A class action copyright infringement lawsuit was filed against Stability AI (Stable Diffusion), Midjourney, and DeviantArt, Inc. (DreamUp) on behalf of artists who allege their works were used to train #AI generative art algorithms. THREAD 🧵 to come 👇 once I read!

It was only a matter of time before the issue of copyright infringement by AI apps was tested by the Courts. In our Nov 2022 @Law360 article, “A Look at Future AI Questions” we asked whether & when AI generated work is an infringing derivative work: /2

The class action plaintiffs in this ND Cal suit, which is - as far as I know - the first asserting these types of claims against generative art AI apps are artists Sarah Andersen, Kelly McKernan, and Karla Ortiz. /3

The named Defendants in this class action IP lawsuit are Stable AI, the company that created Stable Diffusion, as well as MidJourney and Deviant Art (DreamUp app). All of these are AI image generation apps, similar to DALL-E.   /4

The causes of action in the Stable Diffusion Class Action include claims of direct copyright infringement, vicarious infringement, DMCA violations (for the removal of copyright management info),  California right of publicity and related unfair competition claims.  /5

In the main, the Complaint alleges that the defendants AI apps use the works of the plaintiffs without permission to train AI through machine learning to create new AI-generated works, some of which constitute infringing derivative works. /6

The 46-page Complaint, which includes a ToC, leads off stating that “AI IMAGE GENERATORS ARE 21ST-CENTURY COLLAGE TOOLS THAT VIOLATE THE RIGHTS OF MILLIONS OF ARTISTS.” It details how Stable Diffusion operates, as well as the assembly / use of massive training data sets. /7

In the section containing its thesis, the Complaint explains the crux of its claim: The allegations here focus on the ability of some generative AI apps to create art “in the style of” a particular artist based on art by that artist included in its training data set. /8

It’s interesting that this “in the style of” feature appears to be the focus of the claims. This is narrower than what may be possible to allege in terms of copyright infringement/derivative work. There are also questions about whether it is copyright infringement to do so. /9

Also interesting is the allegation of what the derivative work is: “ Derivative Work”…refers to the output of AI Image Products as well as the AI Image Products themselves—which contain compressed copies of the copyrighted works they were trained on.” /10

In describing the set of plaintiffs, the Complaint notes that each has a number of registered copyrighted works included in the Training Data. /11

The support for citing the number of works included in the AI Training Data for each of the plaintiff artists is: /12

Because this is a Class Action, the Complaint also includes class definitions and FRCP Rule 23-related allegations. (Aside from substantive copyright related defenses, compliance with the class action requirements is another route to defend against these claims) /13

The factual allegations are long and detailed. The Complaint notes they offering these apps for free rather than for a paid license had contributed to their rapid adoption. Also noteworthy is the detailed description of how Stable Diffusion works. /14

Significantly, the Complaint alleged that the very point of the complex mathematics behind Stable Diffusion is “to reconstruct copies of the training data with maximum accuracy and fidelity to the Training Image. It is meant to be a duplicate.” /15

The other bottom line point made from this technical description is allegation that the resulting images are derivatives from the original copyrighted images. /16

Another point that the Complaint emphasizes about the operation of Stable Diffusion is the use of highly complex interpolation techniques and “conditioning” where the resulting image may not look like a copy of any single copyrighted source image. /17

Nonetheless, the Complaint alleges, “resulting image is necessarily a derivative work, because it is generated exclusively from a combination of the conditioning data and the latent images, all of which are copies of copyrighted images.” /18

It turns next to where the Training Data comes from, alleging “Though the rapid success of Stable Diffusion has been partly reliant on a great leap forward in computer science, it has been even more reliant on a great leap forward in appropriating copyrighted images.” /19

The Training Data is from a project called Large-Scale AInOpen Network (LAION), which the Complaint alleged pulls data from copyrighted data sets like Getty Images, Shutterstock, etc. /20

I will turn next to a review of the specific Claims here, starting with the copyright infringement claims. /21

But first, I need to 🛌 😴 💤 More tomorrow!!!

. @threadreaderapp unroll