A Hungarian publisher took Gemini to Luxembourg

The case began modestly: Like Company, a Hungarian news publisher, sued Google Ireland before the Budapest Environs Regional Court, saying answers generated by Google's Gemini chatbot reproduced parts of its press articles without permission or payment. On 3 April 2025 the Hungarian court referred the dispute to the Court of Justice of the EU, and on 10 March 2026 the Grand Chamber heard it, the first hearing the EU's top court has ever held on generative AI and copyright.

The elevation to the Grand Chamber is the tell. National courts, notably in Germany, have started answering fragments of the AI copyright question, one lawsuit and one jurisdiction at a time. Luxembourg will answer it once, for all 27 member states, in a judgment every national court must follow.

The four questions that define the fight

The referred questions cover the whole AI pipeline. First, is a chatbot answer that displays text partially identical to a press publication a communication to the public under the press publishers' right and the InfoSoc Directive, and does it matter that the model merely predicts the next word? Second, does training an LLM on such content constitute a reproduction? Third, if it does, is that reproduction covered by the text and data mining exception in Article 4 of the DSM Directive? Fourth, is there a reproduction when the chatbot, prompted by a user, generates a response matching press content in part or in full?

Read together, the questions dismantle the industry's favourite simplification: that training is a quiet, lawful data exercise and outputs are the user's problem. The Court is being asked to trace liability through the whole chain, from ingestion to answer, and to say where the text and data mining exception, the legal foundation European AI training currently stands on, actually ends.

What is actually at stake

The economics are plain. If training on press content is a reproduction that the TDM exception does not cover, or if chatbot outputs are a communication to the public, then AI companies need licences, and licences mean a price for what has so far been free raw material. Europe's press publishers, whose right the referring court invokes, would gain a floor under negotiations that have so far been voluntary and lopsided.

If Google prevails on all four questions, the opposite settles in: training stays inside the TDM exception, outputs stay largely unpriced, and the value of the opt-out that Article 4 grants rights holders becomes the central question, because reserving your rights is only worth something if unreserved content is what keeps training lawful.

The calendar, and the owner's move

Advocate General Maciej Szpunar delivers his opinion on 3 September 2026. Opinions do not bind the Court, but they shape most judgments, and the final ruling is expected some months after. Anyone signing multi-year AI contracts or content licences this year is signing under a law that may look different next year.

If your business publishes content of value, newsletters, research, trade media, product data, decide your TDM position now: an explicit, machine-readable rights reservation costs little and preserves the negotiating position the judgment may create. If your business deploys AI assistants or chatbots, the output questions are aimed at you as much as at Google: ask your vendor in writing who bears copyright liability for generated output, and whether the indemnity survives an adverse ruling.

And put 3 September in the calendar. Not because anything explodes that day, but because the direction of the opinion will tell you which of the two futures to prepare for, months before the judgment makes it law.

Read next: Your AI Now Speaks for You in Law  ·  Software Becomes a Product in December