EU Court adviser: Gemini can reuse an article's facts, but not copy its form

Between June 13, 2023 and May 31, 2024, Like Company, which operates several news portals in Hungary, says its articles were repeatedly reused by Gemini, Google's chatbot. The publisher took the matter to the Hungarian courts. They asked the Court of Justice of the European Union (CJEU) what European law allows an AI to do with press content.

On November 3, 2026, Advocate General Maciej Szpunar delivered his opinion in this case (C-250/25), the first on generative AI before the Court. In his view, a chatbot may report the facts contained in an article. It could, however, infringe the publisher's rights if it reproduces the article's form so closely that readers no longer need to consult the original. The opinion does not bind the Court: the judges are only beginning their deliberations.

AI training set aside

The Budapest-Capital Regional Court had asked four questions, two of them on the training of large language models. The Advocate General proposes declaring them inadmissible. First, nothing proves that Like Company's articles were used to train Gemini. Second, in his view, training is not an “online use” within the meaning of press publishers' rights.

He still gives his opinion “for the sake of completeness”: training a model on protected content is, as a general rule, freely permitted by the text and data mining exception in the 2019 European copyright directive. Legal scholar João Pedro Quintais of the Institute for Information Law (IViR) in Amsterdam analyzed the opinion on the Kluwer Copyright Blog. He notes that this exception requires lawful access to the content and no opt-out by the rightsholder. He also notes that the Advocate General urges great caution toward the idea that models store copies of texts.

When a chatbot's answer could become infringement

Most of the opinion deals with Gemini's answers. The chatbot uses retrieval-augmented generation (RAG): it fetches pages from the internet in real time, often news articles, to build its answer. But, according to the Advocate General, publishers' rights do not protect the facts reported, only their “form of expression.” They protect the investment made, not a monopoly on information.

Two criteria would therefore decide whether there is an infringement:

• access: did the chatbot obtain the article unlawfully, in particular by bypassing a paywall or a protection against automated scraping?
• the answer: does it merely report facts, or does it reproduce the article's form in a recognizable way, to the point that the user no longer needs to read it?

The right of making available to the public would as a general rule be infringed when that form is “systematically recognizable” in the chatbot's answers. The provider would be liable if these practices are inherent in its business model, or if it has not taken appropriate measures to counter them “in a credible and effective manner.” João Pedro Quintais concludes that the Advocate General is judging the design of the service rather than isolated errors.

The access criterion echoes the complaints against Anthropic in Canada, where news publishers accuse it of bypassing their paywalls to feed its AI.

What publishers and chatbot designers can take away

Nothing is settled. But if the Court follows this opinion, a few markers emerge:

• for a publisher, technical protections (paywall, bot blocking) matter: bypassing them would make reuse unlawful;
• for model training, the lever provided by the directive remains the publisher's express opt-out from text and data mining;
• for anyone designing a public-facing chatbot that draws on the press, the proposed test is substitution: an answer that reproduces the article's form to the point of replacing the reading of it would be at risk, not an answer that merely draws facts from it.

The Advocate General also addresses the question of borders. According to João Pedro Quintais's analysis, he considers that offering a model in the Union is not enough to apply European law to training carried out elsewhere, and prefers to leave this point to the legislator. He acknowledges that intellectual property alone cannot answer all the questions raised by AI.

The date of the judgment has not yet been set. It will tell us whether Europe adopts this line between information, which is free, and its presentation, which is protected.