The Dolphins That Were Never There: Advocate General Szpunar on Gemini, Press Publishers and the First AI Case Before the Court of Justice (C-250/25)

Robert Nogacki, Attorney at Law, Managing Partner, Kancelaria Prawna Skarbiec. Warsaw, 4 November 2026

On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Case C-250/25, Like Company v Google Ireland (ECLI:EU:C:2026:919), the first dispute about generative artificial intelligence to reach the Court of Justice of the European Union. Its substance fits into two sentences, each of which will please one side of the argument only half-way. Training a language model on freely accessible press articles falls, in principle, within the European exception for text and data mining, and the press publishers’ right does not reach training at all. A chatbot that fetches an article in real time and serves its contents so completely that the reader no longer needs to visit the publisher’s site may, however, infringe that right, and the provider of the service answers for it where the effect is built into its business model or follows from the absence of safeguards. For publishers, who in Poland have just watched their mediation with Google collapse and the competition authority bring charges against the company, and for every firm building its own AI tools, this is the first map of the terrain they will be walking for years.

An Advocate General’s Opinion does not bind the Court, but the judges follow it in most cases, and in cases about new technology they rarely have a better starting point. The judgment will probably come in 2027. The document deserves a careful reading, because beneath its technical vocabulary lie several decisions about who, in Europe, will be paying for information.

A Lake, a Singer, and a Summary With a Pool in It

It began with dolphins. On 21 July 2023, the Hungarian portal balatonkornyeke.hu, owned by a company called Like Company Kft., published a short item reporting that Kozsó, a Hungarian pop singer, had not abandoned a dream he had cherished for more than twenty years: to bring freshwater dolphins to an aquarium on Lake Balaton. The piece was brief, informational, and borrowed heavily from what the singer had told the tabloid Blikk. Some time later an unknown user asked Google’s chatbot, then still called Bard, whether it could summarise, in Hungarian, the article on balatonkornyeke.hu about the plan. The chatbot obliged with enthusiasm and added details that appeared nowhere in the article: a special pool on the Tihany peninsula, 25 metres by 15 by 5, holding 1,000 cubic metres of water, critics worried about the environment, a project at the draft stage. Google called these hallucinations. The Advocate General called the summary “rather approximative” and counted that exactly one sentence of the original, the first, had made it into the answer. How easily a model invents what it does not know is something we have written about before (Attorney ChatGPT).

Like Company drew the opposite conclusion. If Gemini “knew” the article, it must have been trained on it; if it summarised the article, it had reproduced it and made it available to the public. The company sued Google Ireland before the Budapest Környéki Törvényszék, the regional court for the Budapest area, for a declaration that between 13 June 2023 and 31 May 2024 the company had continuously infringed its rights as a press publisher. Notably, Like Company did not rely on copyright in the texts. It relied solely on the related right of press publishers under Article 15 of Directive 2019/790, a choice the Advocate General records in a footnote and which, as will become clear, decided the fate of half the case.

The Hungarian court sent four questions to Luxembourg: two about training (is it reproduction, and does the text and data mining exception cover it) and two about the chatbot’s answers (is displaying them communication to the public and reproduction for which the provider is liable). The European Commission asked the Court to throw the whole reference out as too vague about the technology. The Advocate General declined, writing that the national court was “evidently lost in the AI landscape” and that the spirit of cooperation required guiding it rather than sending it away (point 35). Nine governments, including Poland, filed written observations, and the hearing took place on 10 March 2026. The Opinion was scheduled for 3 September, then vanished from the Court’s calendar for two months, long enough, as the IPKat blog noted, for conspiracy theories to grow. The finished text runs to 240 points and 243 footnotes, and its author warns at the outset that it significantly exceeds the customary length and that some of its views are provisional, to be refined in future cases (point 3).

That the case was never about training had been predicted in May 2025 by Paul Keller of Open Future, in a post on the Kluwer Copyright Blog titled Do AI models dream of dolphins in lake Balaton?, which the Advocate General cites. The Advocate General found it implausible that an article from July 2023 could have been in the training data of a model released in December 2023, because training large models takes months and the dataset closes long before launch (point 49). On his reconstruction, the chatbot had not remembered the text; it had read it, because the user supplied the domain name and the topic (point 50). How much that reconstruction is worth I return to below; the difference between memory and reading is nonetheless the axis of the whole Opinion.

What a Press Publisher Actually Owns

The press publishers’ right is a creature of 2019, designed with aggregators such as Google News and media monitoring services in mind. Article 15 of Directive 2019/790 gives publishers the rights of reproduction and of making available to the public over their press publications, but only for “online use” by information society service providers. It does not cover private or non-commercial use by individual users, hyperlinking, or individual words and very short extracts, and recital 57 adds that it does not protect mere facts. It expires two years after the year of publication. In Poland the right has been in force since 20 September 2024 under the Act of 26 July 2024, as Article 99⁷ and following of the Copyright and Related Rights Act (ustawa o prawie autorskim i prawach pokrewnych), with two local choices: the authors of the works contained in a publication, in practice the journalists, are entitled to 50 per cent of the publisher’s remuneration from this right (the duty to share derives from the Directive; the rate is Poland’s), and disputes over the amount go to mediation before the President of the Office of Electronic Communications (UKE), who may issue a binding determination if mediation fails. The pricing machinery is, by European standards, unusually elaborate.

The Court had spoken about this right exactly once before, on 12 May 2026, in Case C-797/23 Meta Platforms Ireland v AGCOM, the Italian communications regulator. It held that a national authority may set fair compensation for publishers and that a platform holding the data on revenues must disclose them, since without that disclosure negotiation is a fiction; the mechanism may not, however, turn the exclusive right into a compulsory paid licence, so the publisher remains free to refuse or to consent for nothing. That judgment concerned the nature of the right, not its infringement. The Like Company Opinion is the first attempt to say when the right is broken and what, precisely, it protects: not the news, but its form, that is, the way an article is arranged, structured and phrased (point 199). The Advocate General ties this to the purpose of the provision, which is to protect the investment that goes into collecting, verifying, analysing and presenting facts, and to the freedom of information under Article 11 of the Charter: in a democratic society nobody may be granted an exclusive right over information as such (point 200).

Training: The Questions the Court Is Told Not to Answer, and the Answers It Got Anyway

The most surprising move in the Opinion concerns the scope of the right itself. Training a model is not “online use” of a press publication. Online use, as the Advocate General defines it, is any use that occurs as an integral part of providing an information society service over the internet, at the individual request of a recipient (point 55). Training happens on dedicated infrastructure, with no service provided to anyone, and the link between the model’s observation of an article during training and a chatbot’s later answer is “simply too indirect (and, usually, diffuse)” (points 56 and 57). The consequence is plain and, for publishers, unpleasant: the right they lobbied for in 2019 gives them no weapon against the training of models on their articles. What remains is the copyright of their journalists, usually transferred to the publisher by contract, but that is a different basis and a different lawsuit. The second and third questions are therefore hypothetical, and the Advocate General advises the Court to declare them inadmissible.

He then devotes more than ninety points to them “for the sake of completeness,” with the caveat that he cannot stress their sensitivity enough (point 43). It is the first statement of such length to come out of Luxembourg on the legal status of training, and it deserves an honest summary, because in Luxembourg silence can be as eloquent as words.

First, training involves reproductions. Downloading articles to servers, cleaning and tokenising them, the copies made in working memory during the training loop: all of these are reproductions within Article 2 of Directive 2001/29, because the EU legislature chose a purely technical conception in which the purpose of a copy is immaterial (point 81). The Advocate General rejects the analogy to a human learning to write by reading: it cannot hold, “if only because of the industrial scale of the process” (point 83). He notes in passing, in a footnote, that Anthropic, when training its Claude model, bought physical books and scanned them page by page, as the American court described in Bartz v Anthropic, a case later settled for 1.5 billion dollars.

Second, a trained model is not a database. Its parameters reflect how words are used across the dataset as a whole, not in any particular text, and at that level of abstraction copyright does not exist (point 85). The Advocate General is aware of “regurgitation,” the phenomenon of a model reciting fragments of its training data, but treats it as a glitch caused by overfitting on duplicated texts rather than as proof that copies sit inside the weights (points 88 and 90). The output itself may infringe and must be assessed case by case (point 89), but holding that the model contains reproductions would mean that merely offering a model requires the consent of every rightholder, an “excessive liability burden” (point 92). Here he parts company with the Munich Regional Court, which in its first-instance judgments in GEMA v OpenAI (11 November 2025) and GEMA v Suno (31 July 2026) held that works memorised by a model are reproductions within the model, and sides with the English High Court, which on 4 November 2025 in Getty Images v Stability AI ruled that model weights do not store copies of images; that judgment addressed specific claims about Stable Diffusion, including whether the model itself was an “infringing copy,” and is not a general proposition about every model.

Third, the text and data mining exception in Article 4 of Directive 2019/790 covers training in principle. The definition in Article 2(2), an automated analytical technique aimed at generating information including patterns and correlations, is broad enough to accommodate machine learning, and a study commissioned by the European Parliament’s legal affairs committee during the legislative process expressly flagged that machine learning uses the same algorithms (point 108). The legislature confirmed the intention in recital 105 and Article 53(1)(c) of the AI Act, which presuppose that the exception applies to the training of general-purpose models (point 109). Denmark and Spain argued that in 2019 the legislature had big data analytics in mind, not generative models. The Advocate General replies that this is not decisive. The passage contains a small factual slip: in reporting the argument he writes that GPT-1 was released a few months after the Directive was adopted, whereas OpenAI presented that model in June 2018, ten months before. The error changes nothing in the conclusion, but it shows how thin the layer of technical fact is on which the law is being built.

Fourth, the three-step test of the international conventions is no obstacle. Training is a “special case” because the exception has clear conceptual contours, not because it touches few works (point 115). The existence of a market for training licences does not settle the conflict with normal exploitation, since that argument would abolish every exception (point 123). A model’s capacity to generate content that competes with human work is “too diffuse” a consequence, and copyright does not protect against competition as such (points 128 and 130). Here, however, the Advocate General adds a condition of real practical weight for developers: where a specific model can routinely generate content recognisably resembling specific items in its training data, the developer must show that it took mitigating measures, such as dataset diversification and deduplication and prompt and output filters, failing which the national court should refuse the exception (points 132 and 133). The text and data mining exception, in other words, is not an unconditional right to train but a privilege conditioned on the architecture of the model.

Fifth, the two statutory conditions keep their teeth. Access must be lawful, which means no scraping of pirate sites and “shadow databases” and no circumvention of effective technical measures, including paywalls and bot blocks (points 141 and 142). Rightholders may reserve their rights in a machine-readable way, and the Advocate General reads this pragmatically: a reservation is effective if state-of-the-art robots and associated technologies can understand it reliably at the moment of access (point 148). Robots.txt qualifies, Google-Extended qualifies, metadata embedded in files qualifies, and he does not exclude that a reservation written in plain language in a website’s terms and conditions qualifies too, since language models can now read terms and conditions (point 149). In Poland the collecting society ZAiKS has already filed such a reservation on behalf of the authors it represents. Two things that public debate tends to merge must be kept apart here: a reservation in robots.txt is a declaration a bot is expected to honour, whereas an effective technical measure within Article 6(3) of Directive 2001/29 is a device that actually prevents access; ignoring the first forfeits the text and data mining exception, circumventing the second makes the access unlawful. To this the Advocate General adds a policy proposal: nothing prevents Member States from introducing statutory remuneration for creators for uses covered by the exception, on the model of private copying levies (point 151).

Finally, territoriality. The Gemini models were trained in the United States, and Google argued that EU law does not reach servers across the ocean. The Advocate General concedes that a “centre of gravity” theory of training is simple and predictable but would leave European rightholders with no protection at all (point 65). The effects doctrine, favoured by every government except Germany’s, he finds too elastic, because the effects of training are indirect (points 67 and 69). He does not exclude the middle course proposed by Berlin: EU law covers the scraping of works from servers located in the Union, wherever the training later takes place, and copies of the data may not be exported against the rightholders’ reservation (point 71). He admits that server location is a poor connecting factor in the age of the cloud and calls on the legislature to settle the matter expressly (points 72 and 73).

The call landed at the right moment. The European Parliament, in its resolution of 10 March 2026, adopted by 460 votes to 71, demanded that EU copyright apply to every generative system offered on the internal market regardless of where it was trained, and proposed an opt-out register run by EUIPO and remuneration for past uses. The European Commission is running a consultation on copyright and artificial intelligence whose original deadline, 3 November 2026, coincided with the day the Opinion was published and has since been extended to 10 November. A legislative proposal is hard to imagine before 2027. The Advocate General has therefore written a document that is at once a gloss on the law as it stands and a memorandum to the lawmaker, and the two layers must be read separately.

Retrieval, or the Search Engine That Reads for You

The real dispute begins at point 153, where the Advocate General turns to “what this case is really about.” A language model on its own is a generator of fluent text, not a source of knowledge. Asked about facts, it may answer flawlessly as to form and wrongly as to substance, and when it does not know, it invents (point 154). To cure this, engineers bolted onto chatbots a technique called Retrieval-Augmented Generation, RAG for short, also known as “grounding.” When a user asks something factual, the system searches the Google index, fetches passages from web pages with a bot, appends them to the prompt given to the model, and only then generates an answer, sometimes with links to the sources and sometimes without, and why one or the other remains, as the Advocate General admits, unclear (points 13 and 155). The same technique powers AI Overviews and AI Mode in Google Search.

The difference from a classic search engine is fundamental: the search engine gave a list of links one had to open; the chatbot collects, pieces together, sums up and delivers the content in full sentences (point 156). The Advocate General describes this in the language of economics rather than law. The original business model of the web rested on a delicate quid pro quo: search engines copied sites and displayed snippets, but in return they sent traffic the publisher could monetise. That bargain is called into question when the chatbot delivers the information itself and nobody clicks anywhere (point 159). In a footnote he records the studies on “zero-click” searches and Google’s rejoinder that links inside AI Overviews receive more clicks than ordinary listings. The data favour the publishers. According to a Pew Research Center study from July 2025, based on 68,879 searches by 900 American adults, users clicked a traditional result in 8 per cent of visits when an AI summary appeared at the top of the page and in 15 per cent when it did not, and clicked a link inside the summary in 1 per cent. The study is observational and concerns the search engine rather than a chatbot, so it proves nothing about Like Company’s traffic, but it does show that a link in an answer is not the economic equivalent of sending the reader to the source.

Fetching an article through a chatbot is, in the Advocate General’s view, online use of a press publication, because it occurs as an integral part of a service provided over the internet at the individual request of the user (point 169). Territoriality poses no problem here: a service offered to users in the Union produces immediate and substantial effects there, wherever the provider is established and its servers stand (points 165 and 166). He also proposes a solution to the evidentiary problem that decides such disputes in practice. Since chatbot providers are under no duty to disclose what they scrape, courts should be lenient with the publisher’s burden of proof: it is enough that the content was available online and that the publisher produces a reasonable sample of chatbot answers resembling its publication or linking back to it for use to be presumed, shifting the burden to the provider to prove the contrary (point 163). A single screenshot will rarely do. It is the same direction the Court took with banks in the unauthorised payments case (The Bank Pays First: Case C-70/25 and the Refund of Unauthorised Payment Transactions): whoever controls the technology and the data must do the proving.

On the invisible technical copies created while an article is fetched and processed, the Advocate General is precise. The text and data mining exception does not cover them, because retrieval is not about extracting patterns from a large corpus but about using a specific piece of content to improve a specific answer (point 179). The temporary reproduction exception in Article 5(1) of Directive 2001/29 may: the copies are transient, form an integral part of the process and have no independent economic significance, since their value materialises only in the answer (points 182 to 184). One condition remains: the copies must enable a “lawful use,” and that depends cumulatively on two things, how the article was accessed and what form the answer takes (point 187). If the chatbot circumvented a paywall or a bot block, the use is not lawful, and the circumvention is additionally a tort under Article 6(1) of the Directive (point 188 and footnote 240). If the article’s form of expression remains recognisable in the answer, the answer itself becomes a reproduction.

Facts Are Free, Phrasing Is Not

Here we reach the heart of the Opinion. Whether a chatbot’s answer reproduces a publication is to be judged by the Court’s ordinary infringement test, with no special rule for AI, since such a rule would offend the principle of technological neutrality (point 193). The test has two elements: copying and recognisability. Copying is satisfied by retrieval itself; Google’s argument that the article goes into the prompt rather than the answer does not break the causal chain, and the reproduction right covers indirect copying (point 195). Recognisability requires a side-by-side comparison showing that the protected elements, the form of expression rather than the facts, are perceptible in the answer beyond the threshold of individual words and very short extracts (points 196 and 202).

To this the Advocate General adds a factor he himself calls decisive: whether the answer is a direct substitute for the publication, “because a user, by reading that answer, can dispense with accessing the original article entirely,” thereby undermining the publisher’s investment (points 202 and 207). Verbatim copying, of the kind the New York Times exhibited in December 2023 in its complaint against OpenAI, settles the matter at once, but it rarely happens today, because models paraphrase and Google is said to apply output filters that screen out longer matches with the source (point 203). Paraphrase, however, is not a safe harbour. If it were, the publishers’ right could be circumvented with a single instruction to “put it differently”; so argued Like Company and the governments of France, Italy, Denmark, Spain, Poland, Hungary and Greece, and the Advocate General agreed (point 205). What matters is whether, despite the rewording, the arrangement, structure and general phrasing of the article remain recognisable.

From this follows a distinction with financial consequences for different kinds of journalism. The more distinctive a text, the more easily it is recognised in someone else’s answer; long analyses are by nature more recognisable than purely informative reports, in which information and expression are inseparable and one item looks like the next (point 206). Where the answer reports the same facts in a different structure, as a short summary or as an overview assembled from several sources and reorganised by the model, there is no reproduction, even if the answer diverts readers from the publisher, because the publishers’ right “was never meant to protect those publishers against competition per se” (point 208). The Advocate General concedes a grey area between the two poles, to be resolved case by case (point 209).

For the right of making available to the public under Article 3(2) of Directive 2001/29 the bar is higher. A chatbot makes a publication available only if its form is “systematically” recognisable, to the point of substitution, in answers to related queries (point 213). He thinks this unlikely for a technology whose answers are non-deterministic, varying with the model’s parameters, the wording of the query and the history of the conversation, but allows for it as the result of intentional or negligent design of the retrieval process, which the publisher would have to demonstrate with many examples of ordinary queries rather than prompts engineered to trick the chatbot (point 214 and footnote 221). We have recently described how hidden instructions of that sort can lead even a court astray (A Judge Found Hidden Instructions for AI in a Court Filing and Sanctioned the Plaintiff. It Is the First Such Ruling in the U.S.). Where making available does occur, it uses, in his view, technical means different from publication on a website, since transmission through a predictive generation process differs radically from transmission from server to browser, so the “new public” inquiry becomes unnecessary; should the Court disagree, a new public would appear only where the chatbot circumvented access controls (point 216 and the footnote to it).

Applied to the Balaton dolphins, the test does not favour Like Company. The Advocate General lists four reasons: the company used no technical protection measures; the article’s form of expression is not recognisable in the answer; any reproduction through Gemini does not appear inherent in Google’s business model; and Google seems to have put in place measures mitigating the risk (point 239). A footnote adds the barb that the article itself is a purely informative report of commonplace structure, much of it borrowed from other publications, and point 30 observes that Like Company’s portals “seem far removed” from the publishers of traditional newspapers the legislature had in mind, so the national court must first decide whether this is a press publication at all. Not everyone who publishes is a press publisher, just as not every investor is a consumer (A Charter Signed in a Bank Branch).

Who Is Liable: the User or Google

The publishers’ right applies only to use “by information society service providers.” If the act is attributed to the user, the claim fails, because private use is outside the right altogether. The Advocate General accepts Google’s position that when a user directs a chatbot to perform a copyright-relevant act, primary liability lies, as a rule, with the user; attributing to the provider everything users might devise would amount to strict liability with a chilling effect on the whole technology (points 219 and 220). The provider is nonetheless liable in two situations drawn from the case law on platforms, including the YouTube and Cyando judgment: where such acts are inherent in its business model, and where it has failed to put in place the measures expected of a reasonably diligent operator (point 221). The footnoted example of the first situation is telling: a service designed to deliver extensive, merely paraphrased presentations of individual news articles is functionally indistinguishable from the aggregators the right was created against.

The standard of care is spelled out. Whoever designs and controls a technology as advanced as a chatbot must assume the risks inherent in it, all the more so since the provider earns revenue from it (point 222). The bots used in retrieval must respect the access restrictions publishers put in place, and the process must be tuned so that answers are sufficiently removed from any specific publication retrieved, for instance through output filters. Where such measures operate, occasional mishaps create no liability; where they are absent, liability is warranted. A footnote notes that analogous duties may arise under the Digital Services Act once a chatbot is designated a very large online search engine; the Commission gave ChatGPT that designation on 31 August 2026.

An Open Website Is Not Consent

Google’s broadest argument was that a publisher who leaves its articles open, without paywall or bot blocks, implicitly consents to their use by chatbots. The Advocate General’s reply is three words long: “I do not think that is the case” (point 225). The Court did accept implied consent to hyperlinking in VG Bild-Kunst, but implied consent requires at a minimum that the rightholder was actually and individually informed of the use and had simple means of preventing it (point 227). Neither is present. Chatbot providers did not inform publishers, and how publications are selected remains unclear. A paywall is a serious commercial decision; a publisher may forgo it because it prefers advertising revenue from broad traffic, not because it consents to the reuse of its content (point 228). Bot blocking, for its part, does not distinguish purposes: the bots that ground chatbot answers are not fundamentally different from those that index pages for search, and at Google they appear to be the same bots, so blocking them in robots.txt risks disappearing from search results. From the fact that a publisher tolerates indexing in order to stay visible, no consent to feeding a chatbot follows (point 228). Google’s claim that Google-Extended allows the two to be separated without affecting search “would need to be verified” (footnote to point 228).

Reality has partly outrun the case file. On 3 June 2026, the day the UK Competition and Markets Authority imposed a conduct requirement on its search business, Google announced a Search Console control that removes a site from AI Overviews, AI Mode and the AI features in Discover and stops its content being used for grounding, without affecting ranking; testing began with a subset of UK sites, and on 31 August 2026 the control reached every website worldwide. On 30 October 2026, four days before the Opinion, Google, Microsoft, OpenAI and Anthropic announced that they would honour a single shared token in robots.txt, __usage_ai-training, defined in an IETF draft, which lets a site refuse the use of its content for training in one line. For now it is an announcement and a draft, not an IETF standard: Google plans support in 2027, Microsoft from 1 February 2027, and the token itself is a signal of preference, not a technical barrier. The devil is in the mapping: Google intends to treat the token as a Google-Extended disallow, unless the publisher has written a separate rule for Google-Extended, and therefore also as a refusal of grounding in Gemini, whereas Microsoft confines it to training and handles grounding with a separate tag. The indistinguishability of purposes is exactly what tripped the implied consent argument, and exactly what will be the first question to the expert witness in the next dispute.

Where the Opinion Is Weakest: Five Objections

An honest critique must begin with what is strong. The Advocate General separates categories that public debate merges: model and service, training and retrieval, information and form, competitive harm and infringement, the territoriality of rights and the regulatory duties of providers. He refuses to be pushed either towards “AI reads like a human” or towards “the model is a library of stolen texts.” The weaknesses appear where, having drawn the distinctions, he reintroduces generalisations. The objections that follow are an assessment of the reasoning, not a prediction of the outcome; one can think at once that the Advocate General’s test is too narrow and that Like Company has not made out its claim.

The first objection concerns the threshold of “entire” substitution. Article 2 of Directive 2001/29 protects against reproduction “in whole or in part,” and Article 15 of Directive 2019/790 excludes only individual words and very short extracts; recital 58 warns against interpreting that exclusion in a way that deprives the right of effectiveness. Yet in the operative part proposed to the Court, reproduction occurs where the form of expression is recognisable “to such an extent that the user, by reading that answer, can dispense with accessing the original publication entirely.” Consider a chatbot that reproduces four substantial paragraphs of a piece of investigative reporting but omits the interviews, the photographs and the closing analysis. The reader still has a reason to visit the site, so on a literal reading of the proposed operative part there is no reproduction, although a protected part has been taken. The Advocate General would reply that he speaks of the recognisability of “all or part” of the publication and that his test also requires the taking of form, not merely of facts; the trouble is that the word “entirely” in the operative part attaches to the reader’s dispensing with the original, not to the extent of the taking, and it is that word which narrows the protection of parts. The objection runs the other way too: a single sentence giving a match result fully satisfies the reader’s need and is not a reproduction, because it conveys only a fact. Substitution, then, is evidence of harm and an aid to interpretation, but neither a necessary nor a sufficient condition of reproduction. The best defence of the Advocate General is that he wanted to cut off lawsuits over ordinary reporting of facts and trivial summaries; that aim can be achieved more precisely, by excluding facts and identifying the elements taken, without writing into the law a condition it does not contain. A hypothesis for the future: the Court may keep recognisability and lose the word “entirely” on the way.

The second objection concerns the weights. “The model is not a repository of documents” and “the model fixes no protected sequence” are two different sentences, and the Advocate General passes from the first to the second too smoothly. Article 2 is technologically neutral: it covers reproduction “by any means and in any form,” and twenty points earlier the Advocate General himself explains that the legislature chose a purely technical conception in which the purpose of a copy is irrelevant. It is hard to apply that principle to the training set and abandon it at the weights. The argument from randomness is the weakest: that a retrieval procedure does not always work does not mean the information is absent, just as a scratched disc still contains the album. Nor does the science settle the matter in his favour. The 2021 paper by Carlini and co-authors showed that verbatim sequences present in a single document of the training set can be extracted from a language model, which weakens the story of overfitting on duplicated song lyrics. A preprint first posted in March 2026 and updated since, co-authored by Jane Ginsburg, professor of copyright at Columbia, showed that fine-tuning GPT-4o, Gemini 2.5 Pro and DeepSeek on pairs of summary and excerpt, teaching the models to expand plot summaries into full text, reactivates verbatim recall of up to 85 to 90 per cent of the text of some books, including books not used in the fine-tuning, and that three models from different companies remember the same passages of the same books. It is an experiment on modified models, not on Bard in 2023, and the coverage metric aggregates short spans, so it must not be read as “the model contains 90 per cent of the book.” It suffices, however, to make a general dismissal of the “copy in the weights” category premature. The better path is a test of specific, protected, identifiable and reliably reproducible content, assessed after other sources, including the prompt, have been excluded. The Advocate General lets the problem in through the back door in any case, when he makes the text and data mining exception conditional on measures against memorisation.

The third objection concerns the transience of retrieval copies. The Advocate General assumes that technical copies are deleted at the latest once the answer has been generated (point 182). That is a property of a particular implementation, not of the technique. The original 2020 paper by Lewis and co-authors described RAG as the combination of parametric memory with non-parametric memory built on a document index to which the system returns. Gemini’s own API documentation describes context caching with time-to-live parameters; this does not prove that the consumer Bard used it, but it proves that persistence cannot be excluded from the name of the technique. For the training set the Advocate General rightly refuses the temporary copy exception on account of persistence, function and the independent value of the data (point 96), and admits in a footnote that it “is not at all clear that AI developers discard them afterwards.” For retrieval he applies gentler assumptions to the same criteria. That is not a contradiction if the architectures really differ, but the difference requires proof: what goes into the index, into the cache, into the logs, into quality evaluation and into the next round of training. Who stores what, and for how long, is the question that returns in every AI dispute, including the one about the confidentiality of conversations (Are Your A.I. Chats Confidential? Three Federal Rulings, a Fresh Supreme Court Decision, and the E.U.’s August 18 Deadline).

The fourth objection concerns the user as default actor. In YouTube and Cyando the content was uploaded by autonomously acting users and the platform stored it. In a retrieval system the operator chooses the sources, fetches the text, builds the context and generates the answer, while the user asks about an event. The question is not “did the user type something” but “who supplied the protected content and controlled the act that led to its use.” At least three situations must be told apart: the user pastes an article and asks for a rewrite; the user asks about an event and the system selects and fetches the publication on its own; the user deliberately circumvents safeguards and forces the recital of a specific text. The Advocate General’s scheme handles the first and third well and the second worst, and the second is the one publishers are actually fighting about. He does see the operator’s active role (points 221 and 222), but he resolves it at the level of diligence rather than attribution. Since the publishers’ right covers use by the service provider only, everything turns on this attribution, and the Court would do well to break it down by type of act rather than transplanting a hosting standard of care onto an actively organised service.

The fifth objection concerns facts rather than law. In point 49 the Advocate General sets the date of the article against the release of Gemini 1.0 in December 2023 and describes the period of infringement as June 2023 to February 2024, although in point 15 it ends on 31 May 2024. Bard ran from May 2023 on the PaLM 2 model, and Gemini Pro reached it in December 2023 in English first; the phrase “the Gemini LLMs” therefore does not say which model produced the Hungarian answer. The conclusion that no training took place may be correct, but it requires identifying the model and the date of the answer, not merely an argument from the calendar. The Advocate General himself warns in his fourth footnote that the Opinion was written by a lawyer rather than a computer scientist and that the technology is presented in a “vulgarised” way. That is an honest caveat, and one more reason for the Court to avoid technical generalisations in the operative part and to tie legal propositions to evidentiary tools: Articles 6 to 8 of Directive 2004/48 allow courts to order disclosure and preservation of evidence at the operator, a better route than expecting the publisher to reconstruct a service’s architecture from a chat window. A marginal note: the argument that the exception is a “special case” also because of its narrow circle of beneficiaries (point 116) is weak, since Article 4, unlike Article 3, does not confine beneficiaries to research institutions; the special character must be defended by the definiteness of the use, not the narrowness of the group.

What It Means in Practice, in Poland and Beyond

For publishers the Opinion is a map and a bill at once. Facts remain free, and no judgment will change that. What is protected is what the publisher actually invests in: structure, analysis, its own phrasing. The Advocate General says so in his forecast that artificial intelligence will not displace in-depth journalism but will easily replace texts of low originality, purely informative news and sports reports (point 136). A portal that lives by rewriting wire copy is in a weak position in this system for two reasons: its texts cannot be recognised in a chatbot’s answer, and their production will be taken over by the machine; what remains is protection against unlawful access, that is, paywalls and bot blocks, provided it deploys them. A portal that produces analysis gains both protection and bargaining power, because a chatbot cannot answer questions about the news well without access to the press, and the Advocate General openly hopes this will push providers towards licences (points 236 and 237). The first money is flowing, modestly: according to The Information as reported by PPC Land, Google pays about 100 publishers through a pilot for content used in AI Overviews, AI Mode and the Gemini app, some smaller sites receiving under 1,000 dollars over several months.

Technical measures change the legal status of content, and since 31 August 2026, when the Search Console control reached every site, they cost less visibility than they did in the period the case covers; the shared training token remains, for now, an announcement. A publisher who wants to negotiate should today have a robots.txt reservation for training, a separate decision on grounding, a plain-language reservation in its terms and conditions, since the Advocate General allows for its effectiveness, and a paywall or bot block for content it does not wish to give away. A publisher who has not reserved its rights tolerates, under this Opinion, training on lawfully accessible content and, against chatbots, has two routes: a claim of unlawful access, if it used protective measures, and the recognisability test. Evidence must also be gathered differently from the way Like Company gathered it: not one screenshot of a prompt that names the domain, but a predefined set of ordinary queries, many attempts, every result and every refusal, and a sentence-by-sentence comparison. For making available, repetition counts; for reproduction, a single answer may suffice; but the presumption of point 163 is triggered only by a “reasonable sample.”

The Polish dispute over remuneration runs on a different track, which the Opinion does not close but does illuminate. The Association of Digital Publishers (ZPWC), whose members include Agora, Wirtualna Polska, Ringier Axel Springer Polska, TVN Warner Bros. Discovery and Interia, failed to reach agreement with Google after a year of negotiation and 60 days of mediation before the President of UKE and announced an application for a regulatory determination; an analysis by FehrAdvice puts the remuneration Polish publishers are seeking at up to 514 million złoty, while Google replies that it settles with publishers under its own Extended News Previews model, agreed with more than 5,000 news sites in 24 countries. On 5 October 2026 the President of the Office of Competition and Consumer Protection (UOKiK) brought charges against four Google group companies for abusing a dominant position in negotiations with publishers, because the company withheld the data on how it uses publications and what it earns from them, an infringement that carries a fine of up to 10 per cent of turnover. In parallel, the European Commission has since December 2025 been investigating the use of publishers’ content in AI Overviews and AI Mode under competition law. Google knows what losing to regulators costs, as we have written before (Four Billion Euros Later: The Court of Justice, Google Android, and the Weaponisation of Competition Law). What the Opinion adds to this picture is one clarification: the Polish remuneration rules concern online use covered by the publishers’ right, that is, snippets in search, Google News and Discover; they do not reach the training of models, and they reach a chatbot’s use only where it fetched the publication by circumventing protective measures or where the answer takes over the form of the article. Whoever wants to price “the use of content by AI” as a single line item will have to break it into three different acts with three different legal statuses. Our media law practice has been doing exactly that for publishers and platforms (Media Law).

For firms building their own retrieval-based tools, from customer assistants to press monitoring, the Opinion reads as a checklist. The bot must read robots.txt and must not climb over paywalls. Answers should synthesise several sources, give links and encourage the reader to consult the original, because that is the service the Advocate General describes as synergy rather than substitution (point 234). Output filters screening out longer matches with the source are now the standard of care, not an extra. Logs and caches holding fetched texts should live briefly and have a documented purpose, because persistence forfeits the temporary copy exception. A service whose whole point is to serve up extensive paraphrases of individual articles is the very case the Advocate General describes as inherent in a business model, and it is hard to defend without a licence. These requirements can be treated as a cost or as an advantage: a company able to document the compliance of its architecture will have an easier conversation with publishers and with investors alike.

For developers and deployers of models the Opinion confirms that the text and data mining exception is real in Europe but conditional. Sources and their lawfulness must be documented, reservations honoured in every recognised form, a summary of training data published under the AI Office’s template, and measures against memorisation demonstrated, since the Advocate General makes them a condition of the exception (point 133). The AI Act’s obligations for general-purpose models have applied since 2 August 2025, with providers of models placed on the market earlier given until 2 August 2027 to comply; the Code of Practice of July 2025, which the Opinion cites in several footnotes, remains voluntary, but after this Opinion it is hard to treat it as anything other than a yardstick of what a court may expect from a reasonably diligent developer (AI Act implementation for businesses). European technology companies should also notice the thread of digital sovereignty: the Advocate General expressly allows the EU legislature to favour the training of European models on European content (point 139). In Brussels that argument will weigh more over the coming years than many an expert report.

What Happens Next

Google welcomed the training part of the Opinion with relief, stressing that the text and data mining exceptions are vital for Europe’s competitiveness and that Gemini learns linguistic patterns to create original outputs rather than replicate content. The European Publishers Council answered that publishers should not have to sacrifice visibility in search to protect their intellectual property. Both reactions are understandable and both miss what matters most in the Opinion: the shift of the dispute from “may you train” to “does the answer replace the original.” The first question the Advocate General resolved in favour of the technology, the second in favour of publishers, and only the second decides who pays.

The Court is not bound by the Opinion. It may follow it in full, it may also answer the training questions the Advocate General advised it to leave aside, or it may, as it has often done in precedent-setting cases, narrow its answer to what the Hungarian court strictly needs. Whatever it decides will bind every national court in the Union, which is the whole point of the preliminary reference (The Binding Force of CJEU Jurisprudence in Domestic Tax Law). Further references with an AI element are already waiting in Luxembourg, including a second Hungarian case on the press publishers’ right, Manchester United Fanatics Klub (C-806/26), with questions about training and grounding. Independently of the judgment, the Commission is preparing a legislative initiative, and Parliament demands that EU law reach training wherever it takes place. The Advocate General cautiously advises against that territorial expansion, and he is right about at least one thing: a law that reaches everywhere invites reciprocity, and European model developers have servers outside Europe too. The boomerang effect in law is nothing new (The Boomerang: How Instruments of Repression Turn on Their Makers), and in intellectual property it tends to be especially quick, because every party stands on both sides of the border.

Facts belong to everyone; sentences belong to someone. The dispute about artificial intelligence in Luxembourg began with dolphins that are not there and a pool nobody built, and it ends with the question of who will pay for the sentences someone wrote. Advocate General Szpunar has given an answer that will satisfy nobody completely, which may be precisely why it has a chance of becoming law.

Legal and factual position as at 4 November 2026.

Robert Nogacki

Robert Nogacki is a Polish attorney at law (radca prawny), the founder and managing partner of Kancelaria Prawna Skarbiec (Skarbiec Law Firm), which has operated continuously since 2006.

The law is equal for everyone, but the parties rarely are: on one side stands an organization with time, money, and lawyers, on the other a person with one business, one nest egg, and one life.

Clients rarely come to him with a legal problem. They come with a problem that also has a legal side: an audit that began with a single invoice, money entrusted to someone who has disappeared, a company that has to be passed on before it is too late. Most such matters are decided long before the first letter is written, in decisions made without asking and in deadlines nobody remembered. So he begins by asking how the client got here, not what the client should have done.

He advises entrepreneurs and families from more than a dozen countries, including those whose accounts the tax office has just seized and who do not know what to do tomorrow morning. He defends them in tax audits, customs and fiscal inspections, disputes with the tax authorities, and criminal tax proceedings. He represents victims of investment fraud and Ponzi schemes. He helps families set up family foundations and plan succession, so that a life’s work outlasts a single generation.

Not every case can be won. Every case can be run so that the client knows where they stand. Since 2006 he has represented the victims in the WGI case (Warszawska Grupa Inwestycyjna, the Warsaw Investment Group), one of the longest criminal cases in the history of the Polish financial market, because some things must not be left half finished, even when they take two decades. In the case of the collapsed cryptocurrency exchange Zonda (Zondacrypto, operated by BB Trade Estonia OÜ), he represents several hundred victims in the criminal investigation conducted by Poland’s National Prosecutor’s Office and in the Estonian bankruptcy proceedings.

Kancelaria Prawna Skarbiec is listed in the rankings of Poland’s largest tax advisory firms published by Dziennik Gazeta Prawna and Rzeczpospolita, and it is a four-time recipient (2015 to 2018) of the European Medal awarded by the Business Centre Club and the European Economic and Social Committee. Robert Nogacki publishes regularly, in the press and on the firm’s website, for people who have a problem rather than a law degree, because a legal opinion the client cannot understand protects only the lawyer.

He believes that the best legal advice is the kind that means the client never has to appear in court.