Artists suing artificial intelligence-based art-generating machines have overcome a major hurdle in their first lawsuit over the free and unauthorized use of billions of images downloaded from the internet to train AI systems, with a federal judge allowing the lead case to move forward.
U.S. District Judge William Orrick on Monday upheld allegations of copyright and trademark infringement, marking a significant victory for artists. He found that Stability Diffusion, an AI tool from Stability that can create hyperrealistic images in response to multi-word prompts, was “substantially based on copyrighted work” and may have been created with the intent to “facilitate” infringement. The order potentially exposes AI companies that incorporate the models into their products to potential lawsuits.
The claims of breach of contract and unjust enrichment against the companies, as well as violations of the Digital Millennium Copyright Act for removing information identifying intellectual property, were dismissed. The case will now move to discovery, where artists may be able to uncover information about how the AI companies harvested their copyrighted material and used it to train large-scale language models.
The lawsuit was filed by Carla Ortiz, who has worked on projects like “Black Panther,” “Avengers: Infinity War” and “Thor: Ragnarok,” and is credited with coming up with the main character design for “Doctor Strange.” As AI tools proliferate in production pipelines, concept artists like Ortiz are prepared for further displacement in the future as the technology advances and courts side with AI companies on certain intellectual property issues the tools raise.
The widespread adoption of AI in the filmmaking process will depend largely on how courts rule on the new legal issues raised by this technology. One of the few factors preventing further adoption of this technology is the possibility of courts ruling that training an AI system with copyrighted material constitutes copyright infringement. Another factor is that AI-generated works are not eligible for copyright protection.
The lawsuit, filed last year, concerns the LAION dataset, which was built using 5 billion images allegedly scraped from the internet and used by Stability and Runway to create Stable Diffusion. The suit implicates Midjourney, which used the model to train its AI system, and DeviantArt, which used the model in its image generation tool DreamUp.
In dismissing the lawsuit, Stability and Runway disputed the artists' claims that Stable Diffusion induced copyright infringement and that the Stable Diffusion models themselves constitute infringing works. According to this theory, if third parties use the models provided by Stable Diffusion, the distribution of the models could induce copyright infringement and expose the artists to substantial damages.
Orrick sided with the artists, concluding that the artists had sufficiently argued that Stable Diffusion was made from copyrighted material and that “the way the product works necessarily invokes copies of those works and protected elements.” In a conclusion that could be problematic for AI companies that use the model, he said that Stability and Runway could facilitate copyright infringement, and that Stable Diffusion was “intentionally created to facilitate that infringement.”
In dismissing the copyright infringement claim last year, the court found the case's theory “unclear” as to whether copies of training images stored on Stable Diffusion were used by DeviantArt and Midjourney. The court pointed to the defendants' argument that it is impossible to “compress billions of images into an active program” as Stable Diffusion does.
After the lawsuit was dismissed, the Artists amended one prong of their lawsuit to allege that Midjourney separately trained its products on the LAION dataset and incorporated Stable Diffusion into its products.
In another loss for the AI companies, the court rejected arguments that each of the complaining artists was required to identify in their lawsuits the specific, individual works they claimed were used in training.
“Given the unique facts of this case, including the size of the LAION dataset and the nature of Defendants' products, as well as additional allegations challenging the transparency of Stable Diffusion's core 'open source' software, Plaintiffs do not need that level of detail to state their claims,” the order states.
At the May hearing, DeviantArt warned that if artists' copyright infringement claims against companies that merely used Stable Diffusion but were not involved in its creation were dismissed, several other companies could be sued.
“It's hard to overstate the disruption that a lawsuit against DeviantArt would cause,” said the company's lawyer, Andy Gass. “There are countless parties that could be litigated that are no different from[our company].”
Gass further stated that DeviantArt “did not develop any generation of AI models” and that “all it allegedly did was obtain StabilityAI's Stable Diffusion model, download and upload it, and provide the DreamUp version to its users.”
The court also highlighted that Midjourney created images that resembled the artists' work when it used the artists' names as prompts. This, along with the allegation that the company posted images incorporating the plaintiff's name on its site to promote the features of its tools, provided a basis for proceeding with a trademark infringement claim. Whether consumers would be misled into thinking that Stability's actions were an endorsement of the artists' products could be examined at a later stage in the litigation, the court said.
In a thread on Discord, the platform MidJourney operates, CEO David Holtz posted the names of about 4,700 artists that his company said its AI tools could recreate. Earlier, Stability CEO Prem Akkaraju said the company downloaded large amounts of images from the internet and compressed them in a way that allowed it to “recreate” them.
During discovery, lawyers for the artists are expected to seek information about how Stability and Runway constructed the Stable Diffusion and LAION datasets. They represent several individuals, including Sarah Andersen, Kelly McKernan and Ortiz.
