First AI case at the EU Court: Advocate General says chatbot answers can infringe — training, as a rule, does not

On 3 November 2026 First Advocate General Maciej Szpunar delivered his opinion in Case C-250/25, Like Company — the first generative-AI copyright case to reach the Court of Justice of the European Union. His advice separates two things the AI debate usually blurs: training a model, and what the chatbot later shows the user.

The case

Like Company runs Hungarian news portals. It sued Google Ireland before a Budapest court, relying on the related right that EU law gives press publishers over the “online use” of their publications (Article 15 of the Copyright in the Digital Single Market Directive). It claims Google used its articles to train the models behind the Gemini chatbot, and reproduced and made them available to the public through Gemini between 13 June 2023 and 31 May 2024.

Gemini uses retrieval-augmented generation (“RAG”, or “grounding”): it fetches web content in real time, including news articles, to improve its answers. The Hungarian court referred four questions — two about training, two about the chatbot’s answers.

Training: inadmissible, and generally permitted

The Advocate General advises the Court not to answer the training questions at all. There is no evidence Like Company’s articles were used to train Gemini, and in his view training an LLM is not an “online use” of press publications, so the questions are hypothetical.

For completeness, he adds that using protected material to train an LLM is, as a rule, freely permitted under the EU text and data mining exception — subject to that exception’s conditions, which, as commentators note, include lawful access and respecting an opt-out where rights have been reserved.

Answers: where infringement can occur

The substance of the opinion is about the answers. When a chatbot retrieves an article from a publisher’s site, reuses it to generate a response and gives that response to the user, the publisher’s right “may be infringed in certain circumstances”.

The starting point is that the right protects the form of expression, not the facts: its purpose is to protect the investment in producing a publication, “not to give publishers a monopoly over the news itself”. Within that limit, the Advocate General identifies two factors for the reproduction right:

• Unlawful access — whether the chatbot obtained the article by circumventing effective technical measures, such as a paywall or measures restricting web scraping.
• Recognisable substitution — whether the answer merely reports facts, or reuses the article’s form of expression so recognisably that the user, reading it, no longer needs to access the original.

The right of making available to the public is, as a rule, infringed where that form of expression is systematically recognisable in the chatbot’s answers to related queries.

On liability, the provider is responsible where those uses are inherent to its business model, or where it has failed to put in place appropriate measures to counter them “credibly and effectively”.

The opinion is not binding. The judges now deliberate, and judgment will follow later. Two further Hungarian AI references are already pending.

What this means for you

Strictly, this is about the press publishers’ right, and about one chatbot. But the framework the Advocate General proposes — recognisable expression, substitution for the source, lawful access, and the provider’s own countermeasures — is built on general copyright concepts, and it is the first serious EU test for content served back through AI answers.

For brand owners, publishers of guides and reviews, and anyone whose original material is being summarised by AI systems, three practical points follow:

• Evidence is the case. The question will be whether specific answers reproduce your expression. Capture the prompts, answers, dates and the source pages they track.
• Access controls matter. Paywalls and anti-scraping measures are now relevant to whether retrieval was lawful, not just to traffic.
• Ask the provider what it does. Liability turns partly on whether the provider has credible measures against repeated reuse. A documented complaint, and its outcome, is part of that record.

What the opinion does not do is give anyone control over facts. An AI answer that says the same thing in its own words is outside the right, even if it costs the original a click.