On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Case C-250/25, Like Company v Google Ireland. It is the first generative AI case to reach the Court of Justice of the European Union.
The Opinion concerns Google Gemini's retrieval and reuse of online press publications when generating answers to user questions. The Advocate General proposes that real-time retrieval and reuse may infringe press publishers' rights in certain circumstances.
The legal status is important. An Advocate General's Opinion is independent advice to the Court and is not binding. The judges are now considering the case, and judgment will be delivered later. The Opinion does not itself establish liability for Google or a general rule that every AI-generated news summary infringes copyright.
The case separates training from retrieval
Like Company operates Hungarian news portals. It alleged that Google used its articles to train Gemini and reproduced or made them available through chatbot answers. The Hungarian court referred questions about both LLM training and answers generated through real-time retrieval.
The Advocate General advises the Court to declare the two training questions inadmissible because there was no evidence that Like Company's articles had been used to train Gemini and because LLM training was not an online use of press publications within the specific press-publisher right being examined.
For completeness, he also expresses the view that use of protected material to train an LLM is, as a rule, permitted under the text-and-data-mining exception in Article 4 of the Digital Single Market Copyright Directive. That statement forms part of a non-binding Opinion addressing hypothetical questions in this case. It is not a final CJEU ruling on all AI training, datasets, rights reservations or other copyright claims.
The proposed test for real-time chatbot answers
The Advocate General focuses on retrieval-augmented generation, where a chatbot accesses current online material to improve an answer. He proposes that a press publisher's reproduction right may be engaged where the chatbot accessed the publication unlawfully, particularly by circumventing effective technical measures restricting access or scraping, and reused the publication's form of expression recognisably in its answer.
The distinction between facts and expression is central. The press-publisher right does not create a monopoly over news or reported facts. It protects the specific form in which the publication expresses them and the investment in producing the publication.
The proposed analysis also considers substitution. The Opinion asks whether recognisable reuse goes so far that the user no longer needs to access the original publication. For the making-available right, the Advocate General refers to the protected form of expression being systematically recognisable in answers to related user queries.
These are proposed interpretive criteria, not findings that Gemini satisfied them. The Advocate General's brief guidance indicates that, on the available material, Like Company did not appear to have used effective technological restrictions and that its expression did not appear recognisable in the example answer. The Hungarian court must determine the facts after the CJEU gives judgment.
Provider responsibility depends on product design and controls
The Advocate General proposes that a chatbot provider may be responsible where the relevant uses are inherent in its business model or where it has failed to implement appropriate measures that credibly and effectively counter them.
That approach directs attention beyond model architecture. Relevant evidence may include how the product accesses sources, responds to blocked content, quotes or paraphrases articles, attributes sources, handles repeated prompts and measures whether answers substitute for the original publication.
A general statement that the model predicts tokens would not resolve those factual questions. Product teams need evidence about the complete retrieval and answer pipeline, including search, access, extraction, prompt assembly, output controls and presentation to the user.
What AI businesses should review now
- Separate training, indexing, retrieval and answer generation in legal and technical records. Different stages may engage different rights and exceptions.
- Map which crawlers, retrieval services and agents access publisher websites, including the identity they present and the purpose for which content is used.
- Test compliance with paywalls, authentication, robots instructions and other technical measures. Do not design fallback routes that defeat effective access restrictions.
- Measure whether outputs preserve recognisable wording, structure or editorial expression rather than merely reporting facts.
- Test substitution risk by asking whether an answer provides so much protected expression that a user has little reason to visit the original publication.
- Create controls for quotation length, repeated extraction, source attribution, linking and responses to requests for full articles or close paraphrases.
- Keep versioned evidence of filters, blocked-source rules, evaluation results, complaints, corrections and mitigation effectiveness.
- Review supplier contracts for crawler behaviour, content licences, rights reservations, takedown support, audit evidence and allocation of infringement claims.
- Avoid treating the Opinion as a final judgment or as confirmation that all LLM training is lawful under Article 4.
- Prepare to reassess the product when the Court delivers judgment, particularly if it adopts, narrows or rejects the proposed test.
What press publishers should review
Publishers should document which material is public, licensed, paywalled or subject to technical restrictions. Access controls should be implemented consistently and tested rather than assumed to work.
They should preserve evidence of source articles, timestamps, access restrictions and chatbot answers. A claim based only on overlapping facts will differ from one showing recognisable reuse of protected expression.
The Opinion does not say that technical measures are the only way rights can arise, or that their absence authorises every use. It does show that access conditions, recognisable expression and the practical effect of the answer may become central evidence in this case.
The legal framework and the next step
Article 15 of Directive (EU) 2019/790 gives press publishers rights over certain online uses of press publications by information society service providers. Article 4 provides a text-and-data-mining exception, subject to its own conditions. Other copyright rights and exceptions may also be relevant depending on the material and conduct.
The CJEU has not yet ruled. Its eventual interpretation will guide the Hungarian court and other national courts dealing with similar EU-law questions. Until then, the Opinion is a significant legal signal, not binding law.
For AI founders, the immediate compliance lesson is to distinguish the source of an answer from the wording of an answer. Lawful access to facts, access to protected expression and reuse that substitutes for the original are not the same question.
Legal position and sources reviewed on 4 November 2026. This article provides general information only and does not constitute legal advice on any particular model, publication, dataset or retrieval system.
This article provides general information and does not constitute legal advice on a specific matter. Legal obligations depend on the facts and the applicable law.