EU Court to Consider Lawfulness of AI Training on Copyrighted Works for First Time

03.06.2025

• Legal Framework
• Questions Referred to the EU Court
• Context: Recent Judicial Precedents
• Practical Recommendations
• Contact our lawyer for more details

Background

The Court of Justice of the European Union registered its first preliminary reference directly concerning the permissibility of using copyright-protected works in artificial intelligence training processes.

The case arose from a specific dispute between Hungarian publisher Like Company and Google Ireland Limited regarding responses from the Gemini/Bard chatbot to a specific query about Hungarian singer Kozsó and his plans to release dolphins into Lake Balaton. When a user asked: "Can you provide a summary in Hungarian of the online press publication that appeared on balatonkornyeke.hu regarding Kozsó's plan to introduce dolphins into the lake?", the chatbot provided a detailed response including information from the publisher's copyright-protected publications.

The case concerns the interpretation of Article 4 of Directive (EU) 2019/790, which establishes an exception for text and data mining (TDM) without rightholder consent, provided no machine-readable objection has been expressed.

The reference was made by the Budapest Regional Court (Budapest Környéki Törvényszék) in proceedings between the news publisher and Google Ireland Ltd. (operator of the Gemini/Bard chatbot).

In practice, this provision serves as the key legal basis relied upon by virtually all generative AI developers in the EU. Until now, it has not been officially interpreted by the CJEU, and this case will establish the first precedent.

Legal Framework

AI developers rely on Article 4 of Directive 2019/790, which permits text and data mining without consent if the author has not expressly prohibited this in machine-readable format (TDM opt-out).

However:

• The TDMRep (W3C) specification has limited adoption;
• Most content lacks prohibitions in the required format;
• Consequently, this creates a presumption of lawful training, even without authors' consent.

It is important to note that the EU AI Act directly references TDM exceptions from the DSM Directive in the context of training generative AI models. Article 53 of the AI Act requires providers of general-purpose models to "put in place a policy to respect Union copyright law in particular to identify and respect, including through state of the art technologies, the reservations of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790".

Questions Referred to the EU Court

• Does the display by a chatbot of fragments of protected text constitute communication to the public?
• Does the use of a work in machine learning processes constitute reproduction within the meaning of copyright law?
• If yes - does such reproduction fall under the exception provided for in Article 4 of Directive 2019/790 (TDM exception)?
• Is the AI operator liable if generated responses partially or completely reproduce media material in response to user queries?

Additionally, the court must determine whether it matters that responses result from a process whereby the chatbot merely predicts the next word based on observed patterns.

Context: Recent Judicial Precedents

Case C-250/25 is being considered against the backdrop of the first major judicial decision on TDM and AI in the EU. In September 2024, Hamburg Regional Court dismissed photographer Robert Kneschke's claim against LAION e.V., which had used his photographs to create datasets for AI training. The court held that LAION could rely on the scientific research exception (Article 3 DSM Directive), as dataset creation qualified as TDM for non-commercial scientific purposes.

However, the German court left open the question of applying the commercial TDM exception (Article 4), which is central to the case against Google.

Legal Risks and Precedential Significance of Case C-250/25 for the EU AI Industry

The preliminary ruling in Case C-250/25 could become a key precedent in the legal regulation of training AI models on copyright-protected works.

The EU Court's position will determine:

• Whether the machine learning process will be recognised as reproduction within the meaning of copyright law (Article 2 of Directive 2001/29/EC);
• Whether the TDM exception (Article 4 of Directive 2019/790) extends to training generative models;
• Whether the absence of machine-readable opt-out constitutes sufficient grounds for free use of works without rightholder consent;
• Who bears liability for generating content matching protected elements: the user or AI provider.

The decision will also affect interpretation of new obligations under the EU AI Act, including requirements for general-purpose model providers to provide "sufficiently detailed summaries of content used for model training" and comply with rightholders' opt-out mechanisms.

The consequences potentially affect both AI developers and rightholders (media, publishers, platforms): from content licensing terms to the need for implementing new technical and contractual control mechanisms.

Particular industry concern arises from uncertainty regarding what constitutes "machine-readable" expression of opt-out. Experts note that the absence of clear standards creates legal uncertainty for both rightholders and AI developers.

International Context

It is worth noting that approaches to TDM exceptions vary significantly between jurisdictions. Whilst Article 4 of the DSM Directive permits TDM for any purpose (with opt-out possibility), exceptions in the United Kingdom are limited to non-commercial research, and the United States relies on fair use doctrine. Japan and Singapore have broader exceptions similar to European ones.

The EU Court's decision may influence global standards, considering that many major AI companies provide services in the European market and must comply with European legislation.

Practical Recommendations

For AI companies and platform owners:

• Conduct audits of data sources used for copyright protection;
• Monitor implementation or disregard of TDM opt-out (including TDMRep specifications);
• Consider the possibility of judicial precedent limiting existing practices;
• Assess whether your activities require adjustment of data collection and processing models;
• Prepare for compliance with new EU AI Act requirements on training data disclosure and implementation of opt-out signal recognition technologies.

For rightholders and publishers:

• Explore technical implementation possibilities for machine-readable opt-out mechanisms;
• Consider including TDM clauses in licensing agreements and terms of use;
• Monitor development of TDMRep standards and other technical solutions for expressing copyright reservations.

Our Arbitration & IT Disputes team is prepared to help you assess legal risks of using copyrighted content in AI applications in light of Case C-250/25.

Authors: Kamal Tserakhau, Krystsina Vainilovich