Hungary – Manchester United Fanatics Klub Magyarország v Google Ireland Limited, 20 July 2026

Claimant: Manchester United Fanatics Klub Magyarország (a Hungarian sports association and publisher of the Manchester United supporters’ news portal www.manutdfanatics.hu).

Defendant: Google Ireland Limited.

The date on which the national proceedings were initiated is not specified. The order for reference is dated 9 July 2026 and the request for a preliminary ruling was lodged on 20 July 2026.

The claim concerns the period from 13 July 2023 to 26 November 2025.

The Claimant operates www.manutdfanatics.hu, a periodically updated news portal focusing on Manchester United. It has editorial responsibility for the website, which publishes primarily journalistic and literary content. The Defendant, part of the Alphabet group, provides the generative AI chatbot Gemini (previously Bard), which generates responses to users’ prompts.

According to the technical account set out in the reference, a large language model (LLM) is trained on large datasets, including publicly available web pages and licensed sources. During training, text is divided into numerical tokens and analysed for statistical relationships; the model does not store legible copies of the original training documents. The foundation model underlying Gemini was trained in the United States, without using hardware infrastructure in Hungary.

Gemini can also use a process described as “grounding”, in which it temporarily consults indexed web content through Google Search to obtain factual, up-to-date information, and subsequently clears the extracts from its intermediate memory. Direct quotations may be accompanied by quotation marks and a source link. The Defendant states that a filter is designed to prevent identical reproduction of existing text but accepts that protected content may occasionally appear in responses. The referring court records that the Defendant’s sworn description of Gemini’s functioning is undisputed.

Google-Extended, a mechanism allowing publishers to prevent use of their content for Gemini training without excluding it from Google Search, was introduced on 28 September 2023. Before then, the Claimant could not block AI training separately from Google Search through the robots.txt mechanism. The Defendant does not dispute that Gemini could have accessed the Claimant’s content during training. The Claimant does not currently use Google-Extended and continues to allow technical access for training.

The Claimant alleges that Gemini’s responses can convey sufficient information from its publications to remove the need to visit the original website, thereby depriving it of advertising revenue used to finance the publications.

Claimant’s claim. 

The Claimant maintains that Article 15 of Directive 2019/790 gives press publishers rights requiring authorisation, and remuneration, for the relevant online use of their publications. It argues that Gemini’s responses can reproduce the substantive content of its news reports, even where they are generated as responses or summaries rather than verbatim copies. The resulting loss of visits to the source website undermines its advertising-based business model.

The Claimant further argues that its consent cannot be inferred merely from the accessibility of its publications. Referring to the CJEU’s judgment in Soulier and Doke (C-301/15), it submits that the Defendant did not adequately inform it about the intended uses of its content and therefore cannot rely on implied consent.

Defendant’s defence. 

The Defendant denies that LLM training and the conversion of text into tokens constitute reproduction, and notes that the training took place outside Hungary. It argues that the probabilistically generated responses do not amount to reproduction or making available to the public. Relying on Svensson (C-466/12), it contends that no different technical means or “new public” is involved.

In the alternative, the Defendant invokes the exceptions for temporary acts of reproduction (Article 5(1) of Directive 2001/29/EC) and text and data mining (Article 4 of Directive 2019/790). It also relies on the statutory exclusion for “individual words or very short extracts” from press publications under Paragraph 82/C of the Hungarian Copyright Law.

1. Making available to the public. Whether an LLM-based chatbot makes a press publication available to the public under Article 15(1) of Directive 2019/790 and Article 3(2) of Directive 2001/29/EC when it generates a new synthesis or summary from which the original publication and its content can be recognised.

2. Reproduction during training. Whether training constitutes reproduction under Article 15(1) of Directive 2019/790 and Article 2 of Directive 2001/29/EC when a protected press publication is transformed into tokens and statistical relationships rather than stored as original text.

3. Reproduction when processing prompts. Whether reproduction occurs when a user includes a protected press publication in a prompt and the chatbot processes it in order to respond.

4. Text and data mining exception. If training amounts to reproduction, whether it falls within Article 4 of Directive 2019/790 where the trained model is subsequently used to operate a chatbot that responds to user instructions.

5. Consent by non-prohibition. Whether a publisher that has not expressly prohibited use of its publications on the service provider’s online platform can be considered to have consented to their use for chatbot training, tokenisation and grounding.

6. Reservation of rights. Whether failure to expressly and appropriately reserve rights under Article 4(3) of Directive 2019/790 amounts to authorisation of the use of the press publication for an LLM-based chatbot.