Visual artists who’ve joined together in a category motion lawsuit against among the hottest AI image and video generation corporations are celebrating today after a Judge decided Your copyright infringement case against the AI corporations may proceed to disclosure.
Disclosure: VentureBeat frequently uses AI art generators to create article illustrations, including some mentioned on this case.
The case, registered under number 3:23-cv-00201-WHOwas originally submitted in January 2023. Since then, it has been amended and parts of it repealed several times, including today.
Which artists are involved?
Artists Sarah Andersen, Kelly McKernan, Karla Ortiz, Hawke Southworth, Grzegorz Rutkowski, Gregory Manchess, Gerald Brom, Jingna Zhang, Julia Kaye and Adam Ellis, on behalf of all artists, have accused Midjourney, Runway, Stability AI and DeviantArt of copying their works by offering AI image generator products based on the open-source Stable Diffusion AI model that Runway and Stability AI collaborated on and which the artists say was trained on their copyrighted works in violation of the law.
What the judge decided today
Judge William H. Orrick of the Northern District Court of California, which oversees San Francisco and the guts of the generative AI boom, has not yet made a final decision on the final result of the case. But in his decision released today, he wrote that “the allegations of induced copyright infringement are sufficient” to maneuver the case to the invention phase, where the artists' lawyers could dig into the businesses that make AI image generators and examine documents to divulge to the world more details about their training data sets, mechanisms and the way they work.
“This is a case wherein plaintiffs allege that Stable Diffusion relies in significant part on copyrighted works and that the operation of the product necessarily requires copies or protected elements of those works,” Orrick's decision states. “Whether that is true, and whether it’s the results of a mistake (as Stability claims) or intentional (plaintiffs claim), can be considered at a later date. The allegations of induced copyright infringement are sufficient.”
Artists react with applause
“The judge lets our copyright claims undergo and now we're learning all of the things these corporations aren't telling us through discovery,” wrote one among the suing artists, Kelly McKernan, on her social media account X. “This is a HUGE win for us. I'm SO happy with our incredible team of lawyers and co-plaintiffs!”
“Not only are we pursuing our copyright claims, this order also signifies that corporations using SD (Stable Diffusion) models and/or LAION-like data sets could now be held responsible for copyright infringement, amongst other things,” one other artist and plaintiff within the case, Karla Ortiz, wrote on her X account.
Technical and legal background
Stable diffusion was reportedly trained on LAION-5Ba dataset of greater than 5 billion images collected by researchers from across the Internet and posted online in 2022.
However, as noted within the case itself, this database only contained URLs or links to the pictures and text descriptions. This signifies that the AI corporations would have needed to scrape or screenshot the pictures individually to coach Stable Diffusion or other derived AI model products.
A silver lining for AI corporations?
Orrick gave the AI image generator corporations a victory by dismissing the lawsuits brought against them by the artists under the Digital Millennium Copyright Act of 1998, which prohibits corporations from offering products designed to bypass controls on copyrighted material offered online and thru software (also often called “digital rights management” or DRM).
Midjourney attempted to point to older court cases “concerning jewelry, wood cutouts, and keychains” which found that similarities between various jewelry products and people of earlier artists couldn’t constitute copyright infringement because they were “functional” elements, that’s, they were mandatory to represent certain real-life features or elements or that the artist intended to create, no matter their similarity to earlier works.
The artists claimed that “stable diffusion models use CLIP-driven diffusion,” which relies on prompts resembling the artists’ names to generate a picture.
CLIP, an acronym for “Contrastive Language-Image Pre-training”, is a neural network and AI training technology developed by OpenAI back in 2021, greater than a yr before ChatGPT was unleashed on the world. It can discover objects in images and annotate them with natural language text labels – and is an awesome assist in assembling a dataset for training a brand new AI model like Stable Diffusion.
“The CLIP model, plaintiffs allege, functions as a trade dress database that may retrieve and recreate the weather of any artist's trade dress,” Orrick writes in a single section of the ruling on Midjourney, later stating, “The combination of identified elements and pictures, when considered with plaintiffs' allegations about how the CLIP model functions as a trade dress database and Midjourney's use of plaintiffs' names in its Midjourney name list and presentation, provides sufficient description and plausibility for plaintiffs' trade dress claim.”
In other words, the undeniable fact that Midjourney used each artist names and tagged elements of their works to coach its model may constitute copyright infringement.
But as I even have argued before – from my viewpoint as a journalist, not as a copyright lawyer or expert on this field – it’s already possible and legally permissible for me to commission an artist to create a brand new work within the form of a copyrighted artist's work would appear to undermine the plaintiff's claims.
We will see how well the AI art generators can defend their training practices and model results because the case moves forward. Read the complete document embedded below:
