Earlier this year, three major artists sought class-action lawsuits against AI companies Stability AI, DeviantArt, and Midjourney claiming that the images produced by their generative AI models were dissimilar to their work.
At the time, the artists – Sarah Andersen, Kelly Mckernan, and Kaloa Otiz – alleged that the companies infringed the copyrights of “millions of artists” by training their AI models on five billion images downloaded from the internet without the consent of its original creators.
Artists’ Case Against Midjourney and Stable Diffusion Dismissed on Lack of Evidence
However, the artists were dealt a major blow in the first-of-its-kind lawsuit on Monday after US District Judge Willian Orrick found that the copyright infringement claims against Midjouney and DeviantArt cannot move forward as the accusations were “defective in numerous respects”. Judge Orrick ruled that the plaintiffs failed to provide enough evidence to support their case and granted the defendants a motion to dismiss.
Among the issues discussed were whether generative AI models could run on copies of copyrighted images to create infringing works and if the artist could provide evidence in the absence of identical content created by the tool.
While the lawsuits against Midjounrey and DeviantArt were stalled, Orrick did allow for a separate infringement claim against AI developer Stability AI, creator of Stable Diffusion, to move forward.
In the case against Stability AI, Kelly Mckernan, Karla Ortiz, and Andersen claimed that the company used copyrighted images without their consent to train Stable Diffusion, which they say is a direct infringement of the Digital Millenium Copyright Act (DMCA), the artists’ right of publicity, and the platform’s own terms of service. Judge Orrick acknowledged that the plaintiffs may have an argument on that front but dismissed all three infringements.
Stability AI, which incorporated Stable Diffusion into its AI image generator platform DreamStudio, denied the contention that it stored the copyrighted images on its AI system. The company maintained that the process of training its AI model does not include the wholesale copying of artworks but rather involves the use of parameters such as lines, colors, shades, and other attributes associated with subjects and concepts from those works that could help the system define what an image would look like.
The artists had to establish that their works were specifically used to train Stable Diffusion. They argued that every output image from the AI was derived “exclusively” from source images that are copyrighted, hence, every hybrid image generated by the model is “necessarily a derivative work”.
Attorneys for Stability AI instead argued that the artists should be required to identify which of their copyrighted or “registered” works were copied and compiled into the data used for training Stable Diffusion. However, the judge noted that doing so would be difficult because based on how the generation process works, none of the output images are likely to be a close match for any specific content in the training data. While dismissing the infringement claims, the judge noted that it was impossible for billions of images “to be compressed into an active program” like Stable Diffusion.
Artists Concede That Output Images Are Not a Close Match for any Specific Copyrighted Work
In the latest ruling, Orrick requested the artists to “amend to clarify their theory” related to how AI models contain “compressed copies” of their works and how the algorithms and their instructions could reconstruct those images. This is complicated because the plaintiffs have already conceded that none of the output images returned by Stable Diffusion in response to a particular text prompt was likely to be a close match for any specific image in its training data.
The judge also denied Stability’s motion to dismiss the direct copyright infringement claim. Meanwhile, two of the three artists who filed the litigation have dropped their case because they hadn’t registered their works with the copyright office before suing. The copyright claims will now be limited to Sarah Anderson’s pieces, which have been registered.
To prove that Stable Diffusion had in fact been trained on her material, Anderson provided proof of search results with her name on “haveibeentrained.com” – a platform that allows artists to verify whether their work has been used to train AI models and also offers an option to opt-out to help prevent future unauthorized use.
Orrick also questioned whether Midjourney and DeviantArt, which offers the open-source Stable Diffusion text-to-image generator model through their website and apps, could be liable for direct infringement if the AI system contains only “algorithms and instructions” that can be applied to the creation of images that include “only a few elements of a copyrighted work”.
The judge also asked the artists to clarify their theory against Midjourney as to whether their claims are based on the company’s use of Stable Diffusion or its own independent use of the copyrighted material to train the Midjourney AI model. However, Orrick denied Midjourney’s request to eliminate the class action aspect of the case.
The next hearing in the lawsuit against Stability AI, Midjourney, and Deviant Art is scheduled for November 7, 2023.