Still in the Wrong Lake: European Copyright Society Warns How Confused Vocabulary Could Make Bad Copyright Case Law
The dolphins are back. Or rather, they never left.
In December, we tracked how Like Company v. Google Ireland Limited (Case C-250/25) had turned a pop star’s eccentric plan to introduce dolphins into Lake Balaton into a potential landmark ruling on AI and copyright. We warned that the case rested on a technically fallacious anthropomorphism: treating statistical pattern-matching as reproduction, and a real-time web retrieval as evidence of illicit memorisation.
The case is now before the Grand Chamber of the Court of Justice, with a hearing scheduled for Tuesday, March 10. Fifteen judges at the EU’s highest court will have the first opportunity to rule on a question that has been tearing through AI litigation worldwide: do AI model training and AI outputs infringe exclusive rights under copyright law?
In February 2026, the European Copyright Society (ECS), a body of independent copyright scholars from across Europe, published a formal comment urging the Court to exercise caution. Drafted by Péter Mezei, Martin Kretschmer, Thomas Margoni, Alexander Peukert, and João Pedro Quintais, the opinion is damning, as it goes further than legal technicality. At its core, it is a warning about what happens when a court tries to build doctrine on a foundation of conceptual confusion. A wrong ruling, the ECS warns, could leave Europe with the worst of all worlds: a complex opt-out regime, private licensing as a default, no clear space for research, and no improvement in the actual position of authors and performers.
A reference built on fog
The ECS opinion opens by questioning whether the reference should be heard at all. The Budapest Regional Court’s referral fails a basic admissibility threshold: it cannot consistently identify what technology is actually at stake.
The reference uses “chatbot,” “large language model,” and “search engine” as if they were interchangeable. They are not. An LLM is a trained model. A chatbot is an application that may or may not be built on one. A search engine is a different service entirely, regulated separately under EU law. The referring court acknowledges that Google Search and Google Gemini are “interconnected” and “similar,” then calls them collectively “generative AI chatbots,” without explaining the connection or the similarity in any legally meaningful way.
This matters enormously. These distinctions determine which rights apply, which exceptions cover which activities, and whether the training question (Questions 2 and 3 of the reference) has any factual basis at all. If Gemini’s response to questions about Kozsó’s dolphin plans came from retrieval-augmented generation rather than model training (which the publication timeline and the alleged infringement date strongly suggest, as we noted in December), then there may be no training infringement to adjudicate. The ECS puts it plainly: the reference “does not provide a definition, let alone findings,” of the training process it asks the Court to rule on.
There is also a jurisdictional wrinkle. Google denies that training occurred in Hungary. The referring court does not take a position on this, but if training happened elsewhere, Questions 2 and 3 are entirely hypothetical. The CJEU has consistently said that it is not in the business of answering hypothetical questions.
The wrong subject matter, the wrong law
Even setting admissibility aside, the ECS makes a second, consequential argument: this case is not the AI-and-copyright landmark that commentators are treating it as.
The claim before the Budapest court is based solely on the press publishers’ right under Article 15 of the CDSM Directive, not on authors’ rights in original works. That is a narrower, more limited right, designed to give press publishers leverage against platforms that display snippets of their content, not a general instrument for policing AI training. The ECS is direct: a ruling focused on Article 15 “is not a suitable vehicle for fixing fundamental issues stemming from both AI model training and content generation practices.”
This matters for what the CJEU can and cannot say. Even if the Court rules on the substance, the ruling must be read within the scope of the press publishers’ right, which comes with significant limitations: it covers only “online use” by information society service providers, excludes the use of individual words or very short extracts, and protects the publisher’s organisational and financial contribution rather than the originality of authorial expression. A ruling that conflates these frameworks, treating Article 15 questions as if they resolved questions about authorial reproduction rights, would make confused law from confused facts.
Language as infrastructure, and what happens when it breaks
The ECS opinion, at its technical level, is a complaint about a court that cannot distinguish among the components of a technology stack. But the implications run deeper.
In my book AI Tools, Not Gods, I argue that one of the most persistent risks in AI governance is the migration of metaphor into legal or regulatory status. When a lawmaker or a judge describes an AI as “reading,” “memorising,” or “communicating,” they are not just reaching for a convenient shorthand. They are importing an entire framework of rights, obligations, and moral assumptions that was designed for human actors engaged in human-scale acts. The words carry the legal architecture with them.
The Like Company reference is a case study in this dynamic. The plaintiff’s argument rests on a chain of humanising verbs: the AI ingested, stored, recalled, and broadcast. Each step in that chain corresponds to a legal concept, and each legal concept carries an enforcement mechanism. The problem is that none of the verbs accurately describes what an LLM does. As we have previously explained, training destroys the source text; it does not store it. The model does not “remember” the dolphin story; it adjusts weights that encode statistical relationships between tokens. What might look like recall is pattern completion, not retrieval.
The ECS, writing from a purely legal standpoint, arrives at the same destination from a different direction. Their concern is not philosophical but procedural: a reference built on the interchangeable use of “chatbot,” “LLM,” and “search engine” cannot generate legally precise questions, and imprecise questions generate unreliable doctrine. They are describing, in the language of legal scholarship, exactly what happens when borrowed metaphors do the analytical work that technical understanding should be doing.
What the Court should do
The ECS proposes a reordering of the questions. Training (Questions 2 and 3) should be addressed first, and only if the Court is satisfied that the training activities at issue actually fall within the scope of Article 15. If they do, the TDM exception under Article 4 CDSMD should, in the ECS’s view, cover them, provided the opt-out conditions are met. The display questions (Questions 4 and 1) should be addressed together, and the answer to Question 4 should, on the current facts, be negative.
The most pointed recommendation is the implicit one: the Court should consider declaring the reference entirely inadmissible, leaving room for a future, better-constructed referral that actually describes the technology, identifies the applicable law, and establishes the relevant facts.
That would not be a failure of EU copyright law to engage with AI. It would be EU copyright law working as it should: carefully, on the basis of established facts, resisting the pressure to answer questions the case has not actually posed.
The danger of doctrine built on metaphor
The ECS flags a risk that goes beyond the outcome of this particular case. A rash decision could push Europe toward a licensing economy in which AI systems are offered as a service by non-European multinationals, while the questions of creator consent and remuneration that actually matter for a healthy creative sector remain unresolved. Copyright rules shaped by anthropomorphised descriptions of technology do not protect creators. They protect the parties with the most leverage to negotiate licensing terms, which are rarely the individual journalist, musician, or author.
This is the cost of letting metaphor do legal work. Not only does it produce doctrine that does not fit the technology, it shapes that technology’s development in ways that serve whoever can most effectively exploit the ambiguity. The platform that can afford to license at scale wins. The smaller player that cannot loses. And the creator, whom copyright law exists to protect, ends up somewhere in the footnotes.
The CJEU has an opportunity here. Not to answer a question that has not been properly asked, but to signal that the quality of a legal reference matters, that conceptual precision is not a formality, and that courts cannot build durable doctrine on the shifting ground of anthropomorphised vocabulary.
The dolphins, it turns out, are still in the wrong lake. The question is whether the judges will notice before they rule.
Get the article’s core concepts in a unique auditory summary—listen to the AI-generated track below!
This song was created entirely using artificial intelligence tools. If you enjoy this experiment, keep an eye out for each song illustrating our December articles.
Written by Caroline De Cock, LL.M., Head of Research
