Takeaways

  • Use in commerce: AI-generated outputs containing third-party marks may constitute use in commerce even when the reproduction is unintentional, particularly where the AI tool competes with the trademark owner’s products. The Court left this novel question open while signaling that accidental reproduction of a mark may not automatically shield AI developers from trademark liability.
  • Likelihood of confusion: Distorted or modified marks appearing on competing AI-generated outputs can support a likelihood of confusion claim at the pleading stage, even without evidence of actual confusion.
  • Passing off vs. reverse passing off: Dastar does not bar “passing off” claims where AI outputs bear another’s mark; instead, it bars “reverse passing off” claims where a defendant distributes plaintiff’s work without attribution.
  • Famous marks: Trademark dilution claims can survive at the pleading stage where the plaintiff pleads specific facts supporting household-name status, including search volume, customer counts, and presence in major publications.
  • Training data risks: Training AI models on watermarked or trademarked content may expose them to trademark liability beyond copyright concerns to the extent models reproduce those marks in their outputs.
  • Post-knowledge inaction: Awareness of trademark reproduction in AI outputs without adequate remediation may strengthen a plaintiff’s case for infringement.

Trademark owners cleared an important procedural hurdle against AI companies after a Northern District of California court ruled that Getty Images adequately stated claims for trademark infringement, false designation of origin, and trademark dilution against Stability AI.[1] Getty Images, one of the world’s leading stock photography companies, brought suit against Stability AI, the developer of the Stable Diffusion image-generation model. The lawsuit claimed that Stability AI’s models, trained on more than 12 million Getty images, produced outputs bearing distorted versions of Getty’s watermarks, creating consumer confusion and diluting Getty’s famous marks. In an order granting in part and denying in part Stability AI’s motion to dismiss, U.S. District Judge Trina L. Thompson allowed Getty’s trademark claims to proceed while dismissing only the DMCA false copyright management information claim. This decision has significant implications for AI companies using copyrighted and trademarked content in training data, potentially extending existing trademark liability to the AI development context.

AI image generation systems like Stable Diffusion are typically trained on massive datasets of images scraped from the internet. These models learn patterns from the training data to generate new images in response to user prompts. When training data includes watermarked images, the models may “memorize” and reproduce those watermarks in their outputs—a phenomenon known as “overfitting.” Getty alleged that Stability AI’s models frequently generated images bearing distorted versions of the Getty Images watermark, even when users did not request such marks. Stability AI argued that these watermarks were unintentional byproducts of the training process and could not constitute trademark use or create consumer confusion.

In making its ruling, the Court rejected Stability AI’s reliance on Dastar Corp. v. Twentieth Century Fox Film Corp,[2] holding that Dastar bars “reverse passing off” claims, where a defendant fails to credit the plaintiff, not “passing off” claims where the defendant’s product falsely implies affiliation with the plaintiff. Because the distorted watermarks on AI-generated images could suggest a Getty affiliation, the Court found that Getty’s false designation of origin claim survived.

On the trademark infringement claim, the Court found unpersuasive, at least at the pleading stage, Stability AI’s arguments that the appearance of distorted watermarks was not a “use in commerce” and could not create a likelihood of confusion. The Court noted that “[b]ecause a careful assessment of the pertinent factors that go into determining likelihood of confusion usually requires a full record, dismissal of trademark disputes at the pleading stage is generally disfavored.”[3] The Court found that Getty adequately alleged that its mark had been in use since 1995, that Stability AI’s outputs directly competed with Getty’s images, and that users had reported distorted Getty watermarks on Stable Diffusion outputs. Based on these allegations, the Court concluded that “it would be reasonable to assume that when consumers view images generated by Stab[ility] AI with the distorted Getty’s mark, they might be confused to some extent, which is sufficient to pass muster at this stage.”[4]

On the trademark dilution claims, the Court found that Getty met its burden of pleading that its mark attained “household” status. Getty alleged 2.8 billion annual searches on its website, over 708,000 customers in more than 200 countries, and that its imagery appears “every day in the world’s most influential newspapers, magazines, advertising campaigns, films, television programs, books and websites.”[5] The Court found these specific allegations sufficient to establish fame at the pleading stage. The Court acknowledged that Getty did not allege the extent to which consumers actually recognize its mark—a gap in the third statutory factor—but found that when all alleged facts were weighed together, Getty at least created “facial plausibility that the general public actually recognizes its mark.”[6]

The Court did dismiss Getty’s DMCA Section 1202(a) claim without prejudice for providing false copyright management information, finding that while Getty adequately alleged Stability AI knew its models generated distorted watermarks, the Complaint failed to allege the specific intent required by the statute—“intent to induce, enable, facilitate, or conceal infringement.”[7] The Court distinguished Section 1202(a)’s “double scienter” requirement from the lower intent standard in Section 1202(b) governing removal of CMI.

Notably, in its analysis of the California UCL claim, the Court cited Reddit, Inc. v. Anthropic PBC,[8] another AI-IP dispute over which Judge Thompson presides, in support of the proposition that state unfair competition claims alleging misrepresentation can proceed alongside federal IP claims in the AI context. This citation signals that the Northern District of California, and Judge Thompson in particular, is actively building a body of precedent on AI-IP disputes. Parties should be aware that the reasoning in these prior decisions may be applied in future disputes before this Court.

The ruling in Getty Images v. Stability AI underscores the potential legal risks facing AI companies that train models on content containing trademarks. While the case will proceed to discovery and potentially trial on the trademark claims, the Court’s willingness to let these claims survive the pleading stage signals that “overfitting” and unintentional reproduction of marks may not shield AI developers from trademark liability, at least at the pleading stage. As a practical matter, AI companies may wish to consider additional measures focused on trademark clearance of training data and output monitoring as part of their risk mitigation strategy, and trademark owners should consider proactively monitoring AI outputs and document instances of mark reproduction to build an evidentiary foundation for potential enforcement. Additionally, the Court’s citation to Reddit, signals that the Northern District of California may be actively developing a framework for AI-IP disputes and that this ruling may not be an isolated data point. This case may end up testing the “use in commerce” question on a fuller record, making it one to watch for AI companies and trademark owners alike in assessing trademark exposure from training practices.

[1] Getty Images (US), Inc. v. Stability AI, Ltd., No. 3:25-cv-06891 (N.D. Cal. Apr. 23, 2026).

[2] Dastar Corp. v. Twentieth Century Fox Film Corp., 539 U.S. 23 (2003).

[3] Getty Images, No. 3:25-cv-06891, slip op. at 4.

[4] Id. at 5.

[5] Id. at 7.

[6] Id. at 8.

[7] Id. at 3.

[8] Reddit, Inc. v. Anthropic PBC, No. 25-cv-05643-TLT (N.D. Cal. Mar. 28, 2026)