On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Like Company, C-250/25 – the first generative-AI and copyright referral to reach the Court of Justice of the European Union. The dispute comes from Hungary: the press publisher Like Company sued Google over the alleged use of its articles to train the Gemini large language models and over the reproduction and making available of its content through the chatbot. As The IPKat and the Kluwer Copyright Blog both report, the Opinion runs to 240 paragraphs, and its headline move is procedural: the AG advises the Court to declare the two questions about LLM training inadmissible and to answer only the two about what the chatbot hands back to the user.
The reasoning behind that split matters more than the split itself. According to the Kluwer analysis, the AG found no evidence that Like Company’s articles had been used to train Gemini at all, and considered the example before the national court better explained by retrieval-augmented generation – the chatbot fetching the article in real time and using it as context (paras 46-56; the AG’s blunt view that the main proceedings have “nothing to do with LLM training” is at para 50). He also held that training is not an “online use” of a press publication, so it falls outside the press publishers’ right in Article 15 of the DSM Directive – the only right Like Company invoked. Having said the Court need not answer, the AG answered anyway, in the alternative: training generally involves reproductions under Article 2 of the InfoSoc Directive, but those reproductions are as a rule covered by the Article 4 DSM text-and-data-mining exception, provided its conditions – lawful access, and no reservation of rights – are met (paras 97-152). He read the TDM definition in Article 2(2) broadly enough to cover extracting linguistic patterns, and concluded the three-step test is not, as a rule, breached, with exceptions decided case by case.
Where the AG does see infringement
On the questions he would have the Court answer, the AG proposes two alternative routes to liability for a chatbot that retrieves and reuses a press publication: the provider accessed the publication unlawfully – for instance by circumventing effective technical measures or scraping past them – or the answer reproduces the publication’s form of expression recognisably enough that the user can dispense entirely with reading the original (paras 229 and 240, per Kluwer). Mere repetition of facts is not enough: Article 15 protects the form in which facts are presented, not the facts. But paraphrase is not a safe harbour either, because arrangement, structure and phrasing can remain recognisable. Retrieval copies made during RAG are, on his analysis, outside Article 4 – retrieval is not pattern extraction – though they may qualify as temporary copies under Article 5(1) InfoSoc, which is where lawful access does its work. Responsibility can be attributed to the provider where the conduct is inherent in its business model or where it has not implemented the measures expected of a reasonably diligent operator; a user typing the prompt does not automatically shift the blame. Applying that framework provisionally to the facts, the AG suggested the conditions did not appear to be met here – the article’s expression was not recognisable in the response and Google appeared to have mitigation measures in place – leaving the final call to the Hungarian court (para 239).
What this changes
For publishers and other rightholders, the first lesson is pleading discipline. The training questions did not fail because training is lawful; they failed because the record did not show that these articles trained this model, and because the related right asserted does not reach training in the first place. If you intend to litigate AI training in the EU, assert copyright in the works, not only the press publishers’ right, and build the evidentiary record before filing – crawler logs, access records, dated captures of outputs, and a side-by-side showing recognisable expression rather than shared facts. The AG’s threshold is demanding: the answer must substitute for the original, and for the making-available route the expression must be systematically recognisable across related queries (paras 210-216). A one-off bad answer is not a case.
The second lesson is that the opt-out is now the main lever against training, and it is an operational task, not a legal one. If Article 4 covers training as a rule, then an unreserved, lawfully accessible archive is an archive that can be trained on. That makes an audit of how your rights reservation is actually expressed – machine-readable signals, site terms, and the access controls behind them – the highest-value hour anyone in a content business will spend this quarter. Access controls now pull double duty: they condition the lawful access requirement in Article 4 and they create the circumvention route to liability for retrieval. The AG also floated, as policy, that Article 4 could become a compensated limitation, and he pressed for legislative clarification on territoriality – a question on which, as The IPKat noted, the European Commission’s public consultation closed the same day the Opinion was delivered.
For anyone building or deploying an AI product in Europe, the risk has moved downstream to the retrieval and output layers, and the mitigations are design decisions you can make today: never fetch past an access restriction, keep retrieved text out of the response’s expressive form, attribute and link out, and ensure the answer is not a substitute for the source. Document those safeguards, because on the AG’s framework provider liability keys on the design of the service rather than occasional errors – and documented mitigation is part of what appeared to save Google provisionally at para 239. Non-EU trainers should not read the territoriality analysis as a free pass: while the AG doubted that merely placing a model on the EU market justifies applying EU copyright to foreign training, he did not exclude EU law applying to scraping from servers located in the EU (paras 60-73). Where your crawler reaches, not where your GPUs sit, is the question to answer in your diligence.
Two further consequences are worth tracking. If the Court follows the AG’s view that a trained model does not store the works it learned from – he calls the contrary inference far-fetched at para 90 – then part of the reasoning in the Munich Regional Court’s GEMA judgments against OpenAI and Suno is in question, which bears directly on where in Europe a rightholder should sue. And this is an Opinion, not a judgment: it does not bind the Court, and two further Hungarian referrals on AI and copyright, Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26), are already pending. The sensible response is not to restructure a business on an Opinion, but to fix the two things it makes decisive – what your rights reservation says, and what your retrieval layer does – and to price AI content-licensing deals on the assumption that training may be exempt while output reuse is not. Those are contract and licensing terms that can be revised now, before the judgment lands. US rightholders watching the same question through a different lens should also follow the May 2027 Meta AI training trial, where the evidentiary burden on which works were used is being fought out on the facts rather than on admissibility.