Top Takeaways from Order in the Andersen v. Stability AI Copyright Case
On August 12, an order granting in part and denying in part motions to dismiss a first amended complaint was issued by Judge William Orrick (in the Northern District of California) in Andersen v. Stability AI. While the order is just one early step in a case that still has a long way to go until a final outcome, it provides some insights into how courts may treat claims of direct, input-side copyright infringement, induced infringement, as well as claims for violation of the DMCA for removal and/or alteration of copyright management information. Perhaps most significantly, the order makes clear that generative AI models are unlike any technologies at the center of major copyright disputes in the past and questions regarding how AI models operate and whether they infringe will differ from case to case depending on the model and the types of works ingested.
Background
In early 2023, visual artists Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action lawsuit against Stability AI, Midjourney, and DeviantArt in the Northern District of California, alleging copyright infringement, right of publicity violations, and other claims related to the use of the plaintiffs’ works in training data sets for the AI image-generating platforms Stable Diffusion, the Midjourney Product, DreamStudio, and DreamUp. It was the first lawsuit of its kind, because while there had been suits brought by copyright owners against AI developers in the past, the Andersen complaint involved creators coming together to challenge the unauthorized ingestion of their works by a generative AI company. Many more lawsuits have followed, and there are now over a dozen similar class action lawsuits brought by authors and visual artists against a number of different AI companies.
On October 30, 2023, the district court largely granted motions to dismiss by the defendants but allowed the direct infringement claims to move forward and allowed plaintiffs leave to amend. Then, after a first amended complaint was filed, on August 12, 2024, the court issued an order granting defendants’ motions to dismiss 1202 DMCA claims with prejudice but denying the motions to dismiss trademark, direct copyright infringement, and inducement claims. The order is a bit complex, mostly because the court is addressing motions to dismiss by multiple defendants, but there are a few significant takeaways that show how courts the Northern District of California (and possibly elsewhere) will address generative AI companies’ defense arguments.
Top Takeaways
(1) Arguments that Generative AI is Similar to Past Technologies, Like the VCR, Aren’t Going to Hold Much Weight
A claim against Stability AI for induced copyright infringement was added by the plaintiffs in their first amended complaint, and it accuses the company of inducing infringement by distributing its infringing Stable Diffusion models to third parties. The allegation is supported by statements by Stability’s CEO that the company took 100,000 gigabytes of images and “compressed” them down into a file that can “recreate” any of the images. While Stability challenges the meaning and intent of the statement, Judge Orrick found that whether the company is liable for inducement “depends on how Stable Diffusion works” and “is better addressed on summary judgment, after discovery.”
Judge Orrick then makes a critical distinction that addresses arguments made by many AI companies that liken their technologies to the VCR or other devices from the past that were found to be non-infringing. He explains that the theory of the case is not similar to past contributory infringement cases involving the sale of VCRs where plaintiffs, after discovery had no evidence of defendant’s intent to induce infringement.
“Instead, this is a case where plaintiffs allege 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.”
This is a key distinction—that unlike present generative AI technologies, devices like VCRs were not pre-loaded with and did not ingest and copy protected works before reaching the consumer. They were devices that were capable of copying but had substantial non-infringing uses. For that reason, Judge Orrick found the allegations of inducement sufficient to survive the motion to dismiss.
(2) Copyright Management Information Claims May Be Tough to Prove
Stability also moved to dismiss plaintiffs’ claims under 1202(a) of the Digital Millennium Copyright Act (DMCA) for providing or distributing false copyright management information (CMI). Plaintiffs’ argument is based on the theory that Stability distributes its models under a license that asserts copyright in the model, and that because the models are infringing, Stability is providing and distributing false CMI. The court agreed with Stability that its “generic” license does not suggest any association with plaintiffs’ works and that the 1202(a) claims fails the “double scienter” requirement of alleging Stability knowingly provided false CMI with intent to induce or enable infringement. Thus, Judge Orrick dismissed the claim with prejudice.
Moving to dismiss a claim under 1202(b) for the removal or alteration of CMI, Stability cites to a recent decision in Doe v. Github, a case brought by computer programmers against the software development platform for scraping their code to train its AI tool, GitHub Copilot. In that case, the judge dismissed 1202(b) claims because there were no outputs that were identical to the plaintiffs’ works. This “identicality” requirement, while not in the Copyright Act, has been applied by some courts, including district courts in the Ninth Circuit. Judge Orrick, while recognizing the “issue is unsettled,” agreed with the Doe v. Github decision and dismissed the 1202(b) claim with prejudice because there were no allegations that any Stable Diffusion output was identical to plaintiffs’ works.
The final dismissal of all CMI claims in the Andersen case demonstrates a problem faced by plaintiffs in many of the AI infringement cases. Namely, that many have not provided evidence of identical output that would support a 1202(b) claim. Additionally, the “double scienter” requirement is one that has stymied copyright owners for years, given that proving knowledge and intent is very difficult without smoking gun evidence. While some CMI claims are surviving, they are confined to cases like the one Getty Images brought against Stability where they have shown outputs that, while not identical, contain versions of Getty’s watermark.
(3) Assertions by AI Companies That They Are Just Copying Unprotected Data Don’t Hold Up
Runway AI, which was added as a defendant in the first amended complaint, moved to dismiss two direct infringement claims. One is what Judge Orrick calls the “model theory,” which is based on the allegation that Stability’s product (which Runway incorporated into its own generative AI model) is itself an infringing work. The second is what he terms the “distribution theory,” which is based on allegations that Runway infringes plaintiffs’ exclusive distribution rights by distributing Stability’s infringing product.
Judge Orrick explains that both theories depend on whether plaintiffs’ works are “contained, in some manner” in the Stable Diffusion model. He adds 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.” This is a significant point that could be pivotal in many generative AI cases, as many AI companies argue that they only copy unprotectable “data” that exist as statistical representations in a model. If courts find that these statistical representations are simply a different medium in which expressive elements are fixed, it would be a huge win for copyright owners and cast serious doubt on arguments of noninfringement and fair use.
(4) Decisions in Past Cases—Even Other AI Cases—Are Likely Not To Be Influential
A final point from the court’s order worth mentioning is that it goes out its way to explain that (1) “run of the mill” copyright cases relied upon by the defendants are unhelpful because of the unique nature of generative AI, and (2) even other current generative AI cases are not particularly helpful because of the vast differences in the operation of different generative AI models and what types of copyrighted works are ingested. As to the first point, Judge Orrick notes that Runway’s and other defendants’ reliance on cases where a showing of substantial similarity between works is required when determining whether an inference of copying can be supported is unhelpful “in this case where the copyrighted works themselves are alleged to have not only been used to train the AI models but also invoked in their operation.”
On the second point, Judge Orrick addresses Runway’s reliance on Kadrey v. Meta—a book authors’ class action infringement case—for the position that derivative infringement theories should be rejected. He explains that Kadrey deals with Large Language Models (LLMs), and that the Judge in that case dismissed the claims because “there were no allegations that the outputs of those models could be ‘understood as recasting, transforming, or adapting the plaintiffs’ books’ or otherwise producing outputs that were ‘substantially similar’ to aspects of plaintiffs’ books.” Judge Orrick writes:
“The products at issue here – image generators allegedly trained on, relying on, and perhaps able to invoke copyrighted images – and the necessary allegations regarding the products’ training and operations, are materially different from those in Kadrey. Whether substantial similarity remains a hurdle to specific theories – including any derivative infringement theory – depends in part on what the evidence shows concerning how these products operate and, presumably, whether and what the products can produce substantially similar outputs.”
What Judge Orrick is getting at is similar to something that we at the Copyright Alliance have been stressing when considering the many issues surrounding generative AI. That is, generative AI impacts each copyright industry differently because they operate using different business models, provisions in copyright law apply to them differently, and most important, the AI models that impact them are different. As the order points out, LLMs operate differently than image generators, and they both operate differently than music generator models. It’s tempting to generalize and compare the different models and infringement cases, but each involves a very unique set of facts that must be considered independently. To the extent that some cases involve similar plaintiffs, defendants, and facts, those cases have been consolidated.
Looking Ahead
The order on motions to dismiss in Andersen is just one decision in a case that will likely have many more twists and turns, and it shouldn’t be read as an indication of which parties will prevail. What it does provide is an insight into how a court in a district that is no stranger to copyright disputes will approach and analyze issues surrounding the operation of generative AI models and infringement. But ultimately, as Judge Orrick explains (and we’ve noted numerous times), each case is unique and involves different facts, which will be especially important to recognize when the fair use elephant in the room is litigated.
If you aren’t already a member of the Copyright Alliance, you can join today by completing our Individual Creator Members membership form! Members gain access to monthly newsletters, educational webinars, and so much more — all for free!