What did the EU court adviser decide in the Gemini copyright case?
The first generative AI copyright case to reach the Court of Justice of the European Union produced its first legal milestone on 3 November 2026. Advocate General Maciej Szpunar delivered an opinion in Case C-250/25, a dispute between Hungarian publisher Like Company and Google Ireland over the Gemini chatbot.
The referral came from the Budapest Regional Court. Like Company argued that Google used articles from its news portals to train Gemini and then reproduced them in chatbot answers between 13 June 2023 and 31 May 2024. The Hungarian court asked four questions about training and about the answers Gemini gives users.
Szpunar advised the Court to declare the two training questions inadmissible. In his view, there was no evidence that Like Company's articles were used to train Gemini, and training is not an online use of press publications under the relevant EU rule. He still addressed training for completeness.
The operative part of the opinion concerns retrieval. Gemini uses retrieval-augmented generation, which lets it fetch material from the web in real time and use that material as context for an answer. The Advocate General said the press publisher's right may be infringed in certain circumstances when a chatbot retrieves an article, reuses it and returns the result to a user.
He advised the Court to consider the two output questions together, because they describe the same act from different angles. One asks whether the answer reproduces the article; the other asks whether it makes the article available to the public. Kluwer notes that the Advocate General treated the training questions as inadmissible but rejected the European Commission's argument that the whole referral should be thrown out.
An Advocate General's opinion is not binding on the Court. The judges now deliberate, and a judgment will follow at a later date. Opinions carry weight because the Court often follows its advisers, but they are not verdicts.
Why the case matters for news, links and lost traffic
The dispute sits inside a commercial fight over how news reaches readers. Search engines have long sent visitors to publishers through links, and those visits carry advertising and subscription value. AI assistants increasingly answer questions directly, which can remove the click that pays for the reporting.
Reuters reported that publishers have seen traffic dwindle as AI chatbot summaries replace traditional search results. The news agency also said the tensions have prompted complaints to EU and national antitrust regulators.
EU law gives publishers a specific tool for this fight. Article 15(1) of the 2019 Digital Single Market directive created a press publishers' right over the online use of their publications by information society service providers. The Like Company case tests whether chatbot answers fall inside that right.
The right is a related right, not a general copyright over the news. It was designed to let publishers capture some value when platforms make their publications available online, and it sits alongside the authors' rights in the same articles. Like Company operates various web portals, which is why the dispute covers a range of articles rather than a single story.
A ruling against Google could force the company to pay for the use of news content, Reuters reported. A ruling the other way would leave publishers with less leverage in licensing talks, at a time when referrals from search are under pressure.
The European Publishers Council framed the issue as a choice about substitution. Angela Mills Wade, the council's executive director, said it was important to separate search indexing from authorising AI answers that replace original reporting. "Publishers should not have to choose between search visibility and protecting their rights," she said, according to Reuters.
When can an AI chatbot's answer infringe publishers' rights?
The opinion draws a line between facts and expression. The press publishers' right does not give publishers a monopoly over the news itself; it protects the specific form in which a story is written. Szpunar described the aim as protecting the investment that goes into a publication, not ownership of the underlying events.
For the reproduction right, he proposed two questions. First, did the chatbot access the article unlawfully, for example by circumventing an effective technical measure such as a paywall or anti-scraping defence? Second, does the answer merely report facts, or does it reuse the article's form of expression in a recognisable way?
The two routes interact. Kluwer notes that retrieval creates temporary technical copies, and those copies may fall under the temporary reproduction exception in Article 5(1) of the Information Society directive. Whether that exception applies depends on whether the access was lawful and whether the final answer falls within the publisher's exclusive rights.
Consider two answers to the same query about a court ruling. One restates the verdict, the parties and the reasoning in its own words. The other follows the article's structure, echoes its distinctive phrasing and leaves the reader with no reason to open the page. The first is closer to reporting facts; the second is closer to reproduction.
The second question turns on substitution. If a reader can rely on the answer and skip the original article entirely, the balance shifts towards infringement. An answer that conveys the same facts in different words does not, on this test, reproduce the publication.
The making-available right applies where the expressive form is systematically recognisable across answers to related queries. That wording matters because it targets patterns of behaviour rather than a single unfortunate output.
Liability follows the design of the service. The provider can be responsible where the conduct is inherent to its business model, or where it failed to put in place measures that counter the conduct credibly and effectively. A user's query does not automatically move responsibility to the user.
João Pedro Quintais of the Institute for Information Law in Amsterdam, writing on the Kluwer Copyright Blog, identifies two alternative routes in the opinion. One concerns the lawfulness of the access behind temporary technical copies made during retrieval. The other concerns whether the final answer reproduces protected expression. Lawful access, he notes, does not necessarily make the later reproduction lawful.
That distinction could shape how chatbots are built. A system that fetches a page lawfully but then reproduces its structure and phrasing may still face a claim. A system that summarises facts and points readers to the source sits closer to the safe side of the line.
Is training an AI on news articles legal in the EU?
Training is the part of the case the Advocate General tried hardest to avoid. He called the training questions hypothetical because the record did not show that Gemini had learned from Like Company's articles. His exact phrase, quoted by Kluwer, was that "the case in the main proceedings has nothing to do with LLM training".
For completeness, he said training is freely permitted under the text and data mining exception in Article 4 of the 2019 directive. That exception applies to lawfully accessible works unless rightsholders have reserved their rights through an opt-out. The conditions matter, but the direction of travel is permissive.
The framework therefore leaves room for publisher choices. Google told Reuters that its publisher controls give rightsholders a clear choice and the ability to opt out, and the exception treats a reservation of rights as one of its conditions. Training is not a blanket licence; it is a default that rightsholders can displace.
Kluwer explains the technical layers behind that conclusion. Training involves copies that are reproductions under the Information Society directive. The trained model itself, viewed as statistical relationships, does not ordinarily reproduce the training material. The blog also notes the Advocate General's caution about memorisation, which he described as incidental to machine learning and not entirely avoidable.
The opinion treats the idea that a model stores copies of its training data as "far-fetched", according to Kluwer. That reasoning could complicate parts of German case law, including Munich rulings involving GEMA against OpenAI and Suno, which Kluwer says the Advocate General discussed.
Copyright scholars had urged caution. Kluwer notes that the European Copyright Society argued the press publishers' right is narrow and warned against drawing broad conclusions from an uncertain factual record. The Advocate General shared part of that caution, especially in separating training from retrieval.
This matters because the Court may not decide the training question at all in this case. If it follows the advice to dismiss those questions, the judgment will focus on outputs, leaving the legality of training to other cases and to national courts.
Other disputes are already queued. Kluwer lists two further Hungarian referrals, Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26), alongside the German litigation. Each could produce a different set of questions and a different factual record.
What did Google and the publishers say?
Google stressed the part of the opinion that favours its position. A spokesperson told Reuters that "European law's text and data mining exceptions are vital for Europe's competitiveness and AI development". The company also welcomed the recognition that EU law permits training on publicly available information.
The spokesperson added that Gemini "learns linguistic patterns to create original outputs rather than replicating content". Google pointed to publisher controls that give rightsholders a clear choice and the ability to opt out of training.
Publishers focused on the output framework. Angela Mills Wade of the European Publishers Council said the distinction between indexing and substituting for reporting is central. "Google's powerful gatekeeping role in search makes this particularly important," she said, according to Reuters.
She added that publishers depend on Google to reach readers while Google increasingly delivers AI-generated answers directly to users. In her view, control over access to audiences should not become leverage to secure content for AI services.
The gap between the two readings is not spin alone. It reflects a genuine split in the opinion between what happens during training and what happens when a model answers a question. Google won the first part; publishers gained ground on the second.
The disagreement is also about leverage. Publishers argue that Google's position in search gives it an advantage in negotiations over AI content. Google argues that its publisher controls already give rightsholders a say. The opinion addresses liability, not the fairness of the bargain, so it does not settle that argument.
What are the limits of the opinion?
The first limit is legal status. An Advocate General's opinion is advisory, and the Court can depart from it. Reuters noted that the Court follows the majority of its advisers' non-binding opinions, which makes the text influential but not decisive.
The second limit is factual. Kluwer reports that, applying his own framework provisionally to the Like Company dispute, the Advocate General suggested the conditions for liability did not appear to be met. The article's form of expression was not clearly recognisable in the response, and Google appeared to have implemented measures to reduce the risk.
That provisional view explains why both sides could claim support. Google emphasised the training finding and the cautious application to the facts. Publishers emphasised that the Court was told AI answers can, in principle, infringe.
The third limit is the test itself. Terms such as recognisable, systematically and entirely dispensed from accessing the original are fact-intensive. They will be argued market by market, language by language and product by product, with experts comparing answers to articles.
The referral was also awkwardly drafted. Kluwer quotes the Advocate General's remark that the Hungarian court was "evidently lost in the AI landscape" and notes that he still tried to give it usable guidance. That procedural detail limits how far the judgment can go: the Court is answering questions about a technology the referral described imprecisely.
The fourth limit is territorial. Kluwer reports the Advocate General's view that making a model available in the EU after training elsewhere does not by itself justify applying EU copyright law to the training acts. Scraping material from servers located in the EU may still engage EU law. He favoured a legislative solution to the uncertainty, and Kluwer points to the Commission's planned consultation on the copyright rules.
The opinion also concedes that copyright cannot solve everything. Licensing rules may address substitution, but they do not settle questions about source diversity or misinformation. The Advocate General acknowledged those wider problems while keeping his analysis inside the scope of the press publishers' right.
What happens next in the Gemini copyright case?
The Court of Justice will rule in the coming months, according to Reuters. National courts across the EU will then apply the judgment to their own disputes, including the Budapest Regional Court when it resumes the Like Company case.
The key question is whether the Court adopts the substitution test. If it does, chatbot providers will need to show that their answers do not recognisably reproduce articles or replace the need to read them. Design choices around retrieval, quoting and linking will become legal questions, not just product decisions.
Three further developments are worth watching. First, the pending Hungarian referrals and the German GEMA litigation could clarify training, memorisation and the boundary between a model and its outputs. Second, licensing deals between AI companies and news organisations may set prices before courts set rules. Third, the Commission's review of the copyright framework could change the statutory ground under everyone's feet.
Licensing may move faster than litigation. If the Court endorses the substitution test, publishers gain a clearer basis to negotiate payment for answers that replace articles. If it rejects the test, bargaining shifts back to opt-outs, technical controls and antitrust complaints.
The Commission's review adds another variable. Kluwer points to a planned targeted consultation on the copyright rules, which means the statute may be revised while the case law is still forming. A judgment that reads the current text narrowly could be overtaken by legislation.
For now, the practical lesson is narrow but real. Training on publicly available news is, as a rule, permitted in the EU when the conditions are met. Reusing a publisher's expression in an answer is a different act, and the first legal opinion on that act leans towards the publishers.
Google's Gemini is the named product, but the framework is written for AI chatbots generally. That means the judgment, when it arrives, will be read closely by other chatbot providers.