INTRODUCTION
Sarah Anderson, Kelly McKernan, and Karla Ortiz, all artists, had sued Stability AI Ltd., DeviantArt, Inc., and Midjourney, Inc., claiming that their copyrighted works were unlawfully used to train AI platforms without permission.
On 30th October, 2023, the District Court of the United States ruled that the Plaintiffs’ copyright claims could proceed, but limited them to works registered with the Copyright Office. The Court found the allegations against Stability AI plausible, but required Plaintiffs to clarify their claims against DeviantArt and Midjourney. Various claims such as vicarious infringement, DMCA violations, right of publicity, breach of contract, etc., were dismissed with leave given to the Plaintiffs to amend for further specificity.[i]
Recently a new development came in this case on 12th August 2024 in this case of Sarah Andersen v. Stability AI Ltd.,[ii] which will be analysed in this article. Read decision here .
FACTS
Artists Sarah Andersen, Kelly McKernan, Karla Ortiz, Hawke Southworth, Grzegorz Rutkowski, Gregory Manchess, Gerald Brom, Jingna Zhang, Julia Kaye, and Adam Ellis were the Plaintiffs.
They filed a putative class action lawsuit on behalf of artists challenging the defendants’ creation and/or use of Stable Diffusion, an artificial intelligence software product which makes images from text input. They alleged that Stable Diffusion used Plaintiffs’ artistic works as “training images” and as a result Stable Diffusion could produce output images “in the style” of the training images.
Defendants in this case, as follows: Stability AI Ltd. and Stability AI, Inc. (collectively “Stability AI”), Midjourney, Inc., DeviantArt, Inc., and Runway AI, Inc. These defendants requested to the Court, the dismissal of various claims made by the Plaintiffs.
CLAIMS
The Plaintiffs alleged that five billion images were scraped to create the LAION training sets, which were used by Stability, Runway, and Midjourney to train Stable Diffusion models. These defendants were accused of copying or using versions of the Plaintiffs’ artistic works in their AI products. The Plaintiffs emphasize that the LAION-5B dataset contains only URLs of images, requiring users to download the actual images using tools like img2dataset.
Note: LAION-5B is a dataset that contains 5.85 billion images along with accompanying text, 14x bigger than LAION-400M, previously the biggest openly accessible image-text dataset in the world, generally used to train AI and machine learning technology products.[iii]
On October 2023, the Court had largely dismissed the claims filed by the Plaintiffs. The only claims allowed were against Stability AI. The remaining claims were dismissed with leave to amend, allowing the Plaintiffs to clarify their theories and provide plausible supporting facts.
The Court had also previously rejected Plaintiffs’ claim that Stable Diffusion is a “derivative work” because it supposedly includes compressed copies of copyrighted images and transforms them into new works, finding the argument didn’t meet the necessary legal standards.
Subsequently, the Plaintiffs added defendant Runway AI and seven new Plaintiffs in their amended complaint, but each defendant filed motions to dismiss.
ANALYSIS BY THE COURT
A. Legal Standard
The District Court in United States has to dismiss a complaint if it fails to state a claim for which relief can be granted. To survive this, the Plaintiff has to present sufficient facts to make the claim plausible on its face, as established in Bell Atlantic Corp. v. Twombly.[iv] A claim is plausible when the facts allow the Court to reasonably infer that the defendant is liable. The Plaintiffs must provide more than a speculative possibility of wrongdoing, although specific fact pleading is not required.
If the Court dismisses the complaint, it typically grants leave to amend unless the complaint’s defects cannot be remedied. Factors like undue delay, bad faith, repeated failures to address deficiencies, prejudice to the opposing party, and the futility of the amendment are considered.
B. Addition of New Plaintiffs
Stability AI and DeviantArt objected to the inclusion of seven new Plaintiffs and two new claims in the First Amended Complaint, arguing that these exceeded the scope of the October 2023 order’s leave to amend. The Plaintiffs countered that broad leave was granted to address deficiencies and that Rule 15 permits such additions. The Court acknowledged that new Plaintiffs and claims generally require explicit approval or consent. But the Court in any case assumed an implicit request for leave from the complaint, and therefore permitted the new additions.
C. Stability AI’s motion to dismiss
1. Induced Copyright Infringement – Allowed
Under the theory given by the Plaintiffs, Stability AI was inducing infringement by distributing the models when any third party downloaded, used, or deployed the models provided by Stability AI.
According to Stability AI, the induced copyright infringement claim should be dismissed because the Plaintiffs failed to allege that Stability AI encouraged the use of Stable Diffusion to create infringing outputs. Citing Metro-Goldwyn-Mayer Studios Inc. v. Grokster, Ltd.,[v] Stability AI asserted that the Plaintiffs must show “clear expression or other affirmative steps” taken by Stability AI to foster infringement, which they have not done.
Plaintiffs highlighted a statement by Stability’s CEO, who mentioned that Stability AI compressed 100,000 gigabytes of images into a two-gigabyte file that can “recreate” any of those images. Stability dismisses this as an isolated statement, arguing that it does not demonstrate an intent to promote infringement. Stability also pointed out that Stable Diffusion has a lot of non-infringing uses, where its AI was creating original art that did not reference or invoke any of the Plaintiffs’ works.
The Court further noted that here, the Plaintiffs alleged that Stable Diffusion is built on copyrighted works and operates in a way that necessarily invokes those works, potentially leading to copyright infringement by end users. The Plaintiffs referenced academic articles suggesting that training images could sometimes be reproduced as outputs from AI products.
The Court concluded that the Plaintiffs sufficiently alleged a claim of induced infringement allowing it to proceed, and whether this claim held true or resulted from a glitch (as Stability contends) or by design (as the Plaintiffs argue) would be determined later on in the case.
2. Claims regarding copyright management information (CMI) – Dismissed
- Claim Under 1202(a)
The Plaintiffs alleged that Stability AI had distributed its AI model with copyright management information (CMI) that asserted copyright over the Stability AI Models but failed to acknowledge the Plaintiffs’ copyrighted works, which were included in those models. The Plaintiffs argued that this constituted the distribution of false CMI, as Stability AI’s CMI claimed copyright over the entire model without referencing the Plaintiffs’ copyrighted images that had been incorporated into it.
Stability AI contended that this claim failed because the generic license accompanying the use of Stable Diffusion did not suggest any association with the Plaintiffs’ works. Stability AI argued that the license would not have led a reasonable viewer to believe that Stability AI was claiming rights to or conveying false information about the Plaintiffs’ works.
The Court agreed with Stability AI, noting that the generic license claimed rights to Stable Diffusion as a work, not to the underlying datasets or any specific works used to create them. It was implausible that the license would have been interpreted as conveying any false information about the Plaintiffs’ copyrights. Consequently, this claim was dismissed.
- Claim Under 1202(b)(1)
The Plaintiffs claimed that Stability AI copied works from the LAION-5B dataset, which contained CMI such as watermarks, signatures, and captions. They alleged that the training process intentionally removed or altered this CMI, which would constitute a violation of 17 U.S.C. § 1202(b)(1).
The Court agreed with Stability AI’s position, noting that the Plaintiffs did not allege that any output from Stable Diffusion was identical to their original works. Since identicality is required for a § 1202(b) claim according to the Court, the Plaintiffs’ claim could not succeed. The Court dismissed the Plaintiffs’ DMCA claims with prejudice.
3. Unjust Enrichment – Dismissed
The Plaintiffs claimed that Stability AI unjustly benefited by using their copyrighted works, specifically those in the LAION-5B dataset, to develop, train, and promote the Stability Models, thereby depriving the Plaintiffs of the value of their works. Stability AI argued that this claim is essentially based on the unauthorized use of the Plaintiffs’ copyrighted works, which falls under the Copyright Act and is therefore pre-empted.
The Plaintiffs countered that their unjust enrichment claim focused not just on the use of their works but also on the exploitation of their “artistic personas” and the mimicking of their artistic style through Stability’s AI models.
However, the Court noted that this argument was not reflected in the Plaintiffs’ First Amended Complaint (FAC), which tied the unjust enrichment claim directly to the use of their copyrighted works, making it subject to pre-emption. As a result, the Court dismissed the unjust enrichment claim on the grounds that it was pre-empted by the Copyright Act, but allowed the Plaintiffs leave to amend the claim.
D. Runaway AI’s Motion to Dismiss
1. Direct Copyright Infringement – Allowed
Runway challenged two theories of direct infringement asserted by the Plaintiffs: the “Model Theory” and the “Distribution Theory.”
- Model Theory:
This theory posits that the AI product itself, after training, constitutes an infringing “Statutory Copy” or “Statutory Derivative Work” because it embodies transformations of the Plaintiffs’ copyrighted works.
- Distribution Theory:
This theory asserts that Runway AI infringes on the Plaintiffs’ exclusive distribution rights by distributing their AI product, which the Plaintiffs claim is equivalent to distributing their copyrighted works.
The Court found that these theories depend on whether the Plaintiffs’ protected works are contained within Stable AI’s product in some form. The fact that these works might be embedded as algorithmic or mathematical representations does not preclude the claims at this stage. It was claimed earlier by the Plaintiffs that the training images remain in and are used by Stable Diffusion, relying on statements from Stability’s CEO and academic papers supporting their claims.
As a result, direct infringement claims were allowed, noting that further evidence and possible decisions between competing theories would be addressed at the summary judgment stage.
2. Induced Infringement – Allowed
Runway AI also sought to dismiss the claim of induced infringement, arguing that the Plaintiffs failed to allege that third parties used Runway’s products to commit acts of infringement and that Runway promoted such infringing use. The Court noted that the Plaintiffs alleged Runway’s role in training and developing Stable Diffusion, implying that Runway knew about the product’s potential to use training images in ways that could infringe copyrights. These allegations were deemed sufficient to sustain the claim of induced infringement.
3. DMCA and Unjust Enrichment/UCL Claims – Dismissed
As with Stability AI, the Court dismissed the Plaintiffs’ DMCA claims against Runway AI with prejudice, as the issues raised were similar to those previously addressed. The unjust enrichment and UCL claims were also dismissed but with leave to amend.
E. Midjourney’s Motion to Dismiss
Midjourney moved to dismiss copyright claims, and LANHAM Act claims for false endorsement and trade dress claims asserted against it by the Plaintiffs.
1. Copyright Claim – Allowed
The Plaintiffs alleged that Midjourney trained its product on the LAION400M and LAION5B datasets and incorporated Stable Diffusion into its AI product. These allegations, combined with the claim that Midjourney’s product produces images similar to the Plaintiffs’ artistic works when their names are used as prompts, were deemed sufficient to cross the threshold of plausibility.
Midjourney argued that these examples were inadequate because some of the identified works were unregistered, and the outputs could have been generated from training on unregistered works or non-protected elements. The Court clarified that the Plaintiffs were not using these examples to establish copyright infringement as a matter of law but rather to support the plausibility of their copyright theories.
The Court concluded that the FAC and the exhibits met the plausibility standard, and the validity of the claims would be determined through evidence at the summary judgment stage. Therefore, the copyright claim was allowed.
2. Lanham Act (Trademark Act of the USA) – Allowed
The Plaintiffs asserted claims under the Lanham Act (Trademarks Act in the USA), alleging false endorsement and trade dress violations. Plaintiffs claimed that their names were listed among 4,700 artists by Midjourney’s CEO on Discord, a platform where Midjourney’s AI product operates. This list was promoted as showcasing the artistic styles that the AI could replicate. The Plaintiffs also alleged that Midjourney used user-created images incorporating the Plaintiffs’ names in its showcase site.
- False Endorsement:
To succeed on their Lanham Act claim for false endorsement, the Plaintiffs needed to show they owned a protectable interest in the mark and that the defendant’s use of the mark was likely to cause consumer confusion. Midjourney argued that the Names List uploaded by its CEO didn’t imply endorsement by the artists and referenced to the comment that those names were sourced from Wikipedia. The Court ruled that it was premature to dismiss the claim, as the Plaintiffs plausibly alleged that the use of their names could confuse consumers suggesting endorsement.
- Vicarious Trade Dress:
The Plaintiffs claimed that Midjourney’s AI product, which used a model trained on their works, allowed users to recreate the unique “look and feel” of their artwork, thus infringing their trade dress. Midjourney argued that the Plaintiffs hadn’t clearly identified the specific elements of their trade dress and that these elements were too broad. However, the Court found that, despite some vagueness, the combination of the elements and the way the AI model functioned provided enough detail to support the Plaintiffs’ claim.
The Court also found that the Plaintiffs had adequately claimed that their trade dress had acquired a secondary meaning, linking their art with them in the minds of consumers. Additionally, the Court upheld the vicarious trade dress claim, stating that the Plaintiffs had sufficiently alleged Midjourney’s control over the infringing images through its AI model.
Therefore, the Lanham Act claims of the Plaintiff were allowed and not dismissed by the Court.
F. Deviant Art’s Motion to dismiss
1. Copyright Claim
DeviantArt sought to dismiss the copyright claims against it by arguing that it merely provided AI tools like Stable Diffusion and did not train any AI models. DeviantArt claimed that holding it liable could set a dangerous precedent, affecting millions of users who utilize similar tools.
The Plaintiffs had expanded their allegations from those made previously, asserting that Stable Diffusion retained copies or protected elements of their works, which could be invoked by any version of the tool. The Court found these allegations plausible enough to keep the copyright claims against DeviantArt alive.
While DeviantArt argued that its use of the Plaintiffs’ works should qualify as fair use, the Court determined that this defense and the extent of potential infringement needed further examination, leading to the denial of DeviantArt’s motion to dismiss the claims.
2. Breach of Contract Claim
In the amended complaint, the Plaintiffs reasserted the claim that was dismissed previously by the Court. Therefore the Court again noted there were no allegations that DeviantArt allowed image scraping or violated any Terms of Services provisions by using Stable Diffusion. The Court rejected the Plaintiffs’ argument that DeviantArt’s use of Stable Diffusion exceeded the Terms of Services of DeviantArt.
The Plaintiffs also added a claim for breach of the implied covenant of good faith and fair dealing, arguing that DeviantArt was obligated to protect users’ works. However, the Court dismissed this claim as well, stating that the Plaintiffs did not identify any specific contractual provision that DeviantArt had violated.
Since the Plaintiffs did not contest the Court’s tentative ruling or offer any new facts to support their claim, the breach of contract claim was dismissed with prejudice.
CONCLUSION
The ruling has allowed various claims with regard to Copyright Act such as direct infringement and induced infringement, which shows the trend of Courts, to protect valid copyrights that are owned by artists and right holders, in the face of constant large scale infringements being carried out by Artificial Intelligence companies without any risk of consequences.
Even by admitting claims of the Plaintiff, there is a risk now of losing the case, which will in turn disincentivize AI Companies from venturing into using copyrighted works without the permission of the appropriate right holders.
The arguments with regard to the use of likeness of the Artists each time an AI model creates a work with the look and feel of an Artist seems plausible. There are various images that can be made in the style of these Artists, using their name, such as if I used the prompt Jackson Pollock into an AI model, it will create a similar image as that of the style particularly used by him.
A big deterrent therefore has been established, in this area of text to image AI generative tools, and the effect might be seen throughout the World as foreign jurisdictions including India also start allowing claims by Artists giving them an opportunity to fight against the mass infringement being done by AI companies.
End notes:
[i] https://www.loeb.com/en/insights/publications/2023/11/andersen-v-stability-ai-ltd.
[ii] Case No. 23-cv-00201-WHO.
[iii] https://laion.ai/blog/laion-5b/.
[iv] 550 U.S. 544, 570 (2007).
[v] 545 U.S. 913, 919 (2005).
Image created on Dall-E