certain body of law (here, Copyright Act) provides the sole remedy for claims invoking rights falling within the scope of copyright law. Defendants argue that plaintiffs’ right of publicity claims—that their names were invoked without consent—is not a genuine right of publicity claim because the case is not about their names, but plaintiffs’ persona which are embodied in the artwork, and thereby squarely within the scope of copyright. Additionally, defendants argue, Plaintiffs’ claim that the output images infringe their artistic identities by generating artwork “in the style of” the artist fails to state a claim. Artistic style is akin to an idea, which is not protected by copyright, but which is still within the subject matter of copyright for preemption purposes. Defendants also argue that the claims fail on the merits, because the complaint fails to allege use of the plaintiffs’ identities through their names, and moreover, it was not defendants but users typing in the names, and merely being able to type an artist’s name in the text prompt does not state a claim, lest all Google searches will become actionable. Furthermore, DeviantArt argues that the output images are sufficiently transformed from the source image so as to come under the protection of the First Amendment.
In a set of response papers filed by the plaintiffs on June 2, plaintiffs refute the defendants’ arguments. Among other arguments raised, plaintiffs argue that direct copyright infringement is adequately alleged by, e.g., the scraping (copying) and reproduction of the training images, the creation and distribution of Stable Diffusion/Midjourney product/DreamUp, and the defendants’ distribution of the output images. And plaintiffs claim that the need to show substantial similarity is a red herring and is not a required showing where there is evidence of direct copying (especially at the pleading stage), but regardless, that substantial similarity is met based on their pleading of how the output image is created based on the source image and that defendants’ arguments are based on a deliberate misreading of the pleading. As to vicarious liability, plaintiffs argue that they have adequately alleged direct infringement by users of the AI programs, that defendants have the ability to stop such infringement, and that they are profiting from the programs through subscription fees.
Plaintiffs also argue that their right of publicity claims survive. According to them, the issue is not whether the output images sufficiently represent plaintiffs’ identity, as framed by defendants, but whether plaintiffs’ names are associated with the output images and whether they harm plaintiffs’ identities. And in this regard, plaintiffs argue that Stability knew that plaintiffs’ names would draw users to their product and that the complaint alleges how it used specific class members’ names to attract customers through permitting prompts “in the style of” a particular artist. The claims are not preempted, according to plaintiffs, because the claims are grounded in plaintiffs’ names, not their style. Further, plaintiffs argue that the First Amendment does not apply because the claims are not based on the similarity of the works, but on the use of plaintiffs’ names, so it is irrelevant whether the output images are transformative of the source images.
Shortly after the filing of the above lawsuit, Stability AI was also sued by Getty Images in the UK and in Delaware based on its scraping of at least 12 million Getty Images content without permission. These filings show that with new technological advancement comes a group of people whose interests are affected in meaningful ways, and the legal boundaries between what is permissible and not are both novel and subtle. A robust debate exists for or against AI art, with some pundits opining that AI-generated art constitutes fair use. Whatever one’s position may be with regard to it, AI advancements are affecting the application of intellectual property laws, and for some, their livelihood.