Case C-250/25, Like Company – does a chatbot copy when it predicts?
Every argument about generative AI and copyright eventually reaches the same two questions: is training a copy, and is output a copy? A Hungarian court has now put both to the Court of Justice, and added a third that may decide the case — whether it matters that the machine was only guessing the next word.
Facts
Like Company, a Hungarian press publisher, is suing Google Ireland Limited before the Budapest Környéki Törvényszék. Its complaint concerns Google’s LLM-based chatbot: in response to user prompts, the chatbot displayed text partially identical to articles on Like Company’s websites, in passages long enough to attract the press publishers’ right under Article 15 of Directive (EU) 2019/790. The referring court’s four questions run the full length of the process — from the training of the model on the publisher’s pages, through the text-and-data-mining exception, to the moment a user’s prompt causes an article to reappear on screen. The notice gives no further facts. Lodged on 3 April 2025, this is the first reference to the Court on generative AI and copyright; no Advocate General’s Opinion has yet been delivered.
Questions Referred
1. Must Article 15(1) of Directive (EU) 2019/790 and Article 3(2) of Directive 2001/29/EC be interpreted as meaning that the display, in the responses of an LLM-based chatbot, of a text partially identical to the content of web pages of press publishers, where the length of that text is such that it is already protected under Article 15 of Directive 2019/790, constitutes an instance of communication to the public? If the answer to that question is in the affirmative, does the fact that [the responses in question are] the result of a process in which the chatbot merely predicts the next word on the basis of observed patterns have any relevance?
2. Must Article 15(1) of Directive 2019/790 and Article 2 of Directive 2001/29 be interpreted as meaning that the process of training an LLM-based chatbot constitutes an instance of reproduction, where that LLM is built on the basis of the observation and matching of patterns, making it possible for the model to learn to recognise linguistic patterns?
3. If the answer to the second question referred is in the affirmative, does such reproduction of lawfully accessible works fall within the exception provided for in Article 4 of Directive 2019/790, which ensures free use for the purposes of text and data mining?
4. Must Article 15(1) of Directive 2019/790 and Article 2 of Directive 2001/29 be interpreted as meaning that, where a user gives an LLM-based chatbot an instruction which matches the text contained in a press publication, or which refers to that text, and the chatbot then generates its response based on the instruction given by the user, the fact that, in that response, part or all of the content of a press publication is displayed constitutes an instance of reproduction on the part of the chatbot service provider?
Sources
OJ notice C/2025/3039 (EUR‑Lex) · Case file on CURIA · Directive (EU) 2019/790 · Directive 2001/29/EC
Comment
The reference is built on the press publishers’ right rather than on ordinary copyright, and that is a shrewd choice. Article 15(1) gives publishers the reproduction and making-available rights of Directive 2001/29 “for the online use of their press publications by information society service providers”, and asks nothing about originality; its only textual limit is that it does not apply to “individual words or very short extracts”. Like Company therefore does not have to prove that any particular passage was its journalists’ own intellectual creation — only that the chatbot’s output reproduced more than a very short extract. The notice says the passages were long enough, which puts the case straight onto the questions of substance.
Question 2 is where the ground is least charted. Training a language model involves making copies of its input at some stage; the referring court’s framing — a model “built on the basis of the observation and matching of patterns” — invites the Court to say whether what the model retains is a reproduction at all, or a statistical residue that is no longer the publication. If it is a reproduction, Question 3 asks whether Article 4 saves it. That provision exempts “reproductions and extractions of lawfully accessible works … for the purposes of text and data mining”, defined in Article 2(2) as “any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations”. Training an LLM is, on its face, exactly that. The catch is Article 4(3): the exception applies only where the use “has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means”. So the answer to Question 3 may turn on a fact the notice does not give — whether Like Company’s pages carried an opt-out, and in what form. An affirmative answer would make Article 4 the legal foundation of AI training in the Union; a negative one would make every unreserved training run an infringement waiting for a claimant.
Questions 1 and 4 concern output, and here the sub-question in Question 1 is the one to watch: does it matter that the chatbot “merely predicts the next word on the basis of observed patterns”? On the Court’s approach to reproduction since Case C‑5/08, Infopaq (ECLI:EU:C:2009:465), what counts is the result, not the mechanism — a copy made by a predictive process is still a copy. But the argument has a second life in Question 4, which asks who is responsible when a user’s prompt is what causes the passage to appear. That is not a question about whether reproduction occurred, but about attribution: whether the service provider reproduces the publication when the user, in effect, asks it to.
The reference arrives with unusually fresh guidance on the right it invokes. In Case C‑797/23, Meta Platforms Ireland (ECLI:EU:C:2026:395) the Grand Chamber upheld an Italian scheme under which publishers may claim fair remuneration from information society service providers that use their publications, and which obliges the providers to negotiate and to disclose the information needed to set the price. That judgment treated Article 15 as a right with economic teeth against platforms; Like Company asks whether the teeth reach a service that never links to the publisher at all, but re-states its content from memory. If the answer is yes, the remuneration machinery the Court approved in May will apply to chatbots too. Whichever way the questions fall, this is the reference that will define what generative AI owes to those who wrote its training data — and it will be answered before the AI Act’s own transparency rules for training data have had any time to bite.