What happened
The case, C-250/25, Like Company, began in Hungary. Like Company, which operates web portals, brought proceedings against Google before the Budapest Regional Court. It claims protection under the EU-related right granted to press publishers over online use of press publications. The claim concerns Google Gemini, a generative chatbot that can converse through text or speech and whose language models use retrieval-augmented generation (RAG), also described as grounding, to collect web information in real time. The publisher alleges that Google used its articles to train Gemini’s models and repeatedly reproduced and made them available through the chatbot between 13 June 2023 and 31 May 2024.
The Hungarian court referred four questions to the CJEU. Two ask whether using press publications to train a language model falls within publishers’ exclusive reproduction right or is permitted by the EU text-and-data-mining exception. The other two concern whether chatbot answers can infringe the reproduction right or the right to make a publication available to the public.
Szpunar advises the Court to declare the training questions inadmissible. The release says there is no evidence that Like Company’s articles were used to train Gemini, and that training a model is not itself an “online use” of press publications. He therefore considers those questions hypothetical. For completeness, he suggests that using protected works to train a model is, as a rule, freely permitted under the text-and-data-mining exception. This is a view offered to the Court, not a judgment resolving the issue in the case.
On chatbot answers, the opinion says a publisher’s right may be infringed where a chatbot retrieves a publication, reuses it to generate an answer, and provides that answer to a user. The relevant right does not protect the facts reported in an article as such; it protects the specific form of expression, with the aim of protecting publishers’ investment rather than giving them a monopoly over news. The proposed assessment has two main factors: whether the chatbot accessed the article unlawfully, including by circumventing effective measures such as a paywall or restrictions on scraping; and whether the answer merely reports facts or recognisably reuses the article’s expression to the extent that the user can forgo reading the original. Systematic recognisability in answers to related queries may also infringe the right to make the publication available.
The opinion says a chatbot provider may be liable for potential acts of use when those acts are inherent to its business model or when it has failed to put appropriate, credible, and effective counter-measures in place. The Judges are beginning deliberations; the CJEU will give judgment later. The Advocate General’s opinion is not binding, and the Court does not decide the underlying national dispute itself.
What the archive adds
This is a new fact in a recurring publisher-rights story, not a final ruling. Our earlier note on USA Today’s suit against OpenAI covered a complaint focused on training-data use and model output. The present opinion adds a proposed EU framework for real-time retrieval and answer reuse, while treating the training questions in this particular case as hypothetical.
Trend signal
The opinion supports the pattern Agent-ready software exposes capabilities only weakly: it makes permissioned access and technical controls around web content consequential for a deployed chatbot, but does not prescribe a product architecture. The more direct signal is that agent-like systems’ access to external services may be governed by observable restrictions, rather than prompts alone.
Why now
The opinion was delivered on 3 November 2026 as the first generative-AI case reached the CJEU. It responds to questions referred by the Hungarian court in Like Company’s dispute with Google. The timing reflects the case’s progress to the Advocate General’s opinion; the source gives no timetable for the Court’s eventual judgment.
Why it matters
Teams building retrieval-grounded products should treat content acquisition and answer generation as separate risk surfaces. A model’s training status does not settle whether a live system can lawfully fetch a particular page or how closely an answer can reproduce its expression. The opinion points developers toward controls that affect the full retrieval path: access restrictions, anti-scraping measures, and response behavior that avoids substituting for the source article.
That is a practical design signal, not a settled compliance checklist. The opinion is not binding, and liability depends on circumstances the Court has yet to decide. Still, teams can map where a product retrieves protected material, what permissions or restrictions apply, and whether its answers may preserve recognisable wording or structure. Publishers, meanwhile, have a proposed test that distinguishes reporting facts from reproducing expression and considers whether users are diverted from the original.
Limits and uncertainty
This is an Advocate General’s opinion, not the CJEU’s binding judgment; the Court may not follow it. The source says there is no evidence in this case that Like Company articles trained Gemini, so the training questions are hypothetical. It does not establish that any particular Gemini answer infringed rights, define the technical threshold for recognisable reuse, or provide a complete compliance standard for RAG products. The national court must resolve the dispute after the CJEU rules on the referred questions.
Reader takeaway
The opinion’s useful distinction is between training a model and retrieving an article to answer a user. For retrieval products, the proposed infringement analysis turns on how the content was accessed and whether the answer recognisably reproduces the article’s expression enough to replace the original. Treat this as a significant legal signal, not settled law, until the Court issues its judgment.
Open question
Will the CJEU adopt Szpunar’s proposed distinction and clarify how courts should assess whether a chatbot answer recognisably reproduces an article’s expression?