An adviser to Europe’s highest court has drawn a sharp line through the AI copyright debate, ruling in effect that a chatbot may learn from the news but may not hand back the news itself. Advocate General Maciej Szpunar said on Tuesday that Google’s Gemini chatbot can infringe press publishers’ rights when it retrieves a publication, reuses it to generate an answer, and gives that answer to the user. The opinion is advisory, not a final judgment, but it is the first time the EU’s top court has squarely addressed generative AI and press rights.
What the adviser said
Szpunar was advising the Court of Justice of the European Union in a case brought by Hungarian publisher Like Company against Google Ireland. Like Company alleges that Gemini reproduced and made available substantial parts of its protected press articles without permission. Google disputes that its answers copied the publications or made them available to the public.
The case, C-250/25, was referred to the Luxembourg court by a Hungarian court, which asked four questions about AI and copyright. Two concern training, and two concern the answers Gemini generates from articles available online. Szpunar proposed that the training questions be set aside, because the case file contains no evidence that Like Company’s articles were used to train Gemini.
Training is not the same as answering
Even if the court does reach training, Szpunar said using copyrighted material to train a large language model is, as a general rule, freely permitted under the EU’s text and data mining exception. That exception covers ingesting large volumes of text so a model can learn patterns, provided it was lawfully accessible and not expressly reserved by the rights holder.
Answering is different, because the chatbot is not learning in that moment. It retrieves a specific publication in real time, through retrieval-augmented generation, and turns it into a response for one user. Szpunar argued that this retrieval can infringe a press publisher’s neighbouring right, the separate EU right over online uses of publications under Article 15 of the copyright directive.
Two tests for an infringing answer
The opinion sets out two combined situations in which a Gemini answer could cross the line. The first concerns how the chatbot obtained the article: an answer falls under the publisher’s reproduction right when the chatbot circumvents a paywall or another technical barrier, such as a measure designed to block automated scraping of the site.
The second concerns what the answer contains. Copyright does not protect the mere facts reported by a publication, only its specific form of expression. An answer conveying the facts alone does not infringe, but one that reuses recognisable wording from the publication can do so, especially when it reproduces so much that the reader need not consult the original article.
Liability, in Szpunar’s framing, would not rest with the user asking the question. It could lie with the chatbot provider where reuse of publisher material is inherent to its business model, or where it has not put credible measures in place to prevent such reuse. A pattern of such answers could amount to making the work available to the public.
Our read: the substitution test is the story
The most important word in this opinion is not infringement but substitution. Szpunar is not saying chatbots must stop reading the news, nor giving publishers a monopoly over facts. He is saying that the legal question turns on whether the answer replaces the article, in wording and in function, so the reader never visits the source that paid for it.
That framing echoes the split the Wikimedia Foundation urged in a New York court: training can be fair use, while answers that stand in for the original should respect licensing terms. The two cases come from different legal systems, but they point the way: the hardest copyright problem is no longer the training run, but the answer on the screen.
For Google, the practical risk is therefore about product design, not model training. If the court follows Szpunar, Gemini and similar systems will need stronger guardrails around retrieved material: respecting paywalls and scraping controls, limiting how much recognisable wording an answer can carry, and making sure the response points the reader back to the source rather than replacing it.
Why It Matters
For publishers, an opinion in this direction could be a lifeline, not only in court. A ruling against Google could force payment for the use of press content in chatbot answers, when AI summaries are already diverting readers from the sites that produced the reporting. The money matters, but the traffic matters more, because advertising and subscriptions depend on visits.
The timing matters for search itself. Google is already moving AI-generated brand summaries into a side panel on desktop, a layout change showing how much an AI answer’s placement affects who gets seen. A copyright rule that treats a substituting answer as an infringement would give publishers a second lever, alongside layout, over answers that compete with their own pages.
For users, nothing changes immediately. Advocate General opinions are not binding, and the court’s judges are not obliged to follow Szpunar’s reasoning, though their opinions are usually followed in practice. The judges will deliver their own ruling at a later date, and the Hungarian court will then decide the underlying lawsuit between Like Company and Google.
Background
The dispute reaches back to Gemini’s early period, when Like Company says its articles were reproduced in responses. The Hungarian referral, made in April 2025, followed an oral hearing in March 2026, and Szpunar’s opinion arrived eight months later. Whatever the court decides, it will set the first EU-wide guidance on AI providers’ use of news content in conversational answers.
Source: Reuters report and Delano analysis
0 Comments