ANALYSIS: How Watermarks Might Erode AI’s Infringement Immunity

The first wave of intellectual property lawsuits against generative AI companies has been dominated by copyright claims. A lot of billiable hours have been dedicated to the fair use defense, and whether it applies to this new technology. But a handful of pending cases could open a back door to IP infringement liability: one that’s accessible through trademark law.

Specifically, the charge of trademark dilution—as opposed to the more common trademark infringement claim—might prove a powerful tactic for famous trademark holders to protect their intellectual property without having to wrestle with questions of fair use or substantial similarity that are inherent to copyright litigation.

In a previous analysis, I discussed Getty Images’ copyright infringement claims against an image-based generative AI platform. But that wasn’t the whole picture. Getty also included trademark claims, including infringement and dilution, in its complaint filed last February. And just this past December, the New York Times filed a complaint against OpenAI, alleging copyright infringement and trademark dilution.

According to both complaints, the plaintiffs’ mark had appeared in lower-quality AI outputs, thus diluting their name recognition and tarnishing the brands.

How AI Learned to Replicate Watermarks

Image-based AI platforms train their models on billions of images. Users enter text prompts of what they would like to see, and the platform goes to work, referencing training images and generating a (hopefully) novel image depicting what the user asked for.

These models can fall prey to biases depending on the contents of the training database. For example, if a large portion of the training images contain the same watermark in the same location on the picture, the generated output image might include a facsimile of the watermark. Unlike human beings, AI doesn’t recognize the purpose or use of a watermark; it merely considers it part of the image itself.

This is exactly what happened in Getty Images v. Stability AI and formed the basis of a unique, trademark-based theory of liability. In its amended complaint, Getty alleges that some images generated by the defendants’ platforms contain “grotesque” simulacra of its protected watermarks. To bolster its claim, Getty’s complaint included a side-by-side comparison of one of its stock photos with an AI generated facsimile.

Actual (left) and AI-generated (right) versions of a sports image included in Getty Images’ 2023 complaint against Stability AI (No. 1:23-cv-00135 (D. Del. Feb 3, 2023)).

Getty argues that these generated images, coupled with the synthesized watermarks, constitute both trademark infringement and trademark dilution under the Lanham Act—the latter of which may offer similarly situated plaintiffs a unique theory of liability.

Trademark Dilution Explained

Trademark dilution is a doctrine originating from common law and codified in the Federal Trademark Dilution Act (1995) and the Trademark Dilution Revision Act (2006). At its core, trademark dilution protects the essence and uniqueness of famous marks.

Dilution plaintiffs argue that the very existence of other products that bear marks similar to their own damages the relationship that customers have in their minds with the plaintiffs’ distinct goods and services. Unlike trademark infringement, dilution does not have a “likelihood of confusion” element, and plaintiffs do not need to prove that any actual or potential customers were “tricked” into buying the defendants’ products thinking it was that of the plaintiffs.

Without having to prove likelihood of confusion, dilution claims have a broader reach than traditional infringement. To prevent dilution claims from swallowing infringement, the Lanham Act requires a mark to reach a certain level of fame and distinction before a plaintiff may seek relief for dilution in courts. The exact level of fame required for a dilution claim is subjective and varies across jurisdictions, but generally, the bar is high and reserved only for well recognized, national brands.

Why Getty Images Might Prevail

Getty Images is in a unique position to succeed in their dilution claim against the AI defendants. As one of the largest stock image libraries on the internet, Getty could clear the fame hurdle that would trip up owners of lesser-known marks.

Perceptive users probably wouldn’t confuse the AI-generated image for a Getty original, but that wouldn’t matter in this case, as likelihood of confusion does not apply to dilution claims.

In Getty’s case, the court may find that the existence of subpar, AI-generated images bearing a semblance of the Getty mark constitutes dilution by tarnishment. This wouldn’t be a stretch, as the AI images referenced in the complaint contain distorted figures with uncanny faces, falling below the standard of quality that customers would expect from Getty. Furthermore, the presence of a Getty “watermark” might link these AI images with the Getty name in the minds of anyone familiar with the brand.

Trademark vs Copyright

Regardless of its eventual outcome, this case (as well as the New York Times complaint) presents a trademark-based alternative to copyright claims when it comes to challenging generative AI in intellectual property law.

The majority of the generative AI suits filed in the last few years have been based primarily on copyright claims, not trademark. This is because generative AI mostly deals with images and texts based on previous works of authorship protected by copyright law as opposed to source identifying marks such as company logos and taglines.

Any plaintiff alleging copyright infringement against AI defendants will invariably have to wrestle with a fair use defense, which can be an uphill battle for highly transformative uses such as AI model training.

It’s here where trademark law might offer IP owners respite. There is no comparable fair use defense for defendants to fall back on, so plaintiffs can avoid the fact- intensive fair use inquiry altogether.

Historically, courts have been hesitant to stifle progress by applying older legal doctrines to nascent technologies. If the Getty case is decided on policy considerations, the court could easily find that AI regulation is a question best left to Congress rather than the judiciary. But applying dilution doctrine to AI wouldn’t have the same widespread impact on progress that applying copyright’s fair use doctrine would. Dilution claims are inherently limited in scope—in this case, the holding would apply only to AI-generated facsimiles of distinct, famous marks.

A finding of liability here wouldn’t grind AI development to a halt the way a finding of copyright infringement might. Instead, it would force AI platforms to implement safeguards to prevent the duplication of source-identifying logos and brand names.

For most intellectual property cases in the AI space, trademark dilution won’t be a viable strategy for relief. But what it lacks in universal applicability it might make up for in expediency for protecting famous marks. Dilution could allow these plaintiffs to circumvent well-settled fair use precedent and questions of substantial similarity and customer confusion inherent in traditional infringement.

Bloomberg Law subscribers can find related content on our AI Legal Issues Toolkit and our In Focus: Artificial Intelligence page.

If you’re reading this on the Bloomberg Terminal, please run BLAW OUT <GO> in order to access the hyperlinked content, or click here to view the web version of this article.