Gemini’s live news answers may infringe publisher rights, EU court adviser says

The story

A senior adviser to the Court of Justice of the European Union has proposed a legal test that could make live, grounded AI answers—not only model training—a central copyright risk for chatbot providers. In an opinion delivered on November 3, Advocate General Maciej Szpunar said an AI chatbot may infringe the EU’s related right for press publishers when it retrieves an online publication, reuses it in an answer and supplies that answer to a user under particular circumstances. The opinion concerns Google Gemini and is the first generative-AI case to reach the EU’s highest court.

The qualification is essential: this is not a judgment that Google infringed copyright. Advocate generals independently recommend legal solutions to the judges, and their opinions are not binding. The Court has now begun deliberations in Case C-250/25, Like Company v Google Ireland, and will issue a judgment later. The underlying dispute remains for the Budapest Regional Court to decide in accordance with the EU court’s eventual interpretation.

Like Company, a Hungarian operator of online portals, says Google used its articles to train Gemini and reproduced or made them available through the chatbot between June 2023 and May 2024. The Hungarian court referred four questions to Luxembourg. Two concern training; two concern answers produced after Gemini uses retrieval-augmented generation, or grounding, to collect current information from the web. That distinction lets the opinion separate the creation of a model from a later system action directed at a particular publication.

Szpunar advised the Court to reject the training questions as inadmissible because the national case contained no evidence that Like Company’s articles trained Gemini and because LLM training is not an ‘online use’ covered by the press-publisher right at issue. For completeness, he also suggested that using protected material to train an LLM is, as a rule, permitted by the EU text-and-data-mining exception. That statement is important but should not be reported as a final, universal ruling on European AI training; it appears in a non-binding opinion after the relevant questions were deemed hypothetical.

The live-answer analysis is more concrete. The publisher right does not create ownership of news facts. It protects the investment embodied in a publication’s specific form of expression. Szpunar proposed examining whether the chatbot accessed an article unlawfully—for example, by circumventing an effective paywall or anti-scraping measure—and whether its answer merely conveys facts or recognizably reuses the publication’s expression. An answer becomes more legally exposed when it reproduces that expression so fully that the user no longer needs to open the original article.

He also distinguished reproduction from making content available to the public. The latter right would generally be infringed, in his view, when a publication’s expression appears systematically and recognizably across chatbot answers to related prompts. Potential liability can rest with the chatbot provider when such use is inherent in its business model or when the provider has failed to deploy credible and effective countermeasures. That approach connects copyright analysis to product architecture: retrieval controls, access signals, output filters and the repeatability of responses could all become evidence.

Reuters reported that Google emphasized the other side of the opinion. The company said EU text-and-data-mining exceptions are important for competitiveness, that Gemini learns linguistic patterns to create original outputs, and that publishers have controls to opt out. The European Publishers Council argued that indexing for search should be distinguished from AI answers that can substitute for original reporting. Those positions expose the commercial issue beneath the doctrine: publishers depend on discovery traffic while AI interfaces increasingly answer without requiring a click.

INNOVOX analysis: the opinion’s most consequential move is to treat grounding as its own auditable use rather than as an invisible extension of training. A model may have been lawfully trained yet still generate a problematic answer after retrieving a protected article in real time. Conversely, an answer that independently reports unprotected facts need not infringe merely because it is useful. The difficult boundary is ‘recognizable’ expression and substitution. Courts will need evidence about retrieved passages, prompt-response patterns, citations, access controls and whether safeguards work repeatedly rather than only in curated demonstrations.

What to watch next is the Court’s wording. If the judges adopt this framework, AI providers serving Europe may need stronger separation between search indexing, model training and answer-time retrieval, plus records showing which source was fetched and how much expression survived into the output. Publishers will still have to prove more than lost traffic or factual overlap. The final judgment must clarify the legal threshold, after which the Hungarian court will apply it to Gemini’s actual conduct. Until then, the opinion is a significant roadmap—not a finding of liability.

INNOVOX analysis

The opinion separates two technical stages that are often treated as one copyright problem: model training and retrieval at answer time. Its most consequential idea is that a grounded chatbot’s output can be assessed as a publication-use event, with provider liability linked to business design and effective safeguards rather than only to the model’s training history.

What to watch

Watch the Court’s final judgment, especially its treatment of recognizable expression, paywalls and publisher opt-outs. Product teams should also watch whether the Court defines what safeguards can credibly prevent substitution and whether the Hungarian court finds Gemini’s actual answers met that threshold.