AI chatbots and the press publishers’ right: the two tests proposed by the CJEU Advocate General

On 3 November 2026, First Advocate General Maciej Szpunar delivered his opinion in Case C-250/25, a dispute between the Hungarian publisher Like Company and Google Ireland over the Gemini chatbot. The Court of Justice of the European Union (CJEU) calls it, in its press release No 155/26, “the first generative AI case to reach the Court of Justice”. More regulation stories are in the Society & Law section.

What the Advocate General proposed

Like Company runs several news portals. Before the Budapest Regional Court, it claims that Google used its articles to train the Gemini models and repeatedly reproduced and made them available through the chatbot between 13 June 2023 and 31 May 2024, according to the press release. The Hungarian court referred four questions to Luxembourg.

On the two questions about training, the Advocate General advises the Court to declare them inadmissible: nothing shows that Like Company’s articles were used to train Gemini, and training is not an “online use” of a press publication, because it happens on dedicated infrastructure without any service being provided to a user (points 52 to 57 of the opinion). For completeness, he adds that training is, “as a rule”, permitted under the text and data mining (TDM) exception, as Google and the Estonian and French governments had argued (point 101).

The two questions on what the chatbot sends back to users are the real crux. The publishers’ right “does not protect the mere facts reported in press publications, but only the specific form of expression of those publications”, the Advocate General stresses, as summarised in the Court’s press release: it protects an investment, not a monopoly over the news.

Training versus real-time retrieval: two regimes

Training adjusts a model’s parameters upstream, on a large corpus. Retrieval-augmented generation (RAG), or grounding, happens at each query: the chatbot fetches web pages, often news articles, to write its answer. The Gemini models use this technique, according to the press release.

Source: Advocate General’s opinion and Directive (EU) 2019/790, read on 11 November 2026.

Under Article 4(3) of the directive, the opt-out must be expressed “in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. The Advocate General lists robots.txt, Google’s Google-Extended token, file metadata and even website terms in natural language, provided a machine can understand them (point 149 of the opinion). He also points to the AI Act, which requires providers of general-purpose models to publish a summary of their training content (point 150); the dates of those obligations are in our EU AI Act timeline.

The two tests for chatbot answers

In his proposed answer (point 240 of the opinion), when a chatbot retrieves an article from a publisher’s website to answer a query, there is reproduction if either condition is met:

• unlawful access: the chatbot circumvented effective technical measures designed to restrict access to, or scraping of, the article, such as a paywall or a bot block;
• recognisable form: the article’s form of expression is recognisable in the answer, “to such an extent that the user, by reading that answer, can dispense with accessing the original publication entirely” (opinion, point 240). Rephrasing is no escape (point 205); an answer that only reports the facts is not caught (point 208).

Making available to the public is, as a rule, infringed where that form is “systematically recognisable” in answers to related queries, according to the press release. The provider is liable if these acts are inherent to its business model, or if it failed to take the steps expected of a reasonably diligent operator. Point 222 of the opinion spells them out: programme the RAG bots to respect publishers’ blocks, and keep answers sufficiently removed from any single article, with output filters for instance.

He rejects implied consent: free access is not agreement to feed a chatbot, especially as Google’s bots appear to serve both indexing and grounding (points 225 to 228). In the case itself, subject to the Hungarian court’s findings, Like Company had no technical measures, its Gemini answer showed no recognisable form, and Google seems to have taken appropriate measures (point 239 of the opinion).

Why it reaches beyond Hungary

The answer will interpret EU law for all member states. Nine governments filed written observations, alongside the European Commission (point 21 of the opinion). On paraphrase, the French, Italian, Danish, Spanish, Polish, Hungarian and Greek governments argued with Like Company that the right would be too easy to circumvent if only verbatim copying counted (point 205).

Location does not shield a provider either. For real-time answers, Article 15(1) can protect publishers whenever the chatbot is offered to users in the EU, “irrespective of the place where the chatbot provider is established or of the location of the servers” (point 165 of the opinion).

This is copyright, not data protection: objecting to the use of your personal data for training follows other rules, set out in our guide on AI training and the GDPR.

How France already handled the issue

France transposed Article 15 through the loi n° 2019-775 du 24 juillet 2019 (in French). Article L. 218-2 of the Code de la propriété intellectuelle, France’s intellectual property code, requires the publisher’s authorisation before an online service reproduces or communicates its publications. Article L. 211-3-1 leaves hyperlinks and very short extracts free, but, in its original wording, not when their use “se substitue à la publication de presse elle-même ou dispense le lecteur de s’y référer”, replaces the publication or spares the reader from consulting it (our translation). The second test echoes that idea.

For training, the ordonnance n° 2021-1518 du 24 novembre 2021 (in French) created article L. 122-5-3: text and data mining is open to anyone unless the author has objected “de manière appropriée, notamment par des procédés lisibles par machine”, appropriately, in particular by machine-readable means (our translation). The same text extends the rule to press publications.

The Autorité de la concurrence, France’s competition authority, has issued four decisions on Google and publishers’ rights in four years, as its press release of 20 March 2024 recalls:

Source: Autorité de la concurrence, read on 11 November 2026.

Decision 24-D-03 (in French) also covers Bard, launched in France in July 2023. Google used news content for training, grounding and displaying answers without telling publishers; until at least 28 September 2023, when Google-Extended launched, opting out meant blocking all Google crawling, Search and Google News included. But the authority did not rule on the substance: “The question of whether the use of press publications as part of an artificial intelligence service qualifies for protection under related rights regulations has not yet been settled”, its press release says. That is the question now before the CJEU.

French publishers are also in court. The Alliance de la presse d’information générale, a trade body of general news publishers, and 53 of its members sued Brave before the Paris judicial court for infringing the publishers’ right and their trademarks, according to its press release of 1 June 2026 (in French), cited by the Advocate General in footnote 170.

What publishers and users can do now

If you run a news site in the EU, separate two moves. For training, the machine-readable opt-out remains the key: Google-Extended controls the training of future Gemini models and grounding in the Gemini apps, without affecting Google Search, according to Google’s documentation, which adds that Google intends to support the shared AIPREF training signal in early 2027. Our article on that signal (in French) has the details. For real-time answers, robots.txt may not be enough: the protocol “does not technically prevent scraping”, the Advocate General notes (footnote 153 of the opinion). The first test targets effective technical measures.

The second test has to be proven, and “a single example will rarely be enough” (footnote 173 of the opinion). A publisher that wants to act should log series of queries, with the date, the chatbot and the full answer, to show recognisable and repeated reuse of its articles, not just of their facts.

If you use a chatbot to follow the news, Article 15 of the directive excludes private or non-commercial uses by individual users: the opinion targets the provider, not you. On AI Overviews in Google Search, a separate front, see our article on the publishers’ complaint to the European Commission (in French).

What is still open

“The Advocate General’s Opinion is not binding on the Court of Justice”, the press release states: the judges are deliberating and no judgment date has been announced. The Budapest court will then decide the dispute, and the ruling will bind other national courts facing a similar issue. As of 11 November 2026, the opinion, written in English, is on EUR-Lex as a provisional text. Our editorial method explains how sources are checked.