News publisher rails against lack of compensation for Google AI overviews in EU

(CN) — Europe’s top court wrestled this week with a question dogging newsrooms worldwide: When an AI chatbot generates text that looks like a news article, who deserves credit?

A Hungarian publisher and Google clashed in Luxembourg Tuesday over whether responses generated by large language models, the technology powering modern chatbots, can reuse journalists’ work. The dispute before the Court of Justice of the European Union could become one of the first major copyright battles of the AI era.

Hungarian publisher Like Company says responses generated by Google’s AI chatbot Gemini reproduced parts of its news articles without permission or payment. It is asking a Hungarian court to find Google infringed its press publishers’ rights, prompting the court to ask the EU’s top court how copyright law applies to AI chatbots.

In court, the publisher’s lawyer Gyula Rátz said the example cited by the Hungarian court — a chatbot response referring to an article about a Hungarian singer’s plan to build a dolphin aquarium near Lake Balaton — was only one illustration of a broader problem. The real question, he argued, is whether European publishers’ rights have any practical force when AI-driven tools answer users directly instead of sending them to the original news source.

“The subject of the proceedings is the fact that the defendant used the applicant’s protected press publications in the period covered by the proceedings continuously and without consent and compensation by reproducing it and communicating it to the public,” Rátz told the judges.

Rátz said the case could determine whether publishers retain any real bargaining power when technology companies use journalistic content at scale.

“Today, we are deciding the future of quality journalism. What’s at stake is the fair compensation of publishers in the age of AI,” he said.

Like warns that where an AI model is trained should not matter under EU law, arguing tech companies could otherwise sidestep European copyright rules simply by moving servers abroad.

Google countered that chatbots generate new text rather than retrieve stored articles. The company says the dispute turns on whether a chatbot’s output should be treated as a copy of an existing article or as newly generated language that may sometimes resemble existing text.

“Through automated analysis of texts, the model learns linguistic patterns, enabling it to construct new text by predicting the next word,” Google lawyer Zoltán Szür told the judges.

Because the system generates text probabilistically, Google argues, similarities with existing articles can occur without the model retrieving or reproducing them.

“It predicts the next word based on probabilities, probabilities that the model has learned in the training process. It is inherent if in the nature of a probabilistic construction of text that the outcome may be a text that exists elsewhere. Statistically, it’s something you cannot exclude,” Google lawyer Albrecht Conrad said.

Conrad added there was no evidence in the Hungarian proceedings that Gemini had reproduced a specific article.

The case comes as courts around the world begin confronting how traditional copyright law applies to generative AI.

Judges in Luxembourg are considering questions referred by a Hungarian court about how EU copyright law applies to generative AI. One issue is whether a chatbot producing text resembling a news article could count as making that content available to the public. Another is whether training a large language model itself could amount to reproducing protected works.

Answering those questions requires navigating two pieces of European legislation: The European Union’s 2019 copyright directive, which strengthened publishers’ rights over how their articles are reused online, and the information society directive, which sets the rules for reproduction and communication of copyrighted works.

Publishers argue those rules were designed to prevent large digital platforms from profiting from journalism without paying for it. If chatbots can answer questions using material drawn from news articles, they warn, readers may never visit the original source, weakening the advertising and subscription revenues that sustain reporting.

Technology companies point to provisions allowing text and data mining — the automated analysis of large datasets used to train artificial intelligence systems — unless rightsholders have explicitly opted out.

Lawyers for the European Commission, which intervened in the case, urged judges to distinguish between how AI models are trained and what they ultimately produce. Detecting patterns across large datasets, they said, does not in itself amount to reproducing protected works under EU copyright law, though the legal assessment depends on how the system operates in practice.

The hearing also showed how judges are grappling with the technical mechanics of generative AI before the court’s 15-judge Grand Chamber, led by President Koen Lenaerts, with Octavia Spineanu-Matei serving as rapporteur.

At one point, Spineanu-Matei pressed Google’s lawyers on whether language models risk reproducing existing text.

“Can we infer from this claim that, regardless of a chatbot’s intended purpose, you acknowledge the risk of textual memorization?” the judge asked.

Responding for Google, Conrad acknowledged that statistical systems cannot completely rule out such outcomes but said there was no evidence the model in this case reproduced specific articles.

The model, he explained, “constructs text probabilistically,” generating responses based on patterns learned from large datasets rather than retrieving stored passages.

The judges will now begin their deliberations. Advocate General Maciej Szpunar is scheduled to deliver his opinion on Sept. 3, with a final judgment expected several months later.

Courthouse News reporter Eunseo Hong is based in the Netherlands.