Who’s Afraid of the Big Bad Gen-AI Model?

Developers use artists’ works to train their generative AI models. Those models can then create new visual works, even in the style of a specific artist. So, is this copyright infringement, fair use, or something else entirely? Here, AMBART LLC explains not just the law but also how these AI models are trained.

Case Background
There are several pending copyright infringement lawsuits involving artificial intelligence (“AI”), specifically generative AI. In this blog post, we examine Andersen v. Stability AI,[1] a putative class action in the Northern District of California – for direct and induced copyright infringement, false endorsement, trade dress, and other claims – involving the AI model Stable Diffusion. In Andersen, several visual artists (“Plaintiffs”), including the cartoonist and illustrator Sarah Andersen, allege that Stability AI, Midjourney, DeviantArt, and Runway AI (collectively, “Defendants”) use and promote Stable Diffusion, which was trained on their visual works. Moreover, Plaintiffs allege that Stable Diffusion can produce visual works in their style, and in fact, Defendants allegedly touted that they could generate work “in the style” of a specific artist when prompted by a user. Finally, Plaintiffs allege that DeviantArt’s artwork generator DreamUp results in direct and induced copyright infringement,[2] false endorsement and trade dress violations under the Lanham Act, and DMCA (Digital Millennium Copyright Act) violations.

It’s Math not Magic
Before we discuss Plaintiffs’ claims and the Court’s August 12, 2024 decision on Defendants’ motions to dismiss Plaintiffs’ Amended Complaint, let’s make sure you understand what Stable Diffusion is. Stable Diffusion is a deep-learning, text-to-image generation model, which uses diffusion techniques to create images. (The other common generative AI tool that produces images uses a generative adversarial network (GAN) rather than a diffusion process.) For the end user, that means they enter a textual description of the type of image they would like to see, and Stable Diffusion then generates the image based on the user’s prompt. In order to do this, Stable Diffusion is trained on a large dataset of images with corresponding textual descriptions. 

During the model’s training, the first phase is "Noise Addition,” where signal noise[3] is gradually added to the image, until the original image becomes unrecognizable. (Think of taking a photograph of a dog in a grassy yard and adding “noise” or static to the image until it becomes a collection of colored pixels that does not look like anything recognizeable.) Then, in a reverse process, called “Denoising,” the model learns how to reverse the noise, step by step, to reconstruct the original image. 

But, here is the important part: the model does not store the original image it was trained on. Rather, it remembers the steps it took to “denoise” or restore the image. And so, the next time someone asks for an image of a dog in a yard, the model will not re-produce the exact image it was trained on, but will remember how to produce such an image, while also taking guidance from the user’s description. 

Defendants’ Motions to Dismiss
On August 12, 2024, the Court issued an order granting in part and denying in part Defendants’ motions to dismiss the first amended complaint.

Copyright Infringement Claims
The U.S. Copyright Act confers several exclusive rights to creators, including the right to create derivative works. This means that only the original artist is free to adapt her work, and others must seek a license or some other form of permission. Unauthorized use constitutes direct infringement. Induced copyright infringement, also alleged by Plaintiffs in this case, occurs when one knowingly encourages or influences another party to commit acts of copyright infringement. 

Here, the Court denied Defendants’ requests to dismiss Plaintiffs’ direct and induced copyright infringement claims. 

As against Defendant Stability AI, the Court found that, because the resulting allegedly infringing works were caused by the end users prompting the AI product, it was plausible that the model itself could have been “created to facilitate that infringement” as it was trained “to a significant extent on copyrighted works.” In other words, in order to operate the model and generate an output, the end user’s prompts “necessarily invoke[d]” protected elements of copyrighted material. The Court also pointed to statements that Stability AI’s CEO made (and which Plaintiffs relied on in their arguments), which emphasized the Stable Diffusion model’s skill at replicating works used in the training data set. As against Defendant Runway AI, which distributed and helped to develop the Stable Diffusion 1.5 model, the Court also denied the motion to dismiss the copyright infringement claims, including induced infringement. 

The Court also allowed Plaintiffs’ claim for direct infringement against Defendant Midjourney to proceed. Plaintiffs allege that Midjourney used Stable Diffusion to build its own AI product and (in the development and design phases) trained the model using data sets that included Plaintiffs’ copyrightable works. Plaintiffs allege that their names were included in input prompts and that the models produced outputs substantially similar to their copyrightable works.

As for Defendant DeviantArt, the Court denied the motion to dismiss the direct infringement claim, finding Plaintiffs stated a plausible claim that the Stable Diffusion model exists within DeviantArt’s DreamUp tool. Plaintiffs allege that certain protected elements of their copyrighted works remained in the Stable Diffusion model after training, and that those copyrightable elements were then used to create outputs which mimicked Plaintiffs’ works.

Unjust Enrichment Claims
The Court dismissed Plaintiffs’ state law unjust enrichment claims (with leave to amend) against each of the defendants, finding that they are preempted by the Federal Copyright Act.  Because Plaintiffs claimed their copyrighted works were used “to train, develop and promote” Defendants’ AI models without their consent (which would be copyright infringement) and did not identify an “extra element” that would “change the nature of those state claims to protect something other than rights under the Copyright Act,” the Court dismissed the unjust enrichment claims.  

Lanham Act Claims 
Additionally, the Plaintiffs’ Lanham Act claims against Defendant Midjourney, including false endorsement and vicarious trade dress, survived the motion to dismiss. Plaintiffs allege that Midjourney used their names to advertise the styles that its AI product could produce, in connection with a gallery of user-created images on its website (some in the style of Plaintiffs’ work). This could cause a likelihood of confusion as to whether Plaintiffs’ endorsed the Midjourney AI product. As for the vicarious trade dress infringement claim, the Court concluded that Plaintiffs’ claim was sufficient to survive a motion to dismiss, noting that the “combination of identified elements and images [trade dress]” combined with how the model stored and recalled various styles of art in response to prompts, functionally created a “trade dress database.” Moreover, Midjourney allegedly used Plaintiffs’ names to showcase its AI product, which the court found as a plausible basis for Plaintiffs’ trade dress claim. Additionally, Midjourney’s “joint ownership or control” over the images, provided further basis for the claim. 

Breach of Contract Claim
The Court dismissed Plaintiffs’ breach of contract claim against DeviantArt with prejudice, based on Plaintiffs’ assertion that DeviantArt knew Stable Diffusion was trained on datasets that “had been scraped in part from DeviantArt’s website.”  Because Plaintiffs conceded that DeviantArt “played no role in the scraping or training” of the Stable Diffusion model, the Court dismissed this claim.

DMCA Claims
The Court dismissed each of the Plaintiffs’ Digital Millennium Copyright Act claims with prejudice, for several reasons. For Plaintiffs’ claims against Stability AI and Runway AI for license disclosures, the Court found it implausible that readers of those disclosures would understand that Defendants are claiming rights to or conveying false information regarding Plaintiffs’ copyrightable works. As to the claim regarding the removal or alteration of Copyright Management Information (CMI)[4], the Court found Plaintiffs failed to allege that any output from Stable Diffusion was identical to a Plaintiff’s work, which is necessary to state such a claim. 

Next Steps
The litigation is ongoing. Defendant Midjourney has filed a motion for clarification or, alternatively, limited reconsideration of the Court’s order granting in part and denying in part its motion to dismiss Plaintiffs’ first amended complaint.

But is it Copyright Infringement?
Because the Court did not dismiss Plaintiffs’ copyright claims, those claims may proceed to the merits, and we may one day get some clear answers. But, what is currently unclear is whether using an artist’s work to train a generative AI model results in copying, in the traditional sense. 

Generative AI is defined by the International Association of Privacy Professionals (IAPP) as “a field of AI that uses deep learning trained on large datasets to create content, such as written text, code, images, music, simulations and videos, in response to user prompts.” Note that what distinguishes generative AI models from other forms of deep learning is its ability to generate new content, based on predictions it makes from existing data it was trained on. Users engage with a generative AI model by inputting prompts, such as text, audio, video or images. The prompts cause the model to generate new content, in response to the users’ input. 

Unlike traditional software that relies on pattern detection and labeling and storing vast amounts of data, generative AI models are trained by weights and assigning different weights to different parameters.[5] The specific images they were trained on are not stored or retrieved. As a result, when a user prompts the generative AI tool to create an image or a painting or a comic strip in a specific style, the generative AI model effectively “remembers” having seen that and similar images. In the case of the Stable AI’s Stable Diffusion model, as we explained above, the model remembers the steps it took to denoise an image with a similar description.

Therefore, it is not clear to us that AI models are copying copyrightable works, in the traditional sense. This is less like the situation in which a person looks at an image of a non-textual work (picture Van Gogh’s Irises) and then copies it when prompted. Rather, it is more like a person who remembers having seen numerous Van Gogh works and can produce a work in that style and with that same feeling. Humans, and generative AI models, can accomplish this without “going back” to reference a single, particular Van Gogh (or any other style of artists’) work.

Defenses to Copyright Infringement
Fair Use: Humans who are creating new works, even when they reference an image, invariably add their own artistic touches and personality. In those cases, when faced with claims of copyright infringement, they often argue "fair use.” Fair use is one of the key defenses to copyright infringement. It allows others to use a copyrighted work for certain purposes, including criticism, comment, reporting, and research. The Copyright Act also lists four factors that courts evaluate to decide whether the copyrighted material was used in a way that was sufficiently transformative (and thus a fair use and not an infringement). Those fair use factors are:

1) The purpose and character of the use,

2) The nature of the copyrighted work,

3) The substantiality of the portion used, and

4) The effect of the use on the potential market of the artwork.

Although all four factors are considered, courts have historically given the most weight to the first factor—which judges often characterize as whether the secondary work is “transformative”—and regard this factor as “[t]he heart of the fair use inquiry.” See Blanch v. Koons, 467 F.3d 244, 251 (2d Cir. 2006) (quoting Davis v. The Gap, Inc., 246 F.3d 152, 174 (2d Cir. 2001). The U.S. Supreme Court has instructed courts to consider “whether the new work merely supersedes the objects of the original creation, or instead adds something new, with a further purpose or different character, altering the first with new expression, meaning or message.” Campbell v. Acuff-Rose Music Inc., 510 U.S. 569, 579 (1994).

It will be interesting to see how the fair use test is applied to instances where an AI model is doing the "copying," given that they are not storing images and thus the “substantiality of the portion used” factor is difficult to interpret. That said, the other factors, such as the purposes and characters of the use and the effect on the potential market for the artworks, weigh against these AI models.

Independent Creation: There’s also the question as to whether an independent creation of a work is in fact copyright infringement. Independent creation is a defense to copyright infringement, where defendants can show that, in creating the allegedly infringing work, they did not have access to the plaintiff’s copyrighted work. Proving that a defendant had no access to such work is increasingly difficult to establish. In the current information age, access to copyrightable work is a simple click or tap away. 

When the defense of independent creation is applied to copyrightable work used to train AI models, however, there is now an interesting twist. It is true that the model would necessarily have to “see” the work during its training (which, in a traditional application, would lose the defense, due to the access of the work). But, because the model itself does not retain or “store” the image (as discussed above) then this could be a plausible argument for independent creation. The image is not kept in a “fixed” state, and so the model would not have access to the work in the traditional sense.[6]

Alternatives for Recovery
There may be other laws that, with further case law, could be more developed and better suited to this type of “copying” or misappropriation. For example, the Lanham Act prevents misappropriation of trade dress, which are characteristics used to identify the source of a product and distinguish it from others (think: the shape of a Coca-Cola bottle). Trade dress could potentially provide relief where the Copyright Act falls short, such as securing the artists’ names, reputation and aesthetic style. This could be useful for artists with a very specific style, and where the output of a generative AI model is not the same as the original (so, it is not a direct infringement), but could otherwise confuse consumers as to who made it.

The Takeaway
The use of copyrightable works in the training and deployment of AI technology presents opportunities for innovation, as well as infringement. Balancing technological advancements with protecting artist and creator rights is manifest to cultivating a harmonious environment for tech companies and visual artists alike. We are sure to see more cases like Andersen v Stability AI. We hope each case will produce guidelines and guardrails for creators and tech entrepreneurs alike, and that, in the meantime, our blog post has enabled you to understand the technology and think critically on these issues.

For questions or to discuss, e-mail us at info@ambartlaw.com.

1. Andersen v. Stability AI Ltd., 3:23-cv-00201 (N.D. Cal. 2023).

2. Induced copyright infringement occurs when one knowingly encourages or influences another party to commit acts of copyright infringement. See Intentional Inducement of Copyright Infringements Act of 2004, U.S. Senate 108th Congress, 2d Session (Jul. 22, 2004), available at https://www.copyright.gov/docs/regstat072204.html.

3. Gaussian noise, also known as normal noise, is a type of signal noise that follows a normal distribution. It is created by adding random values to data and is characterized by a bell curve shape, with a mean of zero and a standard deviation of one. See “Gaussian Noise,” GeeksforGeeks (Jun. 17, 2024), available at https://www.geeksforgeeks.org/gaussian-noise/ (explaining the types of Gaussian noise and its normal distribution). 

4. “CMI” refers to information conveyed in connection with copies of a work, including in digital form, that conveys its creator, owner, or use. For a fantastic piece discussing whether the removal of CMI in AI-generated outputs constitutes a violation of the DMCA (and, if training data sets used to develop AI tools is actionable as copyright infringement under the DMCA) see Renata Mitchell, Generative AI, Copyright, and Book Publishing: The Impact of Generative AI on the Book Publishing Industry, 35  NYSBA Entertainment, Arts and Sports Law Journal (2024). 

5.  See Alice Hubbard, Weighing the Options: How Model Weights Can B Used to Fine-tune AI Models, Alliance for Trust in AI (last visited Nov. 4, 2024), available at https://alliancefortrustinai.org/how-model-weights-can-be-used-to-fine-tune-ai-models/. Stable Diffusion models are a new class of multimodal transformer models that combine “flow matching” (using weights) with “diffusion transformer architecture” (which possess strong scalability). See “Scalable Diffusion Models with Transformers”, arxiv.org (Mar. 2, 2023), available at https://arxiv.org/abs/2212.09748;. see also “Flow Matching for Generative Modeling,” arxiv.org (Feb. 8, 2023),  available at https://arxiv.org/abs/2210.02747.

6. This goes into whether or not the copyrighted work in question is sufficiently “stable” to give rise to a claim for infringement. See Flore Brunetti, Training of Generative AI Systems and Copyright Law: A U.S. and European Perspective, 35 N.Y.S.B.A. Entertainment, Arts and Sports Law Journal (2024) (discussing that under U.S. law, a copy “must be sufficiently permanent or stable to implicate the reproduction right,” although there is no “general rule defining how long a reproduction must endure to be ‘fixed’), available at https://nysba.org/publication/entertainment-arts-and-sports-law-journal/2024-vol-35-no-2/training-of-generative-ai-systems-and-copyright-law-a-u-s-and-european-perspective/