A blog post by Victor Hugo Delfante Borborema
The first CJEU Opinion on generative AI and copyright draws an important line between training a model and retrieving articles to answer a user. That distinction may shape how publishers and AI providers argue over access, control and value.
On 3 November 2026, Advocate General Maciej Szpunar delivered an Opinion in Like Company v Google (C250/25), a reference from a Hungarian court concerning Google Gemini. The case asks whether the use of press publications by an AI chatbot can infringe the rights EU law grants to press publishers.
The Opinion’s central move is to separate two activities that are often bundled together in public debate: using material to train a large language model (LLM) and retrieving and reusing an online article to generate an answer. They raise different factual and legal questions.
The case concerns publishers’ related rights
Like Company operates Hungarian web portals and claims that Google used its articles in connection with Gemini. The relevant right is the press publishers’ related right under Article 15 of the Copyright in the Digital Single Market Directive (CDSM Directive). That right covers certain online uses of press publications by information society service providers. It protects publishers’ investment in producing publications; it does not give them ownership of the facts reported in an article.
The reference sent four questions to the Court. Two concern whether Like Company’s articles were used to train Gemini and whether text and data mining (TDM) rules apply. The others concern whether the chatbot’s answers reproduce or make press publications available to the public.
AG Szpunar considered the training questions hypothetical because the case file did not establish that Like Company’s articles had been used to train Gemini. He also reasoned that LLM training, as such, is not an “online use” of press publications within the meaning of Article 15. He nevertheless discussed the TDM exception in the alternative, suggesting that use of protected material to train an LLM is, as a rule, permitted under that exception.
That part of the Opinion should be read with care. The proposed answer does not resolve every question about copyright, training data or the scope of the TDM rules. It addresses the questions referred and the particular press publishers’ right at issue.
Retrieval raises a different question
The more directly relevant issue in the proceedings is retrieval-augmented generation, or RAG. A chatbot using RAG can retrieve material from the web in real time and use it to produce a response. In the Advocate General’s view, that process may involve an “online use” of a press publication, even though it is distinct from the earlier training of the model.
The Opinion proposes a specific way to assess possible reproduction. The national court should consider whether the chatbot accessed the article unlawfully—for example, by circumventing effective technical measures restricting access or scraping—and whether the answer recognisably reuses the article’s form of expression. Similar facts alone would not suffice. The question is whether the answer reproduces distinctive expression to such an extent that a user can obtain the substance of the article without visiting the original publication.
The Opinion also addresses the press publishers’ separate right to make publications available to the public. It suggests that this right may be infringed where the article’s expression is systematically recognisable in answers to related queries. A provider’s responsibility may depend on whether such reuse is inherent in its business model or whether it has failed to adopt credible and effective safeguards.
These are proposed legal tests, not findings that Gemini infringed Like Company’s rights. The Advocate General’s Opinion is non-binding; the Court of Justice will decide the EU-law questions, and the Hungarian court will resolve the dispute.
A useful distinction—and a contested framing
The European Copyright Society (ECS), in a February 2026 comment on the reference, had urged caution about the way the questions were framed. It argued that the reference blurred the differences between a chatbot, an LLM and a search engine, and between model training and the use of content in chatbot responses. The Opinion’s focus on RAG and the evidential gap around training responds to some of those concerns.
The ECS also warned that broad copyright rules could entrench a licensing economy
dominated by large AI providers without necessarily resolving questions of creator consent or fair remuneration. That concern points to a wider policy problem: copyright may protect particular forms of expression, while competition law and platform regulation address questions about access, bargaining power and market structure. Those tools interact, but they do not do the same work.
Like Company highlights how AI-generated answers may shift value between publishers and platforms. If users get what they need from a chatbot response without visiting the source article, publishers may lose traffic and advertising revenue while AI providers continue to draw on their content. AG Szpunar’s proposed framework makes the details matter: whether the system trained on an article or retrieved it in real time, whether the answer repeats facts or recognisably reuses the article’s expression, and what evidence establishes that use.
The dispute also raises a broader question about power in AI-enabled information markets: whether copyright can address publishers’ concerns about access and bargaining, or whether those concerns also call for competition-law analysis. The Opinion offers a way to assess the copyright questions, but it is non-binding. The Court’s judgment will determine which parts, if any, become authoritative EU law.
Sources: AG Szpunar’s Opinion in C250/25; CJEU press release No 155/26; European Copyright Society comment on the reference.