Publication date: 3 November 2026
The first generative artificial intelligence case to reach the Court of Justice of the European Union has produced an important attempt to define how AI chatbots may use news. Advocate General Maciej Szpunar said Google Gemini can infringe a press publisher’s rights when it retrieves an article in real time and reuses protected expression in a way that effectively replaces access to the original publication.
The opinion is not a judgment and is not binding on the court. It nevertheless offers a detailed framework for one of the central disputes between publishers and AI companies: when a generated answer is a lawful account of facts, and when it becomes an unauthorised substitute for journalism.
Facts and expression receive different treatment
The opinion says press publisher rights protect the specific form of expression in an article, not the underlying facts. A chatbot may therefore report information without granting a publisher ownership over the news itself. Risk increases when recognisable wording, structure or presentation appears systematically in answers and gives the user little reason to visit the source.
Access also matters. The Advocate General said publisher rights may be infringed when a chatbot circumvents effective technical measures such as a paywall or controls intended to prevent scraping. This places responsibility on AI providers to respect access restrictions and build credible measures against unlawful retrieval.
Training and live retrieval remain separate questions
The Hungarian case brought by Like Company challenged both the alleged use of articles for model training and Gemini’s responses to user queries. The opinion advised the court to treat the training questions as inadmissible because the evidence did not establish that the publisher’s articles had been used and because the claimed right concerned online use. It added that training is generally permitted under the European text and data mining exception.
That distinction is significant. Model training uses large bodies of material to develop capabilities. Retrieval augmented generation accesses current pages to answer a specific query. Governance, licensing and technical controls may need to treat those processes differently.
Publishers should make access preferences clear and monitor when generated answers reproduce distinctive material. AI providers should retain retrieval records, respect technical restrictions, identify sources and evaluate whether outputs substitute for protected publications. Attribution alone may not cure copying, but it improves traceability and gives users a path to the original reporting.
The court’s eventual ruling will determine the legal effect. Even before then, the opinion provides a practical design principle: AI systems can communicate facts, but they should not bypass access controls or turn a publisher’s original expression into a replacement product.