The growing adoption of artificial intelligence (AI) systems and the increasing importance of AI-generated content have created new challenges for the protection of intellectual property rights. Although the European Union has already adopted the pioneering AI Act, the existing legal framework still contains numerous open-textured concepts whose precise meaning will ultimately be shaped through judicial interpretation. This is particularly true with respect to the relationship between AI models and intellectual property law, including the use of copyright-protected works for AI training and liability for content generated by artificial intelligence.

Accordingly, the case law of the Court of Justice of the European Union (CJEU) and national courts will assume an increasingly significant role. It will be for the courts to determine how concepts such as the reproduction of a work and communication to the public should be interpreted in the context of AI, as well as to define the boundaries of liability of AI providers and users and the scope of protection afforded to rightholders.

This article, forming part of our series on artificial intelligence, examines the proceedings currently pending before the CJEU and national courts that are likely to shape the future direction of AI-related case law.

 

Key Cases – C-250/25 Like Company v. Google and Getty Images v. Stability AI

Among the growing number of AI-related disputes, two proceedings deserve particular attention: Getty Images v. Stability AI, currently before the High Court of Justice of England and Wales, and Case C-250/25 Like Company v. Google Ireland, pending before the Court of Justice of the European Union.

 

C-250/25 Like Company v. Google Ireland

Case C-250/25 Like Company v. Google Ireland is of particular significance for the development of European AI law, as it represents the first major preliminary reference concerning generative artificial intelligence brought before the CJEU.

The dispute concerns responses generated by Google Gemini, which, according to the claimant – a press publisher – constituted detailed summaries of its news articles. The publisher argued that these summaries reduced users’ incentive to access the original publications and consequently undermined its ability to generate advertising revenue.

The Hungarian court referred several preliminary questions to the CJEU concerning, among other matters, whether a chatbot response may constitute a communication to the public of protected content, whether the training of AI models amounts to the reproduction of copyright-protected works, and whether such activities fall within the text and data mining (TDM) exception provided for in Article 4 of the DSM Directive.

During the hearing held on 10 March 2026, particular attention was devoted to the technical functioning of large language models (LLMs). Google argued that its model does not store articles as textual copies but instead operates on statistical representations of data, with responses being generated dynamically in response to user prompts. The claimant, by contrast, submitted that, from the perspective of copyright law, the decisive factor is the output produced by the system. Where a chatbot provides content that effectively substitutes for reading the original article, the publisher’s exclusive rights may be infringed.

The hearing also addressed several additional issues, including the operation of retrieval-augmented generation (RAG) and the territorial scope of EU copyright law where AI model training takes place outside the European Union. Representatives of six Member States – including Germany, France and Spain – as well as the European Commission, also presented oral observations.

The significance of the case for the European Union cannot be overstated. Both the fact that it is being heard by the CJEU and the legal questions referred make it a landmark proceeding. The forthcoming judgment is expected to establish the authoritative interpretation of EU copyright law as it applies to generative AI throughout the Union, particularly with respect to the scope of the TDM exception and liability for content generated by AI chatbots.

The next procedural milestone will be the Opinion of Advocate General Maciej Szpunar, which is expected no earlier than 3 September 2026.

 

Getty Images v. Stability AI (United Kingdom)

Another landmark proceeding is Getty Images (US), Inc. and Others v. Stability AI Ltd., heard before the High Court of Justice, Business and Property Courts of England and Wales, Intellectual Property ([2025] EWHC 2863 (Ch)). The judgment represents one of the first comprehensive judicial analyses of the relationship between generative AI and intellectual property law.

Getty Images alleged that Stability AI had unlawfully used millions of photographs from the Getty Images platform to train its Stable Diffusion model. The original claim included allegations of copyright infringement relating to millions of photographs, infringement of the sui generis database right, and trade mark infringement.During the proceedings, however, the scope of the claim was substantially narrowed. Getty withdrew its allegation that the model-training process itself constituted direct copyright infringement after it became apparent that the training had taken place outside the territorial scope of UK copyright law. It also abandoned its database right claims and the allegation that images generated by Stable Diffusion reproduced specific Getty photographs.As a result, the Court did not determine the legality of using copyright-protected works for AI model training, leaving that fundamental issue unresolved.

The principal issue before the Court was whether the trained Stable Diffusion model itself could be regarded as an infringing copy of the copyright-protected works used during training. The Court concluded that such a characterisation was not sustainable. It held that an AI model does not contain or store copies of the works on which it has been trained. The judgment emphasised that, in order for an object to constitute a copy for the purposes of copyright law, it must at some stage embody or contain a copy of the protected work itself. By contrast, the weights of a neural network are merely numerical representations rather than stored reproductions of individual photographs. The Court stressed that its decision should not be understood as establishing that generative AI models are inherently incapable of memorising portions of their training data. Rather, it found only that Getty had failed to demonstrate that the version of Stable Diffusion at issue exhibited such memorisation. The Court found infringement solely in relation to certain earlier versions of Stable Diffusion, which, on isolated occasions, generated images containing elements resembling trade marks owned by the Getty Images group. However, these instances were found to be incidental and limited in scope. Only a handful of such examples were established, and there was no evidence that similar outputs occurred during ordinary use of the model by end users.

The judgment is particularly noteworthy because it clearly distinguishes between the training phase of an AI model and the trained model itself. While it does not resolve the legality of training AI models on copyright-protected works, it confirms that a trained model should not automatically be regarded as a copy of the materials used during training.

Consequently, the decision leaves unanswered the central legal question of whether the large-scale use of copyright-protected works for AI training is lawful. That issue will continue to be examined in pending proceedings, including parallel litigation before the United States courts. It should also be noted that the judgment is not yet final, as Getty Images has appealed the decision.

 

German case law – the most developed body of AI jurisprudence in Europe

Among the Member States of the European Union, Germany has thus far developed the most extensive body of case law concerning artificial intelligence.

German courts have addressed a broad spectrum of legal issues, ranging from the application of the text and data mining (TDM) exception, copyright protection of AI-generated content, and the use of copyright-protected works for AI training, to liability for infringements of personality rights committed through AI systems, including the protection of an individual’s voice.

When does AI „memorise” a copyright-protected work?

One of the most significant legal questions arising in connection with AI concerns the phenomenon commonly referred to as memorisation. This occurs where an AI model internalises a copyright-protected work during training to such an extent that it is subsequently capable of reproducing that work, or substantial parts of it, without relying on user input or external sources.This issue lies at the heart of GEMA v. OpenAI, currently pending before the Munich Court (Case No. 42 O 14139/24). The German collective management organisation GEMA demonstrated that a chatbot instructed to act as an expert on German music was capable of reproducing verbatim excerpts from the lyrics of well-known songs without accessing internet sources. The Court held that such conduct may constitute both the reproduction of copyright-protected works within the AI model itself and their subsequent communication to the public through responses provided to users.

Crucially, the Court further held that such conduct falls outside the scope of the TDM exception laid down in Article 4 of the DSM Directive, since model memorisation extends beyond the activities covered by that exception. This distinction may prove significant for future AI litigation. There is a fundamental legal difference between using a work solely for data analysis during the training process and retaining that work within the model in a manner that enables its subsequent reproduction.

When does AI training fall within the TDM exception?

The scope of the text and data mining exception was also examined in Kneschke v. LAION, decided by the Regional Court of Hamburg (Case No. 310 O 227/23). The case concerned a photograph bearing a watermark that had been downloaded from a website by LAION and subsequently used to analyse the correspondence between the image and its accompanying textual description. The Court held that downloading the photograph constituted an act of reproduction within the meaning of German copyright law. However, in the circumstances of the case, that act was covered by the TDM exception. Particular weight was attached to the fact that LAION operated as a non-profit organisation and that the resulting dataset was made publicly available free of charge in order to facilitate AI research. The Court did not consider it decisive that the dataset might subsequently be used by commercial entities.

Both cases illustrate two distinct legal issues arising from the use of copyright-protected content by AI systems. In LAION, the Court accepted the use of a protected work as part of a data analysis process falling within the TDM exception. In GEMA, by contrast, the legal issue arose at a later stage, when the model had become capable of reproducing protected content.For AI model developers, the distinction between these two situations may ultimately prove decisive.

 

Can AI-generated content qualify for copyright protection?

German courts have also considered whether content generated with the assistance of artificial intelligence may qualify for copyright protection. In Case C 9786/25, the Munich Court examined AI-generated logos that had subsequently been used on a website. The Court held that the logos themselves did not qualify for copyright protection. At the same time, the Court expressly recognised that works created with the assistance of AI are not excluded from copyright protection per se. Copyright may subsist where the resulting work reflects the author’s personality and embodies the author’s own intellectual creation through a sufficient degree of human creative input.

A similar conclusion was reached by the Higher Regional Court of Düsseldorf (Oberlandesgericht Düsseldorf) in Case I-20 W 2/26, which concerned content generated on the basis of an existing image. The Court held that the generated output constituted neither an independent copyright work nor an original adaptation, as it lacked the degree of individuality and creativity required for copyright protection.

This shows that the mere use of AI neither automatically confers nor precludes copyright protection. The decisive question remains whether the final output embodies a sufficient level of independent human creativity.This principle is likely to become increasingly significant in future disputes concerning authorship and ownership of AI-assisted creative works.

 

Who is liable for what AI says?

The rapid development of generative AI has also raised fundamental questions regarding liability for information disseminated by AI systems. One of the leading German cases in this area is Campact v. xAI (Case No. 324 O 461/25). In that case, the non-governmental organisation Campact alleged that Grok, the chatbot integrated into the X platform, had generated false statements suggesting that the organisation had received financial support from the German government.

The Regional Court of Hamburg ordered the removal of the disputed content, holding that the statements were defamatory and infringed the claimant’s personality rights. Importantly, the Court found that the legal assessment was unaffected by the fact that the statements had been generated by an AI system rather than a human author. The burden of proving the truthfulness of the information published by the chatbot rested with xAI. The judgment reflects an emerging judicial approach according to which AI-generated content is assessed under the same substantive legal standards that apply to human-generated content. The use of artificial intelligence does not, in itself, shield an AI provider from liability.

A similar issue arose in Case 2 O 202/24, decided by the Regional Court of Berlin, although in a different factual context. The proceedings concerned the use, in a political campaign advertisement, of an AI-generated voice closely resembling that of a well-known German dubbing actor. The Court upheld the claimant’s claims, finding that the use of the synthetic voice was capable of misleading the public into believing that the actor endorsed the political message conveyed by the advertisement. The decision illustrates that personality rights – including the right to one’s voice and identity—may be infringed even where no actual recording of the person’s voice has been used. The decisive factor is whether the AI-generated output is capable of creating the false impression of endorsement or association.

German case law demonstrates that many legal questions surrounding artificial intelligence remain unresolved. These include the lawful use of copyright-protected works for AI training, liability for AI-generated content, copyright protection of AI-assisted outputs, and the protection of personality rights in the age of generative AI. Although these decisions are not binding outside Germany, they are likely to serve as an important reference point for the future development of European AI law, particularly as the first disputes concerning the interpretation of the AI Act and existing EU copyright legislation begin to reach the Court of Justice of the European Union.

 

How is AI case law developing in the United States?

No discussion of artificial intelligence would be complete without considering developments in the United States. In debates surrounding AI regulation, the United States and the European Union are frequently portrayed as representing two contrasting approaches: the EU’s comprehensive regulatory model and the comparatively more market-oriented approach traditionally adopted in the United States. Against that background, administrative authorities and the courts play a particularly significant role in defining the legal boundaries of AI deployment. To date, U.S. practice has focused on two principal areas of dispute. The first concerns authorship and the copyright protection of AI-assisted works. The second relates to the legality of using copyright-protected works to train AI models.

 

Who Is the author of an AI-generated work? The Position of the U.S. Copyright Office

In the United States, the U.S. Copyright Office (USCO) plays a central role in defining the scope of copyright protection. In addition to registering copyright claims, the Office issues guidance and policy reports on copyright law, including the implications of artificial intelligence.The Office’s current position is clear – content generated autonomously by an AI system is not, by itself, eligible for copyright protection. However, this does not preclude protection for works in which a human author has made a sufficiently creative contribution.

One issue has been resolved unequivocally, namely: an AI system cannot be recognised as the author of a copyright work.This principle was confirmed in the litigation concerning the artwork A Recent Entrance to Paradise. Dr Stephen Thaler applied to register a work generated by an AI system of his own creation, identifying the AI system itself as the author. The U.S. Copyright Office refused registration, and that decision was subsequently upheld by the United States Court of Appeals for the District of Columbia Circuit (Case No. 25-5233).The Court held that, under U.S. copyright law, authorship is reserved exclusively for human beings. Consequently, an artificial intelligence system cannot qualify as the author of a copyright work. This position broadly aligns with the approach adopted in European copyright law.

A much more nuanced question concerns works produced through a combination of human creativity and artificial intelligence. The leading administrative decision is Théâtre D’opéra Spatial (USCO, SR # 1-11743923581). The work had been generated using Midjourney and subsequently edited by its creator in Adobe Photoshop. The U.S. Copyright Office concluded that the AI-generated elements were not themselves eligible for copyright protection, irrespective of the number of prompts used during the creative proces even though, in this case, the author had reportedly used more than 600 prompts. At the same time, the Office acknowledged that independently created human modifications may qualify for copyright protection where they satisfy the ordinary requirements of originality. In the circumstances of the case, however, the human contribution was considered insufficiently creative in comparison with the AI-generated elements to justify protection for the work as a whole.

The same principle was applied in the well-known Zarya of the Dawn decision. The Office refused copyright protection for the individual illustrations generated using AI, while recognising protection for the elements created by the human author, including the written text and the overall selection, coordination and arrangement of the comic book. Accordingly, copyright subsisted in the creative organisation of the work, but not in the AI-generated images themselves.

The current position of the U.S. Copyright Office therefore draws a relatively clear distinction. Artificial intelligence cannot be recognised as an author, and content generated autonomously by an AI system is not eligible for copyright protection. However, a human creator may obtain copyright protection for their own original intellectual contribution to a work produced with the assistance of AI.

 

Does AI training infringe copyright? Litigation before the Federal Courts

Far more complex questions are currently being considered by the U.S. federal courts in cases concerning AI model training. These disputes are of fundamental importance to the generative AI industry because they seek to determine whether copyright-protected works may lawfully be used for training AI models without the consent of rightholders and without obtaining a licence.

One of the most closely watched proceedings is The New York Times Company v. Microsoft Corporation and OpenAI, Inc., pending before the United States District Court for the Southern District of New York. The New York Times alleges that the defendants unlawfully used millions of its news articles to train GPT models and AI systems built upon those models. The litigation extends beyond the legality of the training process itself. It also concerns the phenomenon of model memorisation, namely the ability of AI models to reproduce portions of their training data. According to the claimant, such capabilities enable AI systems to compete directly with the publisher’s business by providing users with content that substitutes for the original articles. The defendants, by contrast, argue that the use of copyright-protected works for AI training constitutes fair use under U.S. copyright law.

A similar issue arises in Bartz et al. v. Anthropic PBC, pending before the United States District Court for the Northern District of California. A group of authors alleges that Anthropic used their books to train the Claude language model without obtaining the necessary licences. Among the issues before the Court are the legal significance of the source from which training data was obtained and whether materials acquired from unlawful sources may nonetheless fall within the scope of the fair use doctrine. Comparable questions are also being examined in Kadrey et al. v. Meta Platforms, Inc.

The claimants allege that Meta used copyright-protected books to train its Llama family of large language models without authorisation. Another significant case is Andersen v. Stability AI Ltd., Midjourney Inc., and DeviantArt, Inc. In that litigation, a group of visual artists alleges that the defendants unlawfully used copyright-protected artworks to train generative AI models, including Stable Diffusion.These proceedings concern not only the use of protected works during model training but also the broader responsibility of developers of generative AI systems for the impact of those systems on authors, artists and the creative industries.

All of these proceedings revolve around the same fundamental question: can the development of AI models justify the use of vast quantities of copyright-protected works without the permission of their creators? Within the U.S. legal system, the answer will depend largely on the application of the fair use doctrine. Unlike the more specific text and data mining (TDM) exception under EU law, fair use is a flexible doctrine requiring courts to assess each case on its individual facts. The eventual decisions of the U.S. federal courts are therefore likely to establish principles that will influence not only the American market but also the global development of generative artificial intelligence.

Conclusion

The development of artificial intelligence is compelling legislators and courts alike to define the legal boundaries of intellectual property protection and liability from the deployment of AI systems. The case law has focused primarily on four interrelated issues: the legality of using copyright-protected works to train AI models, the copyright status of AI-generated content, the phenomenon of model memorisation, and the liability of AI developers and providers for the outputs generated by their systems. Cases heard by European and American courts show that, despite differences between legal systems, similar questions are arising regarding the relationship between the development of AI and the protection of creators’ rights. While the rulings to date do not yet establish a uniform standard, they do set the direction for the further development of AI law.