Introduction and Background
There is great interest in the question as to whether training and using AI models including generative AI models infringes copyright. Slowly, courts are winnowing out the plausible claims as the 26 or so cases wind through the courts in the United States. In an important decision, Judge William Orrick of the United States District Court for the Northern District of California partially granted and partially denied motions to dismiss in the case of Andersen et al. v. Stability AI Ltd. et al., a case that raises questions about copyright law‘s protection of copyright materials used for training and use of AI models.
The case involves a group of artists who accused Stability AI, Midjourney, DeviantArt, and Runway AI of unlawfully using their copyrighted works to train or use AI models, specifically the AI model known as Stable Diffusion. The decision builds on prior decisions where courts have started to outline legal theories that may be applicable. However, courts are still a long way away from resolving key questions as to which activities may be infringing and whether the fair use defense will insulate at least some activities from being copyright infringement.
A summary of the claims against the defendants is below.
Claims Against Stability AI
Stability AI is accused of directly infringing the plaintiffs’ copyrights by using their artworks without permission in the training of Stable Diffusion. The plaintiffs allege that Stability AI utilized large datasets, specifically the LAION-5B and LAION-400M datasets, which contain billions of images scraped from the internet, including the plaintiffs’ works. These images were allegedly used to train Stable Diffusion, enabling it to generate new images that replicate or closely resemble the style of the original copyrighted works. The plaintiffs further claim that Stability AI’s actions constitute not only direct infringement but also inducement, as the company allegedly distributed the Stable Diffusion models, knowing they would be used by others to create infringing works. Additionally, the plaintiffs accuse Stability AI of violating the Digital Millennium Copyright Act (DMCA) by removing or altering copyright management information (CMI) from their works during the training process.
Claims Against Midjourney
Midjourney is another key defendant in the case. The plaintiffs allege that Midjourney used the same datasets (LAION-5B and LAION-400M) to train its own AI model, which also allegedly generates images in the style of the plaintiffs’ works. The plaintiffs claim that Midjourney’s model incorporates elements of their copyrighted works without authorization, resulting in direct copyright infringement.
Claims Against Runway AI
Runway AI is a newer defendant in the litigation, added in the plaintiffs’ First Amended Complaint. The plaintiffs allege that Runway AI collaborated with Stability AI in the training of the Stable Diffusion model by contributing to the development of the AI’s capabilities. Specifically, Runway AI is accused of using the plaintiffs’ copyrighted works, contained in the LAION-5B dataset, to train its AI models. These models were then allegedly distributed through Runway AI’s platform, where users could generate new images that potentially replicate the plaintiffs’ styles, thus allegedly infringing their copyrights. Like the other defendants, Runway AI faces claims of direct copyright infringement, inducement of copyright infringement, and violations of the DMCA.
Claims Against DeviantArt
DeviantArt is alleged to have participated in the infringement by integrating Stable Diffusion into its own platform, offering a service called “DreamUp.” According to the plaintiffs, DeviantArt allowed users to create AI-generated images using the Stable Diffusion model, which had been trained on their copyrighted works. Although DeviantArt did not directly train the AI model, the plaintiffs argue that by providing this tool to its users, DeviantArt is liable for infringement.
The Court’s Decision
The decision of Judge Orrick follows prior decisions in which several claims against AI companies were allowed to proceed and other claims were struck, mostly with leave to amend. See, Barry Sookman, Generative AI litigation: the Github and Tremblay decisions, Barry Sookman, Resolving GenAI copyright infringement questions: 4 court decisions.
The prior ruling in this proceeding allowed the direct copyright infringement claim against Stability AI to proceed as it related to Stability AI’s alleged creation and use of “Training Images” scraped from the internet and included in the LAION datasets used to train the Stable Diffusion model. However, the court dismissed the plaintiffs’ second theory of direct infringement, which argued that Stable Diffusion was itself a “derivative work” containing compressed copies of billions of copyrighted images. The court also dismissed the plaintiffs’ claims under the DMCA, which accused the defendants of removing or altering copyright management information (CMI) from the plaintiffs’ works. The court also dismissed the unjust enrichment claims against all defendants, stating that these claims were preempted by the U.S. Copyright Act. The court’s order provided the plaintiffs leave to amend their complaint, allowing them to address the specific issues identified in the ruling.
Judge Orrick’s decision with respect to Stability AI, Midjourney, Runway AI, and DeviantArt, are summarized below, focusing on the copyright related claims.
Stability AI
Inducement and Distribution claims. The court allowed the claim of induced infringement to proceed. The plaintiffs alleged that Stable Diffusion is built to a significant extent on copyrighted works and that the way the product operates necessarily invokes copies or protected elements of those works. “The plausible inferences at this juncture are that Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate that infringement by design.” The court caveated its decision stating “[W]hether true and whether the result of a glitch (as Stability contends) or by design (plaintiffs’ contention) will be tested at a later date. The allegations of induced infringement are sufficient.”
The court also did not rule out that Stable Diffusion could be liable for distributing infringing works besides being liable for inducement stating:
Stability argues, first, that this theory is simply a repackaged direct infringement theory, that by “distributing” Stable Diffusion, Stability violates plaintiffs’ exclusive rights of distribution of their works. But whether this is a direct infringement claim (where liability is imposed against Stability for distributing copyrighted works) or more properly characterized as an inducement claim (where liability is imposed because Stability induces or otherwise causes others to copy protectible material) depends on how Stable Diffusion works and is implemented by users other than Stability itself. Any potential overlap – or potential requirement for plaintiffs to elect one claim or another – is better addressed on summary judgment, after discovery.
DMCA Claims. The plaintiffs alleged two primary violations under the DMCA.
False Copyright Management Information (CMI) Claim (17 U.S.C. § 1202(a)). The plaintiffs claimed that Stability AI falsely asserted copyright over the Stable Diffusion models by distributing them under an MIT License, which they argued constituted false CMI because the models allegedly contained elements derived from their copyrighted works.
The court dismissed this claim, finding that Stability AI’s MIT License was a generic assertion of rights over the software, not an explicit claim over the plaintiffs’ specific works. The court ruled that the license did not convey false CMI “in connection with” the plaintiffs’ works, as required under § 1202(a). This claim was dismissed with prejudice.
Removal or Alteration of CMI (17 U.S.C. § 1202(b)(1)). The plaintiffs also alleged that Stability AI removed or altered CMI, such as watermarks and signatures, from their works during the training process for Stable Diffusion. The court dismissed this claim as well, agreeing with the reasoning in a recent case (Doe 1 v. GitHub Inc.) that DMCA claims require the removal of CMI from identical copies of a work. This claim was also dismissed with prejudice.
Unjust Enrichment Claim. The plaintiffs brought a claim of unjust enrichment against Stability AI, alleging that the company unjustly benefited from the use of their copyrighted works to train the Stable Diffusion model without compensation. They argued that Stability AI misappropriated their works to develop, train, and promote its AI models, which allowed the company to reap financial and reputational benefits at the plaintiffs’ expense.
The primary issue the court addressed was whether the unjust enrichment claim was preempted by the U.S. Copyright Act. Stability AI argued that the claim was preempted because it was based on the same underlying conduct as the copyright infringement claims—the use of the plaintiffs’ copyrighted works without authorization. The court agreed, noting that the unjust enrichment claim did not involve an “extra element” beyond the rights protected by the Copyright Act.
Although the court dismissed the unjust enrichment claim, it did so with leave to amend, providing the plaintiffs an opportunity to potentially reformulate their claim.
Midjourney
The plaintiffs alleged that Midjourney directly infringed their copyrights by using their works, included in the LAION datasets, to train its AI model. Midjourney’s AI tool, which generates images based on user prompts, allegedly incorporated and replicated elements of the plaintiffs’ copyrighted works without authorization.
The court noted in its October 2023 ruling, that it had required the plaintiffs to clarify how Midjourney specifically used their works in training its AI model. In response, the plaintiffs’ First Amended Complaint included more detailed allegations, asserting that Midjourney used the LAION-400M and LAION-5B datasets—both containing billions of images, including the plaintiffs’ copyrighted works—to train its AI models. The court found these allegations sufficient at this stage to support the claim that Midjourney directly infringed the plaintiffs’ copyrights by incorporating their works into its training process.
The plaintiffs also provided examples where their names were used as prompts in the Midjourney AI tool, resulting in outputs that allegedly mimicked their artistic styles. The court found that these examples helped demonstrate the plausibility of the plaintiffs’ claim that their works were not only used in the training process but that the resulting AI models retained and reproduced elements of those works. Although the court recognized that proving these claims would require substantial evidence, it determined that the allegations were sufficient to survive a motion to dismiss.
Runway AI
Runway AI faced several claims including direct copyright infringement, inducement of copyright infringement, violations of DMCA, and unjust enrichment.
The court’s analysis of the copyright claims focused on the specific role Runway AI allegedly played in developing, training, and distributing the Stable Diffusion model and its AI tools.
Direct Infringement. The plaintiffs alleged that Runway AI directly infringed their copyrights by using their works in the LAION datasets to train the Stable Diffusion model, specifically version 1.5, which Runway AI allegedly helped develop and distribute.
The court examined the plaintiffs’ claims that Runway AI actively participated in the creation and training of the Stable Diffusion 1.5 model using the LAION datasets, which allegedly contained copies of the plaintiffs’ works. The plaintiffs argued that by using these datasets, Runway AI directly infringed their copyrights. The court found that the plaintiffs provided sufficient factual detail to support their claim that Runway AI was involved in the development and training process, thus making the direct infringement claim plausible.
Similar to the claims against Stability AI and Midjourney, the plaintiffs advanced the “Model Theory” (asserting that the Stable Diffusion 1.5 model itself was an infringing work) and the “Distribution Theory” (claiming that distributing Stable Diffusion was infringing). The court allowed these theories to proceed, acknowledging that the detailed allegations regarding Runway AI’s role in training and distributing the AI model were sufficient to support these claims at the pleading stage. The court noted that these claims would require further factual development to determine their viability.
According to the court:
As with Stability, because Runway does not challenge the use of the images for training purposes, I need not address the other theories of direct infringement. However, I note that both the model theory and the distribution theory of direct infringement depend on whether plaintiffs’ protected works are contained, in some manner, in Stable Diffusion as distributed and operated. That these works may be contained in Stable Diffusion as algorithmic or mathematical representations – and are therefore fixed in a different medium than they may have originally been produced in – is not an impediment to the claim at this juncture. 1 Nimmer on Copyright § 2.09[D][1] (2024) (“A work is no less a motion picture (or other audiovisual work) whether the images are embodied in a videotape, videodisc, or any other tangible form.”).
Inducement of Copyright. The inducement claims against Runway AI were allowed to proceed. The plaintiffs allege that Runway AI helped train and develop Stable Diffusion, and therefore, knew that the product allegedly uses or invokes the training images in its operation. According to the court, those allegations, combined with allegations that Runway actively induces others to download Stable Diffusion by distributing Stable Diffusion through popular coding websites and also by selling its products (including AI Magic Tools) that include Stable Diffusion, were sufficient to meet the minimal requirements for a plausible pleading.
DeviantArt
Direct Copyright Infringement. The plaintiffs alleged that DeviantArt directly infringed their copyrights by incorporating Stable Diffusion, trained on their copyrighted works, into its DreamUp tool. This tool allowed DeviantArt users to generate AI-created images, which the plaintiffs claimed were based on their original works.
Unlike Stability AI and Runway AI, DeviantArt was not directly involved in training the Stable Diffusion model. But, they argued that DeviantArt’s use of the Stable Diffusion model, which had been trained on their copyrighted works, constituted direct infringement. The court dismissed DeviantArt’s motion to dismiss the copyright claim, but did not engage in any precise analysis of the issues. The court did state then following:
The FAC alleges only that DeviantArt incorporates and relies on Stable Diffusion for its DreamUp product. Plaintiffs have added allegations to their FAC, however, regarding how copies or protected elements of their works remain, in some format, in Stable Diffusion and how those works can be invoked by use of all of the Stable Diffusion versions…
Moreover, while plaintiffs admittedly did not include assertions or examples in the FAC of outputs from DreamUp that appear to copy elements of their works (as they did for the other defendants), their added allegations regarding the use of the LAION datasets to train Stable Diffusion and how Stable Diffusion versions operates, including specific examples and academic references to the operation of Stable Diffusion 1.4 used by DreamUp, suffice.
The court also relied on this pleading by the plaintiffs:
On information and belief, by the end of training, Stable Diffusion 1.4 was capable of reproducing protected expression from each of the LAION-5B Registered Works that was in each case substantially similar to that registered work, because— a. In the Carlini Paper, Nicholas Carlini tested Stable Diffusion 1.4 and found that it could emit stored copies of its training images; b. The training procedure for Stable Diffusion 1.4 was very similar to that of Stable Diffusion 1.5, which was shown in Exhibit E: Runway text prompts and Exhibit H: Runway image prompts to be capable of emitting stored copies of protected expression. See also ¶ 394 (“Therefore, like Stable Diffusion 1.5, Stable Diffusion 1.4 also qualifies as an infringing Statutory Copy of the LAION-5B Registered Works. Because Stable Diffusion 1.4 represents a transformation of the LAION-5B Registered Works into an alternative form, Stable Diffusion 1.4 also qualifies as an infringing Statutory Derivative Work.”).
Unjust Enrichment. The plaintiffs further alleged that DeviantArt was unjustly enriched by allowing the use of their copyrighted works to develop and promote the DreamUp tool. They argued that DeviantArt benefited financially from the unauthorized use of their works without providing them any compensation. This claim was dismissed as being pre-empted by the Copyright Act, with leave to amend.
Comments on Judge Orrick’s Order
The decision of Judge Orrick is an important one in that it helps to further refine the copyright and related theories that may be asserted in the case, and presumably in other cases. But, in understanding the scope of the decision, it must be borne in mind that the findings are based on the low standards for pleadings that examines whether the pleading states a plausible claim in law. To succeed on a summary judgement motion or at a trial, the plaintiffs will need to adduce evidence to support the claims. Further, the decision does not address the merits of any possible fair use defense.
Subject to these caveats, Judge Orrick’s decision confirms that the following are plausible copyright infringement theories against one or more of the defendants:
- Copying images for training an AI system may be infringement.
- A trained AI system may be an infringing work. It may contain copies of training images, and even if these are stored as algorithmic or mathematical representations – and are therefore fixed in a different medium than they may have originally been produced, this is not a bar to an infringement.
- The distribution and use of an infringing AI system may implicate the distribution right and may be an inducement to infringe.
