By: Jack Gagner
The “golden age” of artificial intelligence (“AI”) image generation, or at least the buzz surrounding it online, shows no sign of slowing down in 2023. Although AI image generation has been around for years, widely accessible and free online programs have recently delivered it to the internet-savvy masses. Today, “[t]he most common use for [AI] image generators,” technology developed in papers titled “Deep Unsupervised Learning Using Nonequilibrium Thermodynamics” and “Denoising Diffusion Probabilistic Models,” is to create “custom memes.” Custom memes may come at a price though, as long-held fears over the relationship between art and AI image generation have recently hit the U.S. Federal Courts.
AI image generation operates by transforming text provided by the user into detailed images through a process known as “diffusion.” Developed in 2015 by a team of researchers at Stanford, diffusion consists of two phases. First, “noise,” made up of repeated unstructured fluctuations, is added to an image until the image has been “diffused” into random noise. Next, having recorded the steps taken in phase one, the program runs the same sequence in reverse order until the random noise closely resembles the original image. The software then “learns” by repeating this process for millions or billions of images. By storing the steps required to denoise each training image, the software has a library of latent images that it can faithfully reproduce, each with accompanying textual associations. With this store of information, when prompted, the software mathematically blends different latent images into a new, interpolated latent image, which it then converts back into a standard pixel-based image for the user. The final image is thus both derived entirely from data acquired from provided training images and an original use of that data.
A complaint filed on January 13th in the United States District Court for the Northern District of California seeks to challenge the originality of these AI generated images. The suit was brought by three artists claiming that Stability AI, Midjourney, and DeviantArt have violated copyright laws by developing AI image generation software that uses copyrighted images to produce derivative, rather than original, works. According to the complaint, “the rapid success of Stable Diffusion,” the software underlying all three image generation programs, “has been partly reliant on a great leap forward in computer science, [but] even more reliant on a great leap forward in appropriating copyrighted images.” The complaint alleges that by scraping copyrighted images from the internet without permission; copying, compressing, and storing those images in training data; and using a “modern day collage tool” to assemble new images from that data, all works produced by the image generators are derivatives of the copyrighted works in the training data.
Since the complaint was filed, writers have weighed in to challenge both the complaint’s factual and legal claims. Dr. Andres Guadamuz, a reader in intellectual property law at the University of Sussex, believes the complaint does not present an accurate representation of the technology involved. Guadamuz points out that diffusion, and the model that it produces to generate new images, is only part of the AI image generation process. Software like Stable Diffusion also relies on CLIP, a model concerning “a wide range of tasks involving both language and images,” to improve how it understands the relationship between words and associated images – crucial to the generation of images in response to original prompts. Guadamuz also clarifies that the software relies on latent space to store “clusters of representations” associated with textual prompts, rather than storing copies or representations of any original images in the model directly.
Legal considerations include both intellectual property doctrine and far-reaching policy implications. Determining exactly which images, out of the potentially billions of training images, a particular AI-generated work unfairly copied seems a difficult proposition in all but the most obvious cases. Furthermore, the plaintiffs’ theory of liability based on no more than the use of web-scraped images in an AI training model seems a diluted foundation of liability that would fatally undermine the legality of web scraping specifically, and the development of AI more broadly. The suit also has to overcome the idea that copyright protects only specific images, not styles – its characterization of the software as a “21st century collage tool” may be at odds both with the very nature of the suit and established copyright law.
The case-specific factual and legal issues that seem to favor a result for the defendants do little to assuage fears of AI intrusion in the art community. Despite the technical nature of the questions raised by this case, it seems unfair for artists to go unrecognized and uncompensated for the use of their artwork, however attenuated that use may be. As AI image generation software continues to develop and disseminate, the purpose of copyright law to encourage and reward the production of creative works should not be entirely set aside in the name of technological progress.
Student Bio: Jack Gagner is a second-year student at Suffolk University Law School. He is a staff writer on the Journal of High Technology Law. Jack received a Bachelor of Music Degree in Classical Trombone Performance from the University of Toronto.
Disclaimer: The views expressed in this blog are the views of the author alone and do not represent the views of JHTL or Suffolk University Law School.
