EU court adviser says Gemini's answers, not its training, carry Google's copyright risk

On Tuesday, Advocate General Maciej Szpunar gave his opinion in Like Company v Google (C-250/25), the first referral on AI and copyright to reach the Court of Justice of the European Union. Headlines called it a setback for Google, which is half right. The 240-paragraph opinion gives the AI industry much of what it wanted on training data. It then puts a line through the part of the product that increasingly carries the commercial weight: the answer a chatbot shows the user.

The facts are narrow. Like Company, a Hungarian publisher that runs advertising-funded news portals, sued Google Ireland in Budapest. It relied not on copyright but on the press publishers' related right in Article 15 of the EU's 2019 Copyright in the Digital Single Market Directive. Its complaint centres on Gemini answers that allegedly reproduced its articles and made them available to users without permission; Google, for its part, argued the chatbot was not trained in Hungary and invoked the data-mining exception. The Budapest court's referral, lodged with the CJEU on April 3, 2025, asks four questions: two about what a chatbot displays and two about training.

Training: a provisional green light, with conditions

Szpunar advised the court not to answer the training questions at all. Nothing showed the publisher's articles were in Gemini's training data, and the press publishers' right covers only online uses, which training is not, so the questions are hypothetical. He gave a provisional view anyway. Training does engage the reproduction right, he said, and it makes no difference whether content is ingested "as such or as data." But the text-and-data-mining exception in Article 4 of the 2019 directive applies in principle. He rejected the argument that lawmakers could not have had AI in mind when they wrote it.

The conditions carry most of the weight. The exception requires lawful access to the material and no reservation of rights by the rightsholder. Szpunar also suggested that Article 4 could become a compensated limitation, meaning a permission that comes with payment. The IPKat blog notes that the European Commission's current consultation raises the same idea. Broad permission today is no promise about the next revision.

He left one large issue open: where training legally happens. Google argued that the applicable law should be that of the place where its main computing infrastructure sits. Szpunar said that approach had "a significant downside," because it would leave EU rightsholders with no protection against training done abroad. He did not fully accept the alternative of judging training by its effects inside the EU either, and he asked the legislature to settle the question. That matters to every lab that trains models outside Europe and sells them inside it.

Answers: where the exposure moves

Outputs raise no such territorial problem, because the answers go directly to users in the EU. Here the opinion separates facts, which are free, from the way a publication expresses them, which is protected. Szpunar gave two triggers for the reproduction right. One is unlawful access, such as getting around a paywall or a similar technical measure. The other is an answer that reuses the publication's form of expression recognisably enough that the user has no need to read the original. For the separate right of making content available to the public, he said infringement generally occurs when that expression is "systematically recognisable" across answers to related queries. Eleonora Rosati's write-up on the IPKat blog notes that the AG started from the premise that chatbots such as Google's AI Overviews and AI Mode rely on retrieval-augmented generation, not just a trained model, which points the output analysis squarely at answers assembled from fetched pages.

Most of Google's other defences did not survive. Szpunar rejected the idea that publishers implicitly consent by putting content online. He held that the text-and-data-mining exception does not cover chatbot answers, since generating an answer is not data mining. He also said it cannot be assumed that such answers fall under the directive's carve-outs for "mere facts" or "very short extracts." The exception for temporary copies in the older InfoSoc Directive could still cover some of the acts involved.

The liability test may matter most in practice. According to CEA Digital Law's summary, a provider would be liable where infringing uses are inherent to its business model, or where it has not put in place measures that counter them credibly and effectively. The IPKat adds that the provider answers for its own acts, while any liability on the part of users would be assessed under the test the court set in its YouTube rulings (C-682/18 and C-683/18). That is close to a compliance standard that a product could later be checked against.

Why this matters beyond Hungary

The opinion is not binding. The court follows its advocates general in most cases, and a judgment is expected in the coming months. A ruling along these lines could oblige Google to pay for news content that surfaces in Gemini answers. Google's statement focused on training: a spokesperson called the EU's data-mining exceptions "vital for Europe's competitiveness and AI development" and said Gemini produces "original outputs rather than replicating content." Angela Mills Wade, executive director of the European Publishers Council, said publishers "should not have to choose between search visibility and protecting their rights."

Google has faced this fight in Europe before. In March 2024, France's competition authority fined it €250 million after finding it had breached four of seven commitments it made to publishers in a 2022 settlement. Publishers' complaints in that case included that Google had used their content to train Bard, now Gemini, without notifying them. Szpunar's opinion would change what these disputes turn on. Training cases would hinge on opt-outs and lawful access. Output cases would hinge on recognisability, which a publisher can test directly by running queries and comparing the answers with its own articles.

The contrast with the United States is sharp. In late September the Third Circuit upheld a ruling that ROSS Intelligence's copying of Westlaw headnotes to train a non-generative legal research tool was not fair use. That case centred on the training copies and a product that competed directly with Westlaw. The first US appellate ruling on fair use in AI training puts the risk at the training stage, while Szpunar puts the EU risk mainly in the output. A company could satisfy one regime and still be exposed under the other.

What to watch

CEA Digital Law has already flagged the unanswered questions: how recognisability will be measured, what counts as credible and effective prevention, and how a publisher can prove unlawful access. Other AI copyright referrals are pending in Luxembourg, including Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26), so Like Company will not be the last word. For AI search and answer products in Europe, the question is shifting from what went into the training set to what appears on the screen: how much of a source is quoted, how paywalled pages are handled and how sources are credited. The next question is whether those controls would look credible to a judge.