Hungary’s trio of AI referrals on the press publishers’ right

When the Budapest Környéki Törvényszék (the Budapest District Court) referred Like Company v Google Ireland (C-250/25) in April 2025, the case was hailed as the first reference to the CJEU on generative AI and copyright. It no longer stands alone. In July 2026 two further references were lodged: Manchester United Fanatics Klub Magyarország v Google Ireland (C-806/26) and Mobilguru v Google Ireland (C-829/26). The three cases have a lot in common. All come from Hungarian courts. All are brought by Hungarian online publishers against Google Ireland over Gemini (formerly Bard), a chatbot built on a large language model (LLM). And all relate to the press publishers’ right in Article 15 of the Directive 2019/790 (CDSM Directive), read together with Articles 2 and 3 of the Directive 2001/29/EC (InfoSoc Directive) and the text and data mining (TDM) exception in Article 4 of the CDSM Directive.

Like Company (C-250/25)

The publisher and operator of various Hungarian news portals claims that Gemini reproduced its press publications and made them available to the public without consent. It says the chatbot’s answers go well beyond the “individual words or very short extracts” that Article 15(1) leaves free. Google argues that Gemini’s responses do not reach a “new public” (rather, they are accessed by the same public, which is essentially all internet users). It also relies on the temporary copies exception in Article 5(1) InfoSoc Directive and the TDM exception under Article 4 CDSM Directive.

The four questions referred by the Budapest District Court ask whether displaying text partly identical to a press publication in a chatbot’s answer is a communication to the public, and whether it matters that the model “merely predicts the next word”. They also ask whether training an LLM is reproduction and, if so, whether Article 4 covers it. Finally, they ask whether the output generated in response to a user’s prompt is a reproduction by the provider. The Court heard the case on 10 March 2026, with Advocate General Szpunar.

In his Opinion, AG Szpunar said the Court should treat training and live retrieval (RAG) separately. On training, he called the questions inadmissible but said in the alternative that training copies generally count as reproduction and are usually covered by the TDM exception, while a trained model itself is ordinarily not a copy. On chatbot answers, a press publisher’s right is infringed only where an answer recognisably reproduces the publication’s expression so that users can skip the original, or where the content was accessed unlawfully, and he provisionally found neither on the facts against Google. You can read the analysis of the Opinion from João Pedro Quintais here, reviewed against European Copyright Society’s opinion on this matter posted here.

Manchester United Fanatics Klub (C-806/26)

The second reference comes from the Pécsi Törvényszék (Pécs High Court), decision of 9 July 2026. The applicant is a supporters’ association that runs manutdfanatics.hu, a regularly updated news portal on the Manchester United football club. It seeks a declaration that Google infringed its Article 15 rights through Gemini between July 2023 and November 2025.

The way Gemini works is not disputed. It was trained in the US on large web datasets and licensed data. It stores statistical patterns between tokens rather than copies of texts. It “grounds” its answers by briefly consulting the Google Search index and then discarding the extracts. It shows quotation marks and links when it quotes directly, and it has a filter meant to block output identical to existing third-party content, such as that of press publishers.

Google accepted that Gemini had access to the portal during training. It also conceded that protected content may still appear in answers occasionally, and that it keeps no record of prompts or responses.

Until Google-Extended was introduced on 28 September 2023, the applicant could not block AI training without also dropping out of Google Search. It did not use Google-Extended and so allowed both the defendant’s crawler and Gemini to access its press publication for training.

The applicant argued that Gemini’s answers were detailed enough that users no longer needed to visit the portal, so it lost the advertising revenue that funded it. In its view, this was a use requiring consent and remuneration under Article 15. Relying on Soulier and Doke, it added that Google never told it how its content would be used, so Google could not claim implied consent.

The six questions partly mirror Like Company but go further. Does training an LLM amount to reproduction where the publication is broken down into numerical tokens and statistical relationships that do not store the original content and, if so, is it covered by the TDM exception? Is a user’s prompt containing the publication a reproduction? Has the applicant that has not expressly reserved its rights under Article 4(3) CDSM Directive consented to training and grounding? Is there a “making available to the public” where the chatbot answers a prompt with a new text or summary, not previously in existence, from which the original press publication can be recognised as its source? If the Court follows AG Szpunar, Like Company will largely answer this last question. In his view, what matters is whether the publication’s form of expression remains recognisable, even if paraphrased, not merely whether the source can be identified.

Mobilguru (C-829/26)

The third reference comes from the Fővárosi Törvényszék (Budapest High Court), decision of 14 July 2026. Mobilguru publishes a Hungarian-language online magazine about mobile phones. It had not blocked Google through robots.txt. When it tested Gemini, the chatbot recommended Mobilguru as a source and gave, in English, information taken from its articles. Given the URLs of those articles, it produced English summaries and named Mobilguru’s article as the source. The applicant also pointed to the extended news preview (ENP) licences Google had previously concluded with publishers, meaning that it used copyright-protected works, without authorisation, when introducing and operating its AI services. Google answered that no reproduction took place, that there was no new public, that the answers showed only “mere facts” and very short extracts, and that Article 5(1) of the InfoSoc Directive and Article 4 of the CDSM Directive applied in any event.

Article 15 is again at the heart of the case, but the 11 questions (with many sub-questions) go above and beyond the previous two referrals. Article 15 applies to online uses by “information society service providers” (and the court asks whether exclusively so). The referring court first asks whether the “service” under Article 2(a) of Directive 2000/31/EC and Article 1(b) of Directive 2015/1535 covers only Gemini’s answers to prompts or also the training of the model. The referring court notes expressly that this distinction between development and operation was not raised in Like Company, although AG Szpunar has since touched on it in his Opinion.

The later questions turn to how Article 4 and Article 15 interact. What exactly may a TDM actor do without authorisation? Does that include communicating content to the public? What is the scope of the opt-out, and what counts as an “express reservation”? Is robots.txt required, or is a statement in the legal notice, on the home page or below the title enough? Can publishers be expected to use a provider’s technical opt-out that also blocks other uses? The question on the form of the reservation also brings in Article 53(1)(c) of the EU AI Act.

Why it matters

Together, the three references give us a fairly complete map of the issues: training, grounding and RAG, user prompts, outputs and opt-outs. All of them are framed through the perspective of the press publishers’ right in Article 15. That they all come from a single Member State is remarkable. Like Company is likely to be decided first (with the AG’s Opinion already out), but its four questions will not settle the entirety of the latter two referrals. Mobilguru asks whether the training and the operation of an AI model are subject to different legal regimes, a question not raised in the reference in Like Company, although AG Szpunar has now touched on it. Both July references go further on the opt-out under Article 4(3), and Manchester United also asks whether a publisher that has not expressly objected has consented. Hungarian courts, it seems, are determined to leave no stone unturned.

Photo by Zach M on Unsplash