The Case: A Hungarian Publisher Takes on Google
The dispute began when Hungarian media operator Like Company brought proceedings against Google Ireland before a Budapest court.
Like Company alleges that Google used articles from its websites to train Gemini’s large language models and reproduced its content through chatbot responses between June 2023 and May 2024.
Gemini uses a technology called Retrieval-Augmented Generation (RAG), which allows the chatbot to retrieve information from external sources, including news websites, when answering users‘ questions.
The Hungarian court referred four questions to the CJEU, addressing both AI training and the reproduction of protected journalistic content in chatbot responses.
What Does EU Copyright Law Say?
Under Article 15 of the EU Copyright Directive (2019/790), press publishers enjoy specific rights over the online use of their publications by information society service providers.
Advocate General Szpunar drew an important distinction: reporting facts is not the same as reproducing protected expression.
Merely communicating information contained in an article does not necessarily infringe publishers‘ rights. However, infringement may arise where a chatbot circumvents effective technical restrictions, such as paywalls, or reproduces the original publication’s expression recognisably enough to replace reading the article itself.
The Advocate General also suggested that making protected expression systematically available through chatbot responses may infringe publishers‘ rights. AI providers could bear responsibility where such uses are inherent to their business models or they fail to implement effective safeguards.
What About Copyrighted AI Training Data?
The Opinion also addresses a separate and highly contested question: can copyrighted material legally be used to train AI models?
Szpunar recommended declaring the training-related questions inadmissible because there was insufficient evidence that Like Company’s articles had actually been used to train Gemini.
Nevertheless, he indicated that AI training may, in principle, fall within the text-and-data-mining exception under Article 4 of Directive 2019/790.
Importantly, that exception is not unconditional. It requires lawful access and does not apply where rights holders have appropriately reserved their rights.
The distinction between training an AI model and generating potentially infringing outputs therefore remains central to the legal debate.
Final Thoughts
The dispute raises a question extending far beyond Google Gemini: what happens to journalism when readers can obtain its substance without visiting the publisher’s website?
For publishers, the concern is not simply copyright infringement but the economic sustainability of original reporting. For AI developers, access to information remains essential to building useful services.
The Advocate General’s approach attempts to distinguish legitimate information retrieval from the unauthorised exploitation of protected content.
However, this is not yet a judgment. The Opinion is non-binding, and the CJEU must still deliver its ruling.
The eventual decision could help define the boundaries between AI-generated answers, copyright protection and access to information in Europe.
