On 3 November 2026, Advocate General Maciej Szpunar of the Court of Justice of the European Union delivered an Opinion in Like Company v Google (Case C-250/25), brought by a Hungarian publisher: when a chatbot in the EU retrieves press publications in real time and uses them to answer users, it may in certain circumstances infringe publishers' rights. This is the first generative AI case to reach the Court. The Opinion is not binding and no judgment has been given, but it judges model training and live answers separately, which matters directly for retrieval-based chatbots operating in the EU, news publishers, and teams that use AI to read the news.
Training put to one side: two tests decide whether an answer infringes
The Advocate General proposed that the Court should not answer the two training questions in this case, because there was no evidence that Like Company's articles had been used to train Gemini. He also took the view that using copyright-protected works to train a large language model generally falls within the EU text and data mining exception in the Copyright in the Digital Single Market Directive. The answer stage is different: the publishers' right protects the specific form of expression of a news item, not the facts it reports. In his view, an answer may fall within the publishers' reproduction right if the chatbot unlawfully circumvents a paywall or another technical barrier to obtain a publication, or reuses its form of expression in a recognisable way so that the user no longer needs to consult the original. Merely rewording the text does not automatically take an answer outside that assessment.
Four checks before offering retrieval-based answers in the EU
First, lawful access: do not circumvent paywalls, anti-scraping barriers or access restrictions, and keep a record of the authorisation and access route for each source. Second, lawful expression: build answers around the facts, in your own wording, keep quotations restrained, and do not recognisably reproduce the original's expression or substitute for reading the original. Third, keep training and retrieval records separate: record the legal basis and any reservations for training data separately from the sources and access routes used for retrieval answers, and do not use a conclusion about training to justify live retrieval. Fourth, contracts and licences first: where there is a partnership or licence with a publisher, stay within its scope; without one, confirm that the material is publicly available content that may be retrieved directly before launching retrieval answers. Publishers can preserve evidence of retrieval, check their own text and data mining reservations, and update licensing and technical-barrier strategies in light of the forthcoming judgment.
An Opinion, not a judgment, and its scope is the EU only
An Advocate General's Opinion is not a judgment of the Court: the judges may adopt, narrow or reject it. The Hungarian court will still have to decide the dispute in light of the judgment, and the Opinion did not order any change to Gemini, award damages, or set a date for judgment. This case turns on the EU publishers' rights rules, which cannot be applied directly to US fair use or other legal systems; the recommendation not to answer the training questions because of a lack of evidence in this case does not mean that all model training has a general licence. For operators, the prudent course is to treat the Opinion as a risk warning and act now rather than wait for the final judgment. For background, see EU copyright consultation extended to 10 November.
