Authored By: Eshana Warsi
University of Greater Manchester, RAK
ABSTRACT
The surge of generative artificial intelligence technologies that can create fashion design ideas strikingly similar to the known trademark that exemplifies a flagship brand of and/or a well-known luxury brand owner has raised an inherent liability issue in current IP frameworks. This article will explore the key legal issue of whether the existing trademark systems in the United States, European Union, and the international intellectual property regime are sufficient to determine and allocate liability when an AI system produces a design that violates a registered luxury trademark. By adopting a doctrinal and analytical approach, this article challenges whether the direct, contributory, and vicarious trademark infringement doctrines are applicable to AI outputs, questions the ratio decidendi of Getty Images (US) Inc v Stability AI Ltd and other cases, and discusses the EU Act on AI 2024 as a new regulatory tool. The author concludes that such frameworks were designed solely for human infringers and would fail to solve any claims of intellectual property rights infringement by AI without a proper legislative measure.
Keywords: Generative Artificial Intelligence, Trademark Infringement, Luxury Fashion Law, Training Data Liability, Intellectual Property, AI Accountability, Trade Dress Protection
- INTRODUCTION
Trademark law has long played a critical role in protecting the most important asset of the global luxury fashion industry: distinctiveness of visual identity, estimated by the Financial Times at around USD 1.84 trillion in 2025[1]. The commercial value of the interlocking “CC” of Chanel, the Toile Monogram of Louis Vuitton, and the red sole of Christian Louboutin is entirely based on legal exclusivity. The arrival of generative AI products, however, like Midjourney, Adobe Firefly, and Stable Diffusion, has posed a new challenge to the old trademark doctrine. Today, these systems can create fashion designs that mimic, often to disturbing accuracy, the clothing visual language of the luxury houses.
A fundamental legal issue that has not been answered yet is who should be held legally responsible for the infringement of a registered luxury trademark when the infringement is caused by the use of an AI system in creating the design? The developer, the user, or no one at all?
The importance of this question is reflected in Getty Images (US) Inc v Stability AI Ltd[2] where the court had to deal with a structural lack in trademark law as well, such that it was faced with an AI system that was systematically reproducing protected visual material. The doctrine of trademarks, built entirely around the assumption of identifiable intent on the part of a human infringer, has nothing to offer a coherent way to hold a self-governing system liable. The EU AI Act 2024[3] is the most recent legislation to respond to the phenomenon, and the relationship between the EU AI Act and luxury trademark enforcement has been analysed relatively little.
The article consists of four parts: a conceptual foundation of trademark infringement, an analysis of the liability gap resulting from AI outputs, a comparative analysis of the approaches taken across the US and the EU, and recommendations for reform. In essence, the central argument of the submissions was that the existing frameworks are structurally inadequate to address trademark infringement caused by AI, and therefore, there is a need for legislative action.
- BACKGROUND AND CONCEPTUAL FRAMEWORK
To examine the accountability gap that AI-generated fashion products have created, it is crucial to delve into the legal landscape of trademark protection. In the European Union, the EU Trade Mark Regulation[4], like its U.S. counterpart, the Lanham Act 1946[5], forbids the use of an identical or similar mark in commerce if such use is likely to cause confusion for the consumer as to the source of the goods.
In the field of fashion law, trademark protection is not limited to traditional word marks, but also applies to trade dress, which is the unique look or appearance of a product or its packaging. In Wal-Mart Stores Inc v Samara Brothers Inc[6], the United States Supreme Court ruled that product design trade dress must be acquired distinctiveness, a principle which luxury fashion houses have cited heavily. Brands like Hermès, Chanel, and Louis Vuitton have spent the decades building up exactly that kind of distinctiveness, which is the most commercially valuable and legally protected of all in the world.
Historically and doctrinally, there are three categories of trademark infringement liability. A party directly liable for using a trademark infringement directly[7]. Contributory liability, which arises from Inwood Laboratories Inc v Ives Laboratories Inc[8], applies where a person knowingly supplies the means for another person to use an infringing mark. Vicarious liability applies when a person can, and does, control the infringing conduct[9]. Each of the liability categories relies upon the assumption of a human party that can possess knowledge, intent, and control over another party that infringes on the mark. When the infringer is non-human or an AI that is autonomous, all of those assumptions fall apart.
While there has been extensive scholarship on the intersection of AI and copyright, the application of the trademark infringement doctrine to AI-generated fashion outputs is yet to be critically examined, and this article fills such void[10].
- LEGAL ANALYSIS
3.1 The existing doctrine fails: The Liability Vacuum
The core issue that this article will argue is spot-on: Generative AI systems do not currently fit into the three-part test of direct, contributory, or vicarious trademark infringement. When the doctrine is applied to the fashion outputs created by AI, it shows a structural weakness that cannot be filled by the current law without judicial imagination or legislative action.
Because AI generates a design that copies Louis Vuitton’s Toile Monogram in response to a prompt by a user, the AI system has no volition. It does a statistical prediction of data, using training data. It does not mean that it is trying to infringe, nor does it have the capacity to appreciate that it is trying to infringe. It is therefore doctrinally impossible to have direct liability against the AI system itself at the present time.
The question now becomes: who is the one to type the prompt, the user? It is submitted that user liability for direct infringement is also precarious. Where a user tells the AI system to create “a luxury monogram handbag in the style of Louis Vuitton,” but does not tell the AI system to make a copy of the protected mark in the first place, the user has not used the mark in commerce, because the user has asked for something the AI system independently concludes infringes upon the protected mark. This is consistent with the reasoning in Religious Technology Center v Netcom On-Line Communication Services Inc[11], which held that there was no direct copyright infringement upon the mere provision of a service through which copyright infringement takes place. That case was a copyright case, but the volition principle is analytically transferable to trademark direct liability, and directly undermines any argument that there is automatic user liability.
3.2 Contributory Liability and the Knowledge Problem
The standard for contributory trademark infringement, namely, that the defendant either intentionally induced infringement or continued to provide its product to a party whom it knows or has reason to know, to be engaged in trademark infringement, was set forth in Inwood Laboratories Inc v Ives Laboratories Inc[12]. But the ratio of the Supreme Court in Inwood is important: Knowledge is the operating standard. A defendant who provides a tool without knowing of its infringing use cannot be found contributorily liable.
When applied to the context of AI development, it gives rise to the knowledge problem, as described in this article. As part of generating any given output, a Stability AI or Midjourney developer is not aware of what a user might have submitted as a prompt, or what trademark that output might infringe. This developer’s knowledge (if any) is systemic, not specific, meaning that the developer knows that the model can generate outputs that could be protected marks, but he or she doesn’t know about any specific infringing act. It is unclear whether systemic awareness meets Inwood’s knowledge threshold. The Second Circuit in Tiffany (NJ) Inc v eBay Inc[13] decided that contributory infringement does not exist just because it is well known that others are infringing on the product, but only when they have specific knowledge of specific infringing listings. But applied to AI developers, this argument significantly weakens the case for contributory liability and creates an unreliable defendant for luxury trademark owners.
3.3 Support of the European Union Position and the AI Act 2024
The EU’s approach is more structured, but still incomplete, compared with the approach in other countries. It was confirmed in L’Oréal SA v eBay International AG[14] by the Court of Justice of the European Union that an operator who engages in an active element of optimisation or promotion of infringing listings may be liable— beyond the passive intermediary. The active role standard is analytically important for AI developers who use luxury fashion imagery specifically to train the model, indicating intentional design decisions that are more than mere facilitation.
EU AI Act 2024 adds a new level of regulatory burden. Providers of general-purpose AI models must keep technical documentation capable of identifying the source of training data and provide policies to ensure compliance with IP law[15]. This is done in the context of regulation, but it is nonetheless an additional source of evidence for a potential trademark infringement claim.²² Under European law, an argument for contributory liability is significantly strengthened when an AI developer’s training data does not contain any protected luxury trademarks.It’s a materially stronger argument than the current definition of contributory liability in the United States, does not allow, under European law, that the AI developer’s training data did not include any protected luxury trademarks.
3.4 The Problem of Training Data
There is a fundamental question, which, as stated, has not been squarely addressed by either US or EU trademark law: the protected output of a generative AI system is “downstream” of the “upstream” use of training data. Stable Diffusion learns about these protected images, such as the Chanel quilt, the Hermès Birkin silhouette, and more, when trained on millions of images containing each of these visual marks. All further outputs reproducing the same patterns are the results of this training decision. It is therefore submitted that trademark liability analysis must be extended to the training stage, a position which has not yet been articulated by doctrine and which is the greatest reform imperative which has been suggested by this article.
- CASE LAW
4.1 Getty Images (US) Inc v Stability AI Ltd (2023)[16]
Getty Images alleged that Stability AI’s image-generation model had been trained on over twelve million photographs from Getty’s licensed database without authorization, producing outputs that replicated Getty’s watermark alongside protected visual content.
The Delaware District Court declined to dismiss the claim, notably locating potential liability at the training stage rather than the subsequent output stage. The court found that the act of training an AI model on protected material and subsequently deploying it commercially constitutes a sufficiently pleaded act of infringement. This interpretation has significant implications for luxury trademark holders, as it suggests that a developer who deliberately incorporates protected visual identities into training data may incur primary liability for all subsequent infringing outputs, irrespective of the user who initiated their generation. This perspective directly supports the upstream liability argument presented in this article.
4.2 Hermès International v. Mason Rothschild (2023)[17]
Mason Rothschild made and sold NFT art that he called “MetaBirkins”. These digital images of Hermès Birkin bags used fur textures, and Hermès never authorized this use. In a case brought by Hermès against Rothschild in the Southern District of New York, the Court held that the MetaBirkins are commercial speech and not protected artistic expression, and found that Rothschild purposefully took advantage of the association between consumers and the Birkin trademark to benefit financially.
The ratio decidendi of this case has direct relevance for AI-generated fashion output. The Court found that digital copies of luxury trademarks and their corresponding designs are actionable regardless of whether they are an artistic work, especially when there exists a risk of consumer confusion with respect to the source or sponsorship of the infringing item. The same principle applies to AI-generated output where the automatic creation of the output does not automatically exempt it from trademark liability. However, an open question is whether infringement liability exists if the infringing digital copy was created by a machine and not by an individual artist. This article addresses that gap.
4.3 L’Oréal SA v. eBay International AG (2011)[18]
The ruling of the Court of Justice of the European Union concluded that if a platform operator is engaged in any capacity in assisting with the evaluation or promotion of infringing goods, then they cannot claim to be a “passive intermediary” and will therefore be liable for trademark violations. Although this case occurred before generative AI technologies existed, its ratio still applies directly to generative AI development activities. Specifically, if a generative AI developer trains a generative AI system using images of luxury trademarks, optimizes the outputs produced by that generative AI system so that the final result mimics protected visual identities, and sells access to such a system commercially, then that generative AI developer is behaving more like an active facilitator than a passive intermediary. This standard under European law greatly bolsters the argument for developer liability. Thus, the European position is much further advanced than the current position of the United States and suggests another area in which to find alignment through legislation; this is what this article advocates for.
- CRITICAL ANALYSIS
5.1 Current Doctrine’s Structural Deficiencies
The core finding of this article is clear: existing trademark law in both the U.S. and the E.U. was designed with a universe in which there were individual infringers, and the three-part framework of the law of direct, contributory, and vicarious liability requires all three parts to have a volitional actor or an individual party that has the capacity to have knowledge and intent. A human cannot be defined as a volitional actor in any situation where an AI generates or creates content. This creates a vacuum in structural accountability for the holders of luxury trademarks who cannot navigate the current doctrine in order to establish their rights or to protect their brand identity. All three cases have found a human person at least once in the process of infringement: Getty Images, Hermes v. Rothschild, eBay v. L’Oreal. When there is no longer a human actor in the infringement chain, the current doctrine does not provide an answer.
5.2 Differences Between Jurisdictions in Their Judicial Trends
There is a notable difference in the approach to liability for AI developers between the United States and the European Union. The knowledge threshold limitations of Inwood, as well as the requirement that there has been voluntary conduct by a defendant, have given U.S. courts significant hesitation in holding AI developers accountable for upstream infringement of AI-based products unless there is specific, identifiable knowledge of their infringing conduct. Conversely, based on the L’Oréal v. eBay standards established by the European courts and their active role in design choices (especially in respect of curating training data), the European courts are more inclined to hold developers liable based on their deliberate choices. As a result, there exists a dangerous opportunity for regulatory arbitrage; AI developers can potentially use jurisdictions that provide them with the lowest exposure to liability as a means of structuring and/or conducting operations while deploying their systems globally. Luxury brands that own trademarks enjoy an inherently transnational nature and therefore bear the consequences of this inconsistency in AI systems.
5.3 Implications for Policy and Reforms Needed
This analysis sheds light on three key areas for reform. The first area requires trademark law to be modified so that copyright infringement for commercial use of AI-generated content includes the intentional use of protected trademark marks in training data. Second, the documentation requirements outlined in the EU AI Act must include a clear requirement for the clearance of any trademarks prior to commercial use. Finally, establishing an international treaty developed with WIPO that is similar in structure to the WIPO Copyright Treaty[19] will assist with harmonising the standards for AI-related infringement throughout different countries, thus preventing regulatory arbitrage. Failure to address these areas of concern will result in a widening gap from the point of view of accountability as generative AI continues to advance in capability.
- CONCLUSION
With the advent of generative AI, the paradigm of trademark law has been changed forever. The article showed that, under the current law in the US and in the EU, a court is unable to be certain of the liable party when an AI system creates a fashion design that violates a registered luxury trademark. All three of direct, contributory, and vicarious liability each demand a human player – and an independent system will not suffice in this situation for luxury trademark holders to find a satisfactory remedy.
As these three cases demonstrated, the limits of what current doctrine can accomplish can be seen in the combined case of Getty Images v Stability AI, Hermès v Rothschild and L’Oréal v eBay. They also show us how close the law is to acknowledging responsibility for upstream training stages – but not quite there yet.
The recommendations in this article can therefore be summarized into a three-part call to action: recognition of training data liability in applicable trademark law; extension of the AI Act in the EU to include the requirement for trademark clearance before deployment; and WIPO[20] to facilitate the development of an international instrument to harmonise trademark liability with regard to AI. This algorithm is already fit for appropriate use. The law must now come along for the ride!
Bibliograpghy
Table of Cases
- Fonovisa Inc v Cherry Auction Inc 76 F 3d 259 (9th Cir 1996)
- Ford Motor Co v GreatDomains.com Inc 177 F Supp 2d 635 (ED Mich 2001)
- Getty Images (US) Inc v Stability AI Ltd [2023] No 1:23-cv-00135 (D Del)
- Hermès International v Mason Rothschild 590 F Supp 3d 647 (SDNY 2023)
- Inwood Laboratories Inc v Ives Laboratories Inc 456 US 844 (1982)
- L’Oréal SA v eBay International AG (C-324/09) [2011] ECR I-6011
- Religious Technology Center v Netcom On-Line Communication Services Inc 907 F Supp 1361 (ND Cal 1995)
- Thaler v Vidal 43 F 4th 1207 (Fed Cir 2022)
- Tiffany (NJ) Inc v eBay Inc 600 F 3d 93 (2d Cir 2010)
- Wal-Mart Stores Inc v Samara Brothers Inc 529 US 205 (2000)
Table of Legislation
United States
- Lanham Act 1946, 15 USC §§ 1114, 1125(a)
European Union
- Regulation (EU) 2017/1001 of the European Parliament and of the Council on the European Union trade mark [2017] OJ L 154/1
- Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence [2024] OJ L 1689/1
International
- Agreement on Trade-Related Aspects of Intellectual Property Rights (Marrakesh, 15 April 1994)
- WIPO Copyright Treaty (Geneva, 20 December 1996)
Books
- Cornish William, Intellectual Property: Patents, Copyright, Trade Marks and Allied Rights (9th edn, Sweet and Maxwell 2019)
- Scafidi Susan, Who Owns Culture? Appropriation and Authenticity in American Law (Rutgers University Press 2005)
Journal Articles
- Abbott Ryan, ‘I Think, Therefore I Invent: Creative Computers and the Future of Patent Law’ (2016) 57 Boston College Law Review 1079
- Dogan Stacey and Lemley Mark, ‘The Merchandising Right: Fragile Theory or Fait Accompli?’ (2004) 54 Emory Law Journal 461
- Scafidi Susan, ‘Intellectual Property and Fashion Design’ (2006) 1(1) Intellectual Property Law Review 115
Online Sources
- Statista, ‘Global Luxury Fashion Market Size 2025’ (Statista, 2025) https://www.statista.com accessed 6 June 2026
- World Intellectual Property Organization, ‘Artificial Intelligence and Intellectual Property’ (WIPO, 2024) https://www.wipo.int/about-ip/en/artificial_intelligence accessed 6 June 2026
[1] Statista, ‘Global Luxury Fashion Market Size 2025’ (Statista, 2025) https://www.statista.com accessed 1 June 2026
[2] Getty Images (US) Inc v Stability AI Ltd [2023] No 1:23-cv-00135 (D Del)
[3] Regulation (EU) 2024/1689 (EU AI Act) (n 3)
[4] Regulation (EU) 2017/1001 of the European Parliament and of the Council on the European Union trade mark [2017] OJ L 154/1, art 9
[5] Lanham Act 1946, 15 USC § 1114
[6] Wal-Mart Stores Inc v Samara Brothers Inc 529 US 205 (2000)
[7] Lanham Act 1946, 15 USC § 1125(a)
[8] Inwood Laboratories Inc v Ives Laboratories Inc 456 US 844 (1982)
[9] Fonovisa Inc v Cherry Auction Inc 76 F 3d 259 (9th Cir 1996)
[10] Ryan Abbott, ‘I Think, Therefore I Invent: Creative Computers and the Future of Patent Law’ (2016) 57 Boston College Law Review 1079
[11] Religious Technology Center v Netcom On-Line Communication Services Inc 907 F Supp 1361 (ND Cal 1995)
[12] Inwood Laboratories Inc v Ives Laboratories Inc (n 10)
[13] Tiffany (NJ) Inc v eBay Inc 600 F 3d 93 (2d Cir 2010)
[14] L’Oréal SA v eBay International AG (C-324/09) [2011] ECR I-6011
[15] Regulation (EU) 2024/1689 (EU AI Act) (n 2) art 53
[16] Getty Images (US) Inc v Stability AI Ltd [2023] No 1:23-cv-00135 (D Del) (n 5)
[17] Hermès International v Mason Rothschild 590 F Supp 3d 647 (SDNY 2023)
[18] L’Oréal SA v eBay International AG (C-324/09) [2011] ECR I-6011 (n 21)
[19] WIPO Copyright Treaty (Geneva, 20 December 1996)
[20] WIPO Copyright Treaty (Geneva, 20 December 1996)
