Docket

Andersen v. Stability provides an early glimpse into AI lawsuits

The first generative AI decisions are starting to be released, giving us an early look into how courts understand these products. Last week in Andersen v. Stability, a judge in the Northern District of California denied in part and granted in part motions to dismiss by defendant AI companies Stability AI, Midjourney, DeviantArt, and Runway AI, in a suit by artists alleging various common law and copyright violations due to the their works being used as “training images” in the defendants’ models.

At the heart of the complaint is:

The creation of the LAION training sets, where five billion images were allegedly scraped into datasets used by Stability and Runway to train the versions of Stable Diffusion. Plaintiffs state that Midjourney likewise trained its product using Stable Diffusion, and that all four defendants use Stable Diffusion in their AI products; in doing so, those four defendants copy or utilize versions of plaintiffs’ artistic works.  Significantly, plaintiffs allege that the “LAION-5B dataset contains only URLs of training images, not the actual training images. Therefore, anyone who wishes to use LAION-5B for training their own machine learning model must first acquire copies of the actual training images from their URLs using the img2dataset or other similar tool.” They also clarify their theory of direct infringement, adding allegations regarding CLIP-guided diffusion in the training phase but also in use, after training.

Plaintiffs’ first complaint was filed in October 2023, and a first motion to dismiss left only a direct copyright infringement claim standing (with leave to file an amended complaint). The First Amended Complaint brought claims for unjust enrichment and breach of contract, as well as claims under the DMCA, Lanham, Act, and Copyright Act. Ruling on the second set of motions to dismiss, the court dismissed the common law and DMCA claims, and left alive the Copyright and Lanham Act claims.

We thought that the copyright infringement claims, and the court’s willingness to let both direct and induced copyright infringement claims proceed, was worth calling out as it gives a good insight into the ways courts understand claims around generative AI products, and shows the level of depth at which AI companies will be examined.

Direct v. Induced Copyright Infringement

In their motion to dismiss, Stability first argued that plaintiffs’ induced infringement claims were simply a repackaged version of the direct infringement claims. As explained by the court, a direct infringement claim would impose liability on Stability for distributing the plaintiff artists’ copyrighted works by, for example, having its model produce works that looked almost exactly like the plaintiffs’ works. An inducement claim, on the other hand, would find liability for Stability if they were found to have induced others to copy protected material. In other words, there could be a finding of inducement if Stability encouraged users to recreate the plaintiffs’ copyrighted works with their AI product.

The court declined to find that the inducement claims were merely a repackaged version of the direct infringement claims, holding that any overlap between the two types of claims would ultimately depend on how Stability actually works, and how users use Stability. Because these questions cannot be answered without a full review of the factual record in discovery, the court denied Stability’s motion to dismiss on this point.  

The court next addressed Stability’s argument that the inducement claims must fail because plaintiffs did not allege any facts that Stability had any intent to promote infringement by its users. In rejecting this argument the court fist pointed to a statement by Stability’s CEO, and quoted by plaintiffs in their complaint, that Stability took 100,000 gigabytes of images and compressed them into a file that can “recreate” any of its images.

The court then differentiated Stability’s case from a Supreme Court case in which plaintiffs had alleged that selling VCR’s constituted inducement as VCR players could easily be used to copy and distribute copyrighted works. In that case, the court ruled that plaintiffs had no evidence of a defendant’s intent to induce infringement simply by selling VCRs, because intent could not be “based on presuming or imputing intent to cause infringement solely from the design or distribution of a product capable of substantial lawful use, which the distributed knows is in fact used for infringement.”

For the current case, the court ruled that not only did plaintiffs allege Stability knew that its product could be used for infringement, but also that the “plausible inference at this juncture” is that “Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate infringement by design.” When combined with the CEO’s statements as well as articles by academics on how training images can reproduce copyrighted works, the Court found that plaintiffs had adequately pled a claim for inducement.

It’s still too early to say what exactly this all means for AI and copyright law going forward, but one thing that is clear is that AI companies are not operating in some wild west, completely out of the reach of the law. As we’ve written before, the courts act as a powerful regulatory force constraining AI developer behavior by thoroughly examining their products for legal compliance. While this might not be and probably isn’t the optimal form of regulation, it does mean that legislators do not need to rush to create regulations simply for the sake of regulations.  

Additional reading: the motion to dismiss

If you’re a lawyer this is probably the end of the interesting part of the article for you. If you aren’t a lawyer, you may be asking yourself things like wait, why is the court drawing inferences in favor of the plaintiff? I thought things were supposed to be neutral?

Well lucky for you, we wrote a little bit more on how, and why, courts rule the way they do when they take on a motion to dismiss.

A motion to dismiss occurs at the beginning of a case, before the parties have engaged in discovery and therefore before there’s much of a factual record. At this early stage, all the plaintiff has to rely on to make their case are things they were able to learn themselves, or facts they have a good reason to believe to be true. One of the main goals of discovery is allowing the plaintiff a fair shot to prove why they were harmed, provided they can make plausible enough allegations at the motion to dismiss stage to get there. Otherwise, defendants may be able to flout the law because they hold all the necessary evidence, like internal documents or communications, to hold them liable.

This lack of a factual record informs the standard by which the motion to dismiss is to be judged by the court: taking everything the plaintiff is saying as true and drawing all reasonable inferences in their favor, has the plaintiff laid out enough facts establishing that the defendant is liable for the conduct alleged? The court is not asking if the plaintiff has shown the defendant is liable, but merely have they shown enough that it makes sense to get into discovery. So if I were to sue someone for fraud, my complaint would need to lay out facts establishing all the elements of a fraud claim, namely that the defendant (1) made a material false statement that (2) I relied on that (3) caused me damages. The court does not care (at this point) if the facts actually back me up, but rather is simply asking if I were able to prove the things I am alleging, would they form the basis of a legal claim. This is the court saying to the plaintiff “this sounds plausible, let’s find out what the facts actually say.”