Vinge Insights

When does generative AI infringe copyright? Insights from the Advocate General's Opinion in Like Company v Google

5. Patent 16.9

Background

On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Case C‑250/25, Like Company v Google Ireland. This case is the first in which the Court of Justice of the European Union (the “CJEU”) will rule on how copyright rules relate to generative AI, and it has attracted significant interest from both rights holders and the AI industry.

The background to the case is a dispute in a Hungarian court between Like Company, which operates a news portal, and Google. Like Company alleges that Google, through its AI system “Gemini”, has infringed the company’s exclusive right to its press publications – both by using the articles without consent to train the language models that power Gemini, and by partially reproducing the articles in the system’s responses, such as when a user requests a summary of an article.

The questions referred by the Hungarian court to the CJEU concern the permissibility of using press publications to train the language models that power an AI system (“AI models”) (Questions 2 and 3), and whether it constitutes an infringement of the exclusive right to press publications when parts of such a publication are reproduced in an AI system’s response to the user (Questions 1 and 4).

Summary of the Advocate General’s Opinion

Questions concerning the use of press publications in the training of AI models

Advocate General Szpunar considers that the CJEU should not answer Questions 2 and 3 regarding the training of AI models, as they are hypothetical in the present case. According to Szpunar, Like Company has not demonstrated that the company’s press publications were in fact used to train Gemini. Szpunar notes that the AI system could likely have generated summaries of the material without such training, since the AI system can also retrieve information from the web in real time, i.e., after the user has asked a question but before the answer is generated.

Nevertheless, Szpunar sets out his views on these issues without intending to provide a definitive answer. In this way, Szpunar seeks to establish a framework for the debate and draw the CJEU’s attention to the fact that it should proceed with caution in its adjudication of the case. Szpunar notes that the trained model does not itself contain any actual copies of the training material (i.e. the model cannot “memorise” a work from the training stage) but, rather, consists of abstract statistical relationships between words. That said, Szpunar is of the opinion that a form of reproduction occurs during training. Therefore, the training of generative AI as such falls, as a starting point, within the scope of the rights holders’ exclusive right of reproduction and requires their consent, unless an exception applies.

Accordingly, the central issue is whether the exception for text and data mining (the “TDM exception”) applies. The exception permits the copying of lawfully accessible material for automated analysis for the purpose of extracting information such as patterns and correlations, provided that the rights holder has not reserved its rights in an appropriate manner. Although the exception was introduced with traditional data analysis in mind, Szpunar believes that it, in principle, covers the training of AI models, since the material is then analysed automatically to extract the statistical relationships that become the model’s parameters. This presupposes that the material was lawfully accessible, that the rights holder has not reserved its rights, and that the AI developer has taken appropriate measures to ensure that the model does not generate results that are identical to the training material. According to Szpunar, the so-called “three-step test” - which stipulates that exceptions to copyright may only apply in certain special cases that do not conflict with the normal exploitation of the works or unreasonably prejudice the interests of rights holders - does not prevent such an application of the TDM exception.

Questions concerning whether the reproduction of press publications in an AI system’s response to a user constitutes infringement

With regard to the questions concerning Gemini’s output, Szpunar believes that unauthorised reproduction of a press publication may occur if an AI system retrieves information from the publication in real time and subsequently reproduces parts of it in its response to the user. Szpunar argues, however, that no infringement occurs if (i) the AI system has had lawful access to the publication – meaning the system has not circumvented a technical restriction or a paywall – and (ii) the AI system’s response reproduces only the factual content of the publication rather than its form of expression.

In the Opinion, Szpunar also clarifies under what conditions AI providers can be held liable for the reproductions carried out through their AI systems. The basic principle is that the user is primarily liable when the AI system reproduces an article at the user’s instruction. However, the AI provider may also be held liable if the infringement is inherent in the provider’s business model, or if the provider has not taken the measures that can be expected of a reasonably diligent operator to prevent infringement – for example, through filters which ensure that the responses do not resemble the retrieved articles too closely.

Some reflections on the implications of the Advocate General’s Opinion

Although Szpunar’s Opinion is not binding, and it remains to be seen how the CJEU will choose to answer the Hungarian court’s questions, the Opinion offers several interesting indications of how the copyright issues raised by AI development may be addressed.

To begin with, Szpunar advocates a middle ground regarding infringing output that may be considered relatively lenient towards AI providers: even though reproducing more than factual content may constitute infringement, the provider is, as a starting point, not liable for isolated instances of infringement as long as it has taken reasonable measures to prevent them. It remains to be seen how this reasoning would apply to, for example, art or music, and whether the CJEU will follow the Opinion. However, the Opinion can be interpreted as an expression of concern that copyright and related rights will hinder AI development in Europe – a concern that Szpunar himself addresses when he highlights the importance of digital sovereignty and European AI models, and which likely reflects the fear that Europe will fall behind the United States and China in this regard.

If the CJEU follows the Advocate General’s Opinion and does not address the questions regarding training, the case risks being a disappointment for those hoping for clarifications. It is likely precisely these questions that are most important for both AI providers and rights holders, as they concern the very foundation of how a generative AI model is constructed.

Szpunar’s general reasoning regarding training does, however, offer a hint as to how the CJEU might view these issues. Of particular interest is his rejection of the idea that an AI model can “memorise” a work. In the ruling of the Regional Court of Munich (Landgericht München) in GEMA v OpenAI, the German court found that ChatGPT’s ability to reproduce song lyrics verbatim demonstrated that the texts had been memorised by the model, that this memorisation constituted an unauthorised reproduction, and that the TDM exception did not cover such copies. Under Szpunar’s view of memorisation, the same conclusions would not have been possible. Szpunar’s approach is closer to that taken by the High Court of England and Wales in Getty Images v Stability AI and is difficult to reconcile with the German court’s conclusions. The discrepancy between the courts’ interpretations of technical facts illustrates the complexity of the AI-related issues that the CJEU must address, and we look forward to seeing how the CJEU chooses to tackle these issues.