IP Law Daily, COPYRIGHT NEWS—European Advocate General: AI training can fall within copyright exception, (Nov 5, 2026)

By Matthew Hersh, J.D.

The European Court of Justice's First Advocate General weighs in, though his opinion is advisory and nonbinding.

The training of artificial intelligence language models on copyrighted content can be protected under European law as long as that content is not harvested in violation of online access controls and reasonable measures are taken to ensure that the content is not regurgitated to users, the court’s First Advocate General has argued. The lengthy submission by the court’s advocate general, in a case currently pending before the tribunal, also addresses a wide range of other legal issues arising from the use of copyrighted content by AI models (Like Company v. Google Ireland, Ltd., Opinion of European Advocate General, Case C-250/25, (Nov. 4, 2026)).

The opinion arises out of a lawsuit pending in the Hungarian courts. Like Company Kft., a Hungarian news company, sued Google for infringement in a Budapest court. The gist of the lawsuit was that Google infringed in two ways: first, in scraping the articles from the internet and using them to train its Gemini artificial intelligence model; and second, in reproducing the articles in response to user queries. As an example, the news company pointed to a 2023 article it published on the (perhaps-quixotic) efforts of a popular Hungarian singer, who goes by Kozsó, to introduce dolphins into Lake Balaton; according to the lawsuit, the Gemini chatbot reproduced aspects of the article when asked to give a summary of it.

The Hungarian court decided to stay the proceedings and to seek guidance from the European Court of Justice. That court then received written observations from the news company, Google, the European Commission, and several European national governments. The court’s Advocate General, Poland-born Maciej Szpunar, in turn issued an independent, nonbinding opinion analyzing the legal issues, which is summarized below.

The Court of Justice has now begun deliberating on the case; a judgment is expected at a later date, at which point the matter will be returned to the Hungarian courts for their continued consideration.

LLM Training—admissibility of the issue. AG Szpunar first discussed whether using news articles for training violated the news company’s rights. The AG first noted that the question was a hypothetical one and should not be addressed by the court, for two reasons. First, the news company never actually demonstrated that its articles were scraped during Gemini’s training process. Indeed, even in the example it gave of the Lake Balaton article, the article in question was published five months before the release of the Gemini version that allegedly reproduced that article—making it unlikely that Google had scraped the article and incorporated it into its training data. It was therefore more likely, the AG concluded, that to the extent Gemini based its answers on the article, it did so because it reviewed the article in real-time on the web, rather than resorting to its training corpus.

Even more importantly, the AG stressed, the issue was cabined by the limited scope of the news company’s lawsuit. News publishers who own the copyright to the works they publish enjoy the full range of exclusive rights in those articles. But regardless of copyright ownership, news publishers also enjoy a separate, more limited right against the “online use” of their publications by online service providers. The Hungarian news company sued only under this separate right—and that presented a problem, the AG concluded, because the use of news articles for LLM training did not constitute an “online use” of those articles. Rather, the AG noted, “such use occurs as part of the development of a piece of AI technology on dedicated computer infrastructure, without any information society service being provided to a user via the internet.”

But even though the AG concluded that the training question should not be addressed by the court, he nonetheless gave his opinion on the issue “for the sake of completeness and in order to provide a proper framework for the debate.”

LLM Training—jurisdiction. The AG next turned to a more knotty question. Google’s training took place principally on computer infrastructure in the United States, not in Europe. Did EU artistic and literary property law even reach that training at all? Limiting the regulation of training only to the nation where that training occurred, the AG observed, certainly had the advantage of simplicity, and therefore offered “a high degree of legal certainty to LLM developers.” But this would mean that European rightsholders would not receive “any protection when their works and other subject matter are used for the purposes of model training outside the Union.” Another consideration, the AG noted, was that the LLMs in question were made available to European users and therefore had harmful effects on European rightsholders. But this argument too had shortcomings, the AG noted, because after all, the issue here was not use but merely training—and that took place only in the United States.

Ultimately, the AG begged off the question, but not before suggesting that consideration be given to a “sort of middle ground” offered by Germany. Here, although the training took place in the United States, the scraping of the information used for the training could have involved servers located in Europe. At a minimum, if the owner of that information stored it on access-controlled servers in Europe, it would make sense for EU law to apply. “Copies of protected works and other subject matter, as data,” the AG noted, “should not be transferred outside the European Union for training purposes where the rightsholders oppose such a use.” All things considered, the AG concluded, it would be better for the European legislature to address the matter, but “some refined version” of the German solution might be worth consideration.

LLM training—whether it involves reproduction. The AG then addressed whether LLM training involved the making of reproductions of protected works. There was “little doubt” that it does, the AG found. For one thing, digital reproductions of protected works are usually made during the creation of the training dataset, when works are scraped from the internet and downloaded to the AI developer’s server. Moreover, various reproductions were made during the training process itself. Although some commentators objected that these digital reproductions do not satisfy a “functional” definition of the concept of reproduction, in that they were used only as data for the LLM to learn and not communicated to be read or otherwise enjoyed by anyone. But the AG rejected that argument because the reproduction right under EU law “was broadly defined in that provision to cover any digital reproduction of a protected subject matter,” regardless of its function, the AG noted.

The AG also addressed the somewhat knottier question of whether the trained model itself contained reproductions of the works. Importantly, the AG noted, a trained LLM is not a database, in the sense that it is “not a repository of (parts of) works or other subject matter that the model accesses and ‘collates’ to produce an output.” Rather, the model contains only information extracted from the training dataset, namely individual tokens, and the statistical patterns and relationships connecting them. To be sure, the AG noted, LLMs sometimes “regurgitate” the information they are trained on—but that regurgitation is largely a function of the extent to which statistical relationships have been formed between words, not the replication of those words per se within the database. In any event, the AG emphasized, reproductions of protected works and other subject matter generated by the model as outputs were, in themselves, actionable, so there was no reason to infer merely from those outputs that the model itself contained reproductions of the protected works.

LLM Training—the merits. The AG finally turned to the question of whether LLM training involved an unlawful use of the works. Here the question turned on whether such training fell within an exception under EU law for “text and data mining”—a “seminal issue in EU artistic and literary property law,” the AG noted. At bottom, the AG concluded, “I tend to think that the TDM exception does, in principle, apply to the reproductions of protected works and other subject matter carried out for the purposes of LLM training.

The AG reached his conclusion in various steps. To begin with, he rejected the argument, tendered by some European governments, that the TDM exception applied only to “Big Data”-type analysis for the purpose of scientific and other research—in other words, “the automated analysis of a large corpus of data in order to gain insight into that corpus (which can then be used for various purposes, including predictive or user behaviour analytics).” The problem with this argument, the AG noted, was that it was not consistent with the intent of the EU legislature in adopting the exception. Indeed, he noted, the EU legislature decided to adopt “a broad exception covering text and data mining techniques which are widely used both by private and public entities to analyse large amounts of data for various purposes, including ‘the development of new applications or technologies.’” That was broad enough, the AG concluded, to encompass AI training.

The AG then turned to the three-step test applied to determine whether the TDM exception applied. Under EU law, the AG noted, the exception must be restricted to (i) “certain special cases” which (ii) “do not conflict with a normal exploitation” of the work or other subject matter used and (iii) “do not unreasonably prejudice the legitimate interests” of the rightholders concerned. The “special case” test was easily met, the AG concluded, because the case in question—LLM training—had “sufficiently well-defined contours” and therefore was “narrow from a conceptual...point of view.”

As to the second step of the test, conflict with normal exploitation of the work, application of this test required sensitive considerations. To be sure, the AG allowed, AI-generated works could compete with works made by humans. But copyright law and the related rights “are not designed to protect rightholders against competition per se,” the AG emphasized. Moreover, he emphasized, the mere capability of a trained LLM to generate content that mimics human-made works generally and could compete with them to some extent is, “too diffuse a consequence” to be relevant to this factor. To be sure, it may affect the exploitation of human-made creations “in a more abstract manner”—but that was not the result of the use of specific works in and of themselves but was “the result of training generally.” A different question arose, however, when an LLM generated outputs that closely reproduced particular works in its training data, potentially substituting for those works. Ultimately, the AG recommended, this issue would have to be considered on a case-by-case basis—and developers wishing to benefit from the exception “should demonstrate that they have taken the measures envisaged in the preceding point in order to mitigate the risk of their model generating outputs that resemble specific pieces of their training material too closely.”

Finally, as to the third step, the AG found, exempting the use of protected works and other subject matter for the purposes of LLM training under the TDM exception “does not, as a matter of principle, unreasonably prejudice the legitimate interests of the rightsholders concerned.” Critically, he noted, the TDM exception applies only to the use of works to which developers have lawful access, and only where rightsholders have not expressly reserved their rights against such use. Thus, to the extent that rightsholders used appropriate measures to indicate disapproval of the use of their works for training, he concluded—such as the common “robots.txt” protocol, or references within the terms and conditions of websites—they were able to prevent such use under the exception. Even so, the AG acknowledged that exceptional circumstances might require a different conclusion about unreasonable prejudice in the training of a particular model.

Reproductions during the RAG process. The AG then turned to another, less often evaluated, part of the AI chatbot experience, namely the accessing of protected web content in response to user queries. In this step, known as Retrieval-Augmented Generation, or RAG, Gemini will sometimes conduct a Google search and incorporate information from that search within its response to a user query. Does that search, and the review of resulting web pages, constitute a reproduction of those pages under EU law? The AG found that this required a “nuanced” answer. Importantly, he found, such a RAG process “is not carried out for the purpose of extracting information from a large dataset (such as, in LLM training, the statistical patterns and relationships between words that become the model’s parameters).” Rather, it consists in “retrieving some specific piece of content from the World Wide Web in order to use it directly to ‘augment’ the accuracy of a specific answer.” Ultimately, he concluded, although RAG involves technical reproductions, those reproductions could qualify for the temporary reproduction exception under EU law, provided that chatbot accessed the material lawfully, without circumventing effective technical restrictions, and its ultimate response did not infringe the publisher’s rights. Under those conditions, exempting the technical reproductions would not “unreasonably prejudice the legitimate interests of the press publishers concerned.”

Reproductions during output. Finally, the AG turned to “the visible part of things,” namely the potential reproduction of protected content during output. Here, the AG noted, traditional doctrines of copyright law (as codified in the idea/expression dichotomy in this country) applied. Thus, the publishers’ right extended only to “the specific expression of information related to news in articles contained in a given collection.” In other words, he emphasized, “it grants the publisher exclusivity over the form of expression in those articles (namely the way in which those articles are arranged, structured and phrased), not the facts reported therein.” The right also excluded individual words and “very short extracts” from protection. Importantly, though, where the answer provided to the user by the chatbot “not only reports the same facts but also reflects a distinctive arrangement, structure and general phrasing of an article which forms part of a press publication, the latter could be recognisable in that answer despite some degree of rephrasing.” The decisive consideration, the AG explained, was whether the answer effectively substituted for the original article, allowing users to dispense with reading it and thereby undermining the publisher's investment. That would require a case-by-case analysis, the AG concluded.

MainStory: TopStory Copyright AINews TechnologyInternet GCNNews