chevron-down Created with Sketch Beta.
May 09, 2024 Feature

Generative AI Litigation: A One-Year Check-In

Luke Rushing and Samuel Lowry

As generative artificial intelligence (Gen AI) closes out its first year in the mainstream, any moderately apprised attorney is aware that recent developments in this technology will likely spur new types of litigation; the question is, how? These are still early days; there are likely to be significant developments still to come, so the categories of cases are only partially crystalized. But throwing caution to wind, there are so far five main legal theories pursued in cases revolving around Gen AI: privacy, copyright, trademark, right of publicity (and facial recognition), and tort cases. Despite the differentiation, a helpful rubric for analyzing cases from all five categories is to consider whether the harm alleged is focused on the AI input, the output, or both. Typically, input-focused claims revolve around training data, not user input, and often allege impropriety in how such data were gathered or that they were accessed in violation of intellectual property rights. Output-focused cases often allege that AI outputs violate copyrights and trademarks, or produce content that is otherwise tortious.

Before diving further into the individual categories, some common features jump out. One uniting feature across the cases is the positions of the parties: In all of the major cases filed so far, the defendant has been a company involved in Gen AI. Another nearly unifying feature is California as the forum of choice, although federal courts in Delaware and Georgia also have seen cases filed. In fact, nearly all the major cases so far have been filed in the same federal district, the Northern District of California. California state law—especially California’s Unfair Competition Law, codified at Business and Professions Code § 17200, which has been pleaded in over half the cases examined for this article—is also highly prominent, but Illinois state law has been pleaded multiple times and Delaware and Georgia state law have each been pleaded once.

Gen AI in Privacy Cases

Two cases in this category have been filed thus far, both by anonymous plaintiffs against tech giants. In each case, the anonymous plaintiff was represented by the same law firm, which helps explain why both cases (one now voluntarily dismissed and the other anticipating a motion to dismiss the amended complaint) were filed in the Northern District of California and alleged violations of California’s Unfair Competition Law.

These cases are largely input-focused; the predominant claim in each is that the plaintiffs had economic and privacy rights in data that were scraped by OpenAI and Google, respectively. Google’s motion to dismiss made a wide-reaching counterargument: that the plaintiffs had failed to allege Article III standing because (1) they had not identified specific private data that had been violated and (2) they did not allege the personal information whose value was negatively impacted. Unfortunately, no court has yet weighed these theories. Google’s motion to dismiss provoked an amendment to the complaint from the anonymous plaintiffs; Google has informed the court of its intent to dismiss the amended complaint as well.

GEN AI and Copyright Cases

As of this writing, four major cases have been filed in this category: one by photographers whose work was trained on by Stability AI, one by coders whose work was trained on by Github, and two related cases by authors whose work was trained on by OpenAI. This category of claim often includes elements of input-based damages and output-based damages. For example, in all four cases, the plaintiffs alleged that a model had been trained on their copyrighted works, which were accessed without permission or proper license. However, in all four cases, the plaintiffs also alleged that outputs from the model could either resemble their works too closely or exactly reproduce their works without proper attribution (potentially violating the Digital Millenium Copyright Act).

In the Github case, the court noted that the input-focused claims were based mostly on privacy and property rights (but not explicitly based in copyright). The court found that the plaintiffs had not sufficiently alleged harm to these rights and consequently dismissed the input-based claims. However, the court found that the plaintiffs had adequately alleged that the model had the potential to output works similar to the plaintiff’s works. Thus, the court found that Article III standing was appropriate for the output-based claims.

In the Anderson v. Stability AI case, on the other hand, every claim was dismissed except for the input-based claim. In that case, several plaintiffs sued over the use of their photographs as training data for the Stability Diffusion Model. The court found that the complaint was largely deficient and dismissed all but one claim. Specific to copyright, the court found that only one plaintiff had actually alleged a registered copyright in a work that was accessed by defendants without license. Only that claim was permitted to proceed. With regards to the output-related claims, the court dismissed with leave to amend after finding that the plaintiffs had unintentionally handicapped themselves by admitting that “none of the Stable Diffusion output images provided in response to a particular Text Prompt is likely to be a close match for any specific image in the training data.” This admission helps explain why the output-based claim failed in this case but succeeded in the prior case, Doe 1 v. Github. In Github, rather than admitting that outputs would not resemble the work at issue, the “Plaintiffs argue[d] that, ‘[g]iven the number of times users may use Copilot, it is a virtual certainty [that] any particular plaintiff’s code will be displayed either with copyright notices removed or in violation of Plaintiffs’ open-source licenses for profit.’” A clear lesson emerges here: Plaintiffs seeking to recover for output-based claims would do well to include specific, credible allegations that pieces of their works will be included in model outputs.

Both cases regarding authors, Tremblay and Silverman, have pending motions to dismiss. Coupled with the leave to amend granted in both the GitHub and Stability AI cases, further developments will shed significant additional light on the viability of copyright-focused claims.

GEN AI and Trademark Cases

Getty Images v. Stability AI contains allegations of both copyright and trademark infringement, and I have chosen here to break out the trademark claims for a closer examination of the subject. Getty Images owns and licenses approximately 12 million images across the web. Their trademark is ubiquitous on internet images, so much so that images produced by Stability’s model have often incorporated variants on the mark. Under Getty’s theory, Stability infringes on Getty’s mark when its models reproduce their trademark. Furthermore, because images produced by Stability are often lower quality than typical photographs (e.g., the distorted human faces and unresolved details that are hallmarks of generated images), Getty argues that their trademark is diluted when it is wrongfully associated with generated images. Stability AI initially moved to dismiss the complaint on procedural and technical grounds before the parties agreed to conduct jurisdictional discovery. The case is stayed while that discovery is still ongoing.

Right of Publicity and Facial Recognition

This category of cases revolves around the use of facial data in image-generating (or image-modifying) software. Neither of these cases necessarily involves AI; in each instance, the AI component could simply have been any technology. Nonetheless, cases involving AI only tangentially will likely represent a significant amount of AI case law, so they still merit examination. In Prisma Labs, the defendant corporation owned and operated an app that allowed users to upload photos that would be converted into artistic images. Plaintiffs alleged that the software also stored those facial data without informing its users so that the data could be used to train Prisma’s neural network. These allegations were brought almost entirely under Illinois state law, which is particularly plaintiff-friendly for biometric law, and did not allege any damages resulting from the Gen AI output. That case has been ordered by the court into arbitration.

Similarly, in Young v. NeoCortext, the defendant corporation operated an app that altered users’ photos. But in this instance, the app also offered a premium service that advertised a model that could swap users’ faces with those of celebrities. One of the celebrities promoted was reality star Kyland Young, who alleged that he never consented to the use of his likeness and had not received compensation despite the defendant’s profits. Young sued for one claim of California’s right of publicity law. The defendants moved to dismiss the sole claim, but the court denied the motion. The most relevant holding for the AI industry was that replacing Young’s face—while leaving his body and using AI to realistically merge the user’s face into Young’s body—was not transformative enough at the motion to dismiss stage to act as an affirmative defense. Purveyors of generated images should take note: If the purpose of transforming a photo still involves retaining an element that is protectable by a third party, that third party may be able to claim damages.

GEN AI and Tort Damages

Finally, Walters v. OpenAI raises the question of whether AI can produce content that is defamatory. A journalist investigating a legal case asked ChatGPT for information regarding the case; ChatGPT responded that Mark Walters was accused of embezzlement. In reality, Mr. Walters was accused of no such thing and was not related to the complaint being investigated. Walters brought a state law claim for libel in Georgia state court; the case was removed to federal court and then remanded back to state court, where it was subject to a motion to dismiss from OpenAI. OpenAI argued that the journalist who encountered the falsehood did not take it to be accurate and that the plaintiff could not prove actual malice from the output of a statistical model. The state court recently denied the motion to dismiss but without further analysis of the arguments at issue.

What to Watch in Year Two of AI Litigation

Mainstream generative AI is currently in its second year and already there are multiple avenues for litigation surrounding it. Here are some of the key takeaways as we enter the second year:

It’s Still Early Days

Many of the cases discussed above remain in their infancy. For the few that have had any consideration on the merits, that has only been at the motion to dismiss stage, leaving plenty of space for additional rulings. Expect substantial new developments in all of these cases over the following year.

Federal Courts in California Are the Forum of Choice

Perhaps unsurprisingly given that the center of gravity of the tech world lies in Northern California, nearly every complaint filed has been in a California court, and almost all of those were in the Northern District of California.

Federal, California, and Illinois Law Are the Main Jurisprudence at Play

The federal copyright regime, the federal trademark regime, and the Digital Millenium Copyright Act are all regularly cited in AI cases. Furthermore, many plaintiffs have relied on California state law, especially regarding privacy and unfair competition, as well as Illinois law regarding privacy in biometric data.

Specificity Is Key

This is perhaps a practice tip applicable to all attorneys. In part due to the large number of cases decided on the motion to dismiss standard, whether plaintiffs have pleaded sufficiently specific harms is an issue courts have regularly grappled with. Successful plaintiffs have gone out of their way to allege their copyright registrations, that their works or private data were actually used in training data, and that outputs may resemble their data, work, or likeness. Unsuccessful plaintiffs have simply alleged in general that the entire internet was scraped for data, so their works must have been included, or any work in the training set might be contained in an output.

Transformation Should Be Comprehensive

Many AI services have capitalized on AI’s ability to transform images and text. The ruling in Young implies that transformations should be comprehensive and should not intend to maintain elements of the original, or the affirmative defense of transformation may not apply.

* * *

As generative AI enters its second year, keep your eyes on this space for significant updates on the state of the law.

    The material in all ABA publications is copyrighted and may be reprinted by permission only. Request reprint permission here.

    Luke Rushing


    Luke Rushing began his career in commercial litigation focusing on entertainment law. In the years since, he has added significant civil rights and maritime work to his portfolio. As generative AI has surged in usage and in the zeitgeist, he has spoken on the subject for law firms, interviewed numerous thought leaders in the field, and served as a vice chair to the ABA’s Committee on AI and Robotics.

    Samuel Lowry

    Huth Reynolds LLP

    Samuel Lowry is a legal analyst at the boutique firm of Huth Reynolds LLP and assists with issues ranging from maritime law to generative AI. He graduated from Harvard College in 2023.