An adviser to the European Union's highest court has set out when an AI chatbot's use of press articles infringes copyright, and the answer turns far less on what the model was trained on than on how recognisably it can reproduce what it has read.
Advocate General Maciej Szpunar delivered his opinion on 3 November in Case C-250/25, Like Company v Google, the first generative-AI copyright case to reach the Court of Justice of the European Union. It arrives on a reference from a Hungarian court, which put four questions to the Court. The analysis below follows the account given by Joao Pedro Quintais of the Institute for Information Law in Amsterdam, writing on the Kluwer Copyright Blog.
The four questions ask whether displaying press publication content in a chatbot response amounts to making it available to the public; whether training a large language model involves acts of reproduction; whether the text-and-data-mining exception in Article 4 of the Copyright in the Digital Single Market Directive covers those reproductions; and whether displaying press content in a response is itself a reproduction by the chatbot's provider.
On training, the Advocate General considers the relevant questions inadmissible on these facts. There is no evidence in the file that Like Company's articles were used to train Gemini, and he regards real-time retrieval as the better explanation of the example before the Court. "The case in the main proceedings has nothing to do with LLM training," he writes.
He answers one of them for completeness all the same. Reproductions made for the purpose of training are, as a rule, covered by the Article 4 exception where its conditions are met, meaning lawful access and no applicable opt-out. He reads the definition of text and data mining broadly. He also takes the view that training is not online use within the meaning of Article 15, the press publishers' right, and so sits outside it altogether.
On whether the model itself holds copies, he advises "the utmost caution", and describes the inference that a copy of the training material is stored inside the model as "far-fetched".
Outputs are treated differently. Retrieval-augmented generation creates technical reproductions that the text-and-data-mining exception does not cover, though these may fall within the temporary reproduction exception in Article 5(1) of the InfoSoc Directive.
The operative test is narrower still. Reproduction occurs where the chatbot either accesses the publication unlawfully, by circumventing effective technical measures or by scraping, or produces an answer whose form of expression remains recognisable, to the extent that users can dispense entirely with the original. Facts alone, or very short extracts, are not enough: Article 15 protects the form of expression and not the facts.
Acts may be attributed to the provider where they are inherent in its business model, or where it has failed to take the measures expected of a reasonably diligent operator. On the facts as presented, the Advocate General's provisional view is that the conditions for liability do not appear to be met, with the final assessment left to the referring court. The opinion is not binding; the Court will now deliberate. Two further Hungarian references are already pending.
Reduced to its working part, the test asks how much of the article the machine can still produce on demand. Copying it is lawful. Having read it is lawful, and any suggestion that a copy survives inside the model is far-fetched. What is not lawful is handing it back in a form the reader recognises.
This makes the protected interest a very particular thing. It is not the article. It is the reader's remaining reason to go and look at the article.
A compliance strategy follows from that, and it is an unusual one. A provider reduces its exposure by reducing the fidelity of its own output. The more exactly a system can render what it has read, the closer it comes to the line. The more it paraphrases, blurs and loses the thread, the further it sits from it.
Our analysts note where this places the standard of the reasonably diligent operator. Diligence, on this reading, is the discipline of building a system that has read everything and can repeat none of it accurately, and of auditing it regularly to confirm that it still cannot.
The directive is explicit that facts are not protected, only their form of expression. A chatbot may therefore return every fact in a press article, in the correct order and at the correct length, provided it does so in prose the publisher would decline to claim as its own.
The Department of Unnecessary Research has extracted the only operational guidance available to developers before the Court rules, and offers it without comment. The safest model is the one that read the article and did not take it in.