Training an artificial intelligence model on copyrighted books, articles and other works should, as a rule, be lawful in the European Union, provided the developer avoids pirate sources, respects paywalls, bot blocks and machine-readable opt-outs, and can show it took steps to stop the model reproducing specific works. That is the provisional view Advocate General Maciej Szpunar set out on November 3, 2026, in an opinion on a question he told the Court of Justice it does not need to answer.

The opinion, in case C-250/25 Like Company v Google Ireland Limited, concerns a Hungarian news website's claim against Google's Gemini chatbot. Szpunar advises the Court to declare the two questions about model training inadmissible, because nothing indicates the Hungarian articles were used to train Gemini and because the right the publisher relies on does not reach training at all. He then answers them anyway, "for the sake of completeness and in order to provide a proper framework for the debate," across some 95 of the opinion's 239 numbered points. "I cannot stress the sensitivity of those questions enough," he writes.

PPC Land reported the opinion's findings on chatbot answers on November 3. The training analysis matters well beyond the Hungarian case, and the opinion explains why. The reproduction right and the text and data mining exception apply to every category of rightholder, from authors to record producers, so whatever the Court says about press publications "could be transposed mutatis mutandis to other subject matter." The use of protected works for training, the opinion says, "is massive, and at the heart of the development of generative AI." The opinion's introduction carries a caveat, though: on the hypothetical matters, Szpunar writes, "I will sometimes express a provisional view only, reserving the right to refine the latter in future cases."

In Short

An adviser to the EU's top court has explained how he thinks copyright law should treat the training of AI models like the ones behind Gemini and ChatGPT. In his view, copying articles and books to train a model is generally allowed in the EU, unless the owner has said no in a way machines can read, the developer got past a paywall or bot block, or the model ends up producing near-copies of specific works. That leaves the opt-out as the main lever for publishers who want to be paid, while AI companies would have to prove they took steps to stop their models memorising content. The opinion is not binding, and the court may not rule on training at all in this case.

Why training falls outside the dispute

Like Company Kft., which runs advertising-funded news portals in Hungary including balatonkornyeke.hu, claims before the Budapest Környéki Törvényszék (Budapest Regional Court) that Google, through Bard and then Gemini, infringed its rights between June 13, 2023 and May 31, 2024. The only right it invokes is the press publishers' right created by Article 15(1) of the 2019 Copyright in the Digital Single Market (CDSM) Directive. Its evidence on training was that Gemini could summarise a July 21, 2023 article about the singer Kozsó's plan to bring dolphins to Lake Balaton, so the model must have "known" it.

Szpunar finds that implausible. Gemini 1.0 was released in December 2023, and "the training of such large LLMs generally takes months and the final dataset is usually collected well before they are deployed." The chatbot more likely fetched the article in real time, since the user named the domain in the query.

The second reason matters more for news publishers. Article 15(1) covers only the "online use" of press publications, which Szpunar defines as "any use of such a publication that occurs as an integral part of the provision of an information society service via the internet, at the individual request of a recipient of that service." Training happens "on dedicated computer infrastructure, without any information society service being provided to a user via the internet," and the link to what a chatbot later tells its users "is simply too indirect (and, usually, diffuse)." The Commission's original proposal would have covered the "digital use" of publications; lawmakers narrowed it to "online use."

The upshot is that the press publishers' right gives news publishers no protection against model training. Copyright in the articles themselves, held by journalists and often transferred to publishers by contract, remains available, which is why the rest of the training analysis concerns copyright and related rights generally.

Four stages and a trail of copies

The opinion divides training into four stages. First, a corpus is collected, often through "automated, large-scale scraping" of the web by bots, sometimes by the developers themselves and sometimes by entities that specialise in compiling datasets. Quality matters as much as quantity: books, scientific journals and newspapers "are considered high-quality sources for training," a footnote says. Another footnote records that, to train Claude, "Anthropic notoriously collected countless hard copies of books, of which it tore and digitalised every page," citing the June 2025 ruling in Bartz v Anthropic.

Second, the dataset is cleaned, converted into a standard format and broken into tokens, each assigned a numerical identifier. Third, during the "training loop," the neural network processes those sequences of numbers, creating and deleting temporary copies in memory, and absorbs statistical relationships between tokens as parameters. Fourth, the model is fine-tuned and deployed.

Each stage produces copies, and Szpunar finds "little doubt" that they count as reproductions under Article 2 of the 2001 InfoSoc Directive. Some experts argue that copying works "as data" to extract patterns is not reproduction in any meaningful sense, comparing it to a person learning to write by reading. The opinion rejects that. EU law took "not a functional, but a purely technical approach," so "the purpose for which such a reproduction is made is immaterial." And the analogy fails on its own terms: "This is not comparable to a human learning to write by reading the works of others, if only because of the industrial scale of the process."

Google relied on the exception for temporary copies in Article 5(1) of the InfoSoc Directive. Szpunar accepts that it can cover the transient by-products of the training loop, but not the dataset. Those copies are stored on servers for months, are created so that training can happen rather than as an incidental part of it, and have value of their own. "Such a dataset has commercial value in itself - which is evidenced by the fact that many entities now specialise in the creation and marketing of such datasets," the opinion says. A footnote adds that "it is not at all clear that AI developers discard them afterwards."

A model is not a database

Several parties argued that a trained model itself contains copies of the works it learned from. Szpunar urges "the utmost caution" on that point.

"A trained LLM is not a database," he writes. Its parameters reflect how words are used "not in a particular text in the dataset, but in the dataset generally," and Google told the hearing that the parameters cannot be reverse-engineered to reveal the works used. The difficulty is regurgitation: models sometimes reproduce passages verbatim when prompted with the opening of a text. The opinion attributes that to "overfitting," which occurs when a text such as the lyrics of a famous song appears many times in the training data. Such memorisation "appears incidental to the machine learning process, and not entirely avoidable."

How often it happens is disputed, though in January 2026 Stanford researchers reported extracting entire copyrighted books from AI models. German courts have drawn a firm conclusion from the phenomenon. On November 11, 2025, the Munich Regional Court ruled in GEMA v OpenAI that GPT contained reproductions of at least nine German songs. According to the opinion, the evidence was outputs generated from prompts that the collecting society's lawyer designed to force the model to recite the lyrics. The same court took the same approach in GEMA v Suno on July 31, 2026.

Szpunar sees it differently. Regurgitation involves randomness, and the user's prompt is itself an essential component of it: "it is the combination of the model's parameters and of the user's prompt that may lead to the model generating reproductions of its training dataset." Seen alone, the parameters are "merely preparatory material." A footnote takes the opposing argument to its limit: it "could mean that, when a human 'memorises' a work by reading it repeatedly, that work is 'reproduced' as electrical patterns in his or her neurons." Among the authorities he cites for the view that parameters are too abstract to be copies is the UK High Court's November 2025 judgment in Getty Images v Stability AI.

The commercial stakes are large. If models contained copies, "the mere offering of such a model would constitute an act of exploitation" of every work used in training, requiring each rightholder's consent and creating "an excessive liability burden for the developer." Should the Court ever take that route, a footnote suggests, the exception for incidental inclusion of a work in other material might offer a way out.

Rightholders keep two remedies under his approach. An output that reproduces a protected work can infringe in its own right, assessed case by case. And the lawfulness of training itself will depend on whether developers took measures against memorisation, a condition the opinion develops later.

Where the training happens

Before any of that can matter, there is a preliminary problem. Gemini's models were trained essentially on computer infrastructure in the United States. Hungarian law governs Like Company's claim, which Google does not dispute, but copyright is territorial: a state's law protects only against acts sufficiently connected to its territory.

Google argued that only the country hosting the main training infrastructure, the "centre of gravity" of training, can regulate it. Szpunar credits the approach with simplicity and legal certainty, but notes its "significant downside": EU law could offer no protection when European works train models abroad. Every government that submitted observations except Germany's took the opposite view, that EU law reaches training anywhere once the model is used in the internal market. Extending EU law that far on the basis of effects alone, he writes, "would raise questions with respect to the principle of territoriality and the legal certainty offered to model developers established in third countries."

Germany proposed a middle course: applying EU copyright to the scraping of works from servers located in the Union, wherever training later takes place. "I would not exclude such an interpretation," Szpunar writes. "Copies of protected works and other subject matter, as data, should not be transferred outside the European Union for training purposes where the rightholders oppose such a use." He acknowledges the weakness. A server's location "may be unknown, particularly in the age of cloud technology," although rightholders could store works on servers in member states and developers "could probably programme their bots to identify a server's location during the scraping process."

He reads the AI Act the same way. Providers of general-purpose models placed on the EU market must have a policy to comply with EU copyright law and respect opt-outs, "regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place." When developers scrape content from servers in Europe, the opinion reasons, EU law may apply and require them to honour opt-outs, even if the training itself happens abroad. The opinion also quotes the explanatory statement of Axel Voss's European Parliament report on copyright and generative AI, dated February 25, 2026: "we cannot allow AI models to be trained just anywhere in the world using European copyright-protected data only for them to be then made available in Europe." Parliament went on to adopt a resolution on copyright and generative AI on March 10, 2026, the same day the Court heard this case. Szpunar's own preference is for the legislature to settle the territorial question "in a clear manner."

Is training "text and data mining"?

The central question is whether training falls under the text and data mining (TDM) exception in Article 4 of the CDSM Directive. The opinion states the stakes plainly: "If those reproductions fall under the TDM exception, the use of protected works and other subject matter for the purposes of such training is, as a rule, freely permitted in the European Union. By contrast, if that exception does not apply, then that use falls under the rightholders' sole purview - and, thus, requires their prior consent and compensation."

"I tend to think that the TDM exception does, in principle, apply," Szpunar writes. Article 2(2) defines text and data mining 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." Google and the Estonian and French governments argued that training fits: a dataset is analysed by an automated technique to generate the statistical patterns that become the model's parameters.

The Danish and Spanish governments countered that text and data mining means "Big Data" analysis, a different activity the legislature had in mind in 2019. The opinion's account of that argument says GPT-1 was released "a few months after the adoption of the CDSM Directive." OpenAI in fact published GPT-1 in June 2018, about ten months before the directive was adopted on April 17, 2019.

Szpunar accepts that training differs from Big Data analysis but finds it "not decisive." The definition does not limit what kind of information may be extracted or what it is later used for. The legislative process, he notes, turned a narrow exception proposed for scientific research into an additional broad one for private and public entities, including for "the development of new applications or technologies." A 2018 study for the European Parliament's legal affairs committee already observed that "although classical TDM and machine learning have different utility, it should not be overlooked that both use the same key algorithms to discover patterns in data." Above all, Google and the German, Estonian, French and Hungarian governments pointed to recital 105 and Article 53(1)(c) of the AI Act, which in his reading show the legislature considers the exception to cover training.

If the exception applies, it covers scraping, storage of the dataset and the other copies made during training. The directive allows copies to be kept "for as long as is necessary" for mining, which the opinion reads as "pending and during LLM training." A footnote adds that this would not allow developers to keep a dataset longer than training requires, nor a third-party scraper to keep copies "forever and for other purposes." The exception cannot cover a model's outputs.

The three-step test

Greece and Spain argued that exempting training would breach the "three-step test" in the Berne Convention and the WTO's TRIPS agreement, under which exceptions must be confined to certain special cases, must not conflict with a normal exploitation of the work and must not unreasonably prejudice rightholders' legitimate interests. Szpunar sides with Germany, Estonia, France and Hungary that the test does not, in principle, rule the reading out. But his reasoning attaches a condition that developers may find demanding.

On the first step, he rejects a purely quantitative approach. Private copying and temporary copies are mass uses that still count as special cases; what matters is that the use is narrowly defined, here the use of works "as data" to extract information.

The second step is where, in his words, "the debate concerning LLM training is the most intense." He splits the objection in two. On the input side, he accepts that a market for training licences "already exists" and that free use may shrink it. But the argument that a use could otherwise be licensed "cannot be decisive in itself. Otherwise, it would justify doing away with any exception or limitation to copyright and the related rights, no matter how important the public interest pursued by the EU legislature thereunder." The exception's own conditions, lawful access and the opt-out, preserve the ability to license.

On the output side, the fear that AI-generated content will dilute and displace demand for human work is, in general, "too diffuse a consequence." A model trained on open data, or on a licensed catalogue, would have the same capability. And "copyright law and the related rights are not designed to protect rightholders against competition per se." Reading the second step broadly would also strip the legislature of the option of creating a statutory licence or remuneration scheme. "Such a radical outcome surely cannot be endorsed lightly," the opinion says.

Model-specific substitution is treated differently. Where a model "could routinely generate content that resembles specific pieces of its training dataset to such an extent that the latter is 'recognisable' in the former," developers can be expected to take measures such as "proper dataset diversification and deduplication, implementation of prompt and output filters, and so on." Because the three-step test must be checked whenever the exception is applied, that becomes a burden of proof. "Developers wishing to benefit from that 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," the opinion says. "Failing that, the national court adjudicating the case should refuse the application of the TDM exception."

He links that duty to the AI Act's requirement that providers maintain a copyright compliance policy, and to Measure 1.4 of the General-Purpose AI Code of Practice, under which signatories commit to technical safeguards against infringing outputs. Google said in July 2025 it would sign the code; Meta said it would not.

The third step weighs harm against benefit. On harm, the opinion is direct about which kinds of writing are exposed: "LLM-generated content will not displace the demand for human-made, highly original novels and other masterpieces of literature. Nor will it displace demand for in-depth, analytic pieces of journalism. However, it could easily replace human creators as regards texts with low originality, such as purely informative news or sports reports, romance novels (and so on)." A footnote observes that press publishers themselves increasingly use generative AI to draft such reports. The opinion also warns of the web being flooded with AI-generated "slop" and of "widespread public distrust of, rather than curiosity in, unknown content."

On benefit, it cites innovation, the transaction costs of licensing every work individually and, more unusually, industrial policy. "Digital sovereignty should not be forgotten in that equation," Szpunar writes. The legislature "could take the view that the new policy of limiting the European Union's dependence on foreign technology involves the development of European AI models, and that the latter should be fostered, rather than hindered." Training on European works "also ensures the representation of European culture in AI." His conclusion is that the legislature "was accordingly at liberty to tip the scales more in favour of innovation."

He leaves room for exceptions decided case by case. A footnote gives one example: exempting the use of works "to train a model to generate racist content" could unreasonably prejudice rightholders.

Lawful access: no pirate libraries, no hacking

The first condition in Article 4 is that works be "lawfully accessible." For content on the web, Szpunar reads that as access to lawful sources only, meaning pages where the rightholder made the content available or consented to it. A broader reading in recital 14 of the directive, he argues, belongs to the separate scientific research exception.

Where rightholders use effective technical measures, "such as a paywall or a measure designed to prevent access by robots specifically," content obtained by circumventing them is not lawfully accessible. A footnote notes that "various providers offer 'bot-management' services" for this purpose, and another rejects any reading of the InfoSoc Directive that would let exception beneficiaries break through: Article 6(4) "does not grant model developers a right to hack."

He also recognises a practical limit. Bots cannot tell whether a page was posted by the rightholder, and too strict a reading "could make it impossible for developers to have recourse to such web scraping." The balance he proposes is a duty "not to scrape obviously unlawful sources (such as pirate websites and shadow databases) and not to circumvent such measures." Signatories of the Code of Practice already commit to excluding sites recognised by EU courts or authorities as persistent, commercial-scale infringers. The stakes of getting this wrong are visible in the United States, where Anthropic, after the ruling that its pirated library was not covered by fair use, agreed to pay at least $1.5 billion to settle the authors' claims in September 2025. Paywalls are a live issue too: Common Crawl, whose archive feeds AI companies, has supplied paywalled content despite publishers' objections.

Opting out: robots.txt, Google-Extended and plain language

The second condition gives rightholders their main lever. Under Article 4(3), the exception applies only if use "has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online." A reservation of rights lets a rightholder license training "in exchange for appropriate remuneration," the opinion says.

Critics cited in the opinion object that the burden falls on rightholders, who may need to opt out on every site hosting their work in the absence of a central EU register, that the rules are uncertain, and that developers have been shown to ignore opt-outs. Szpunar answers that the legislature "may well require that rightholders show a certain degree of diligence with respect to the management of their online content," and rejects the argument that an opt-out is a formality banned by the Berne Convention.

His test for a valid reservation is functional. It qualifies "where state-of-the-art robots and associated technologies are capable of 'understanding' it with a sufficient degree of reliability at the moment of access."

The robots.txt protocol meets that test, though the opinion lists its weaknesses. It "does not technically prevent scraping," relying on good faith, and "some bots have been shown to ignore robots.txt directives." It requires publishers to know crawlers' names, which some developers do not disclose. And it "may not easily be used to distinguish between robots and the purposes for which crawling is carried out (model training, RAG, or web search indexing)." Cloudflare has documented undeclared crawlers it attributed to Perplexity, which denied training AI models. The AI Act helps, Szpunar argues, because Code of Practice signatories commit to crawlers that follow robots.txt and to disclosing the crawlers they use. Data protection regulators have moved in a similar direction: in October 2026, the European Data Protection Board told AI developers to avoid sites that block scraping with robots.txt.

Google-Extended, which Google released on September 28, 2023, rests on robots.txt. According to the opinion, it lets publishers stop Google's bots from scraping their content to train Google's models and to ground Gemini's answers, and it "allegedly does not affect Google Search." In the part of the opinion dealing with chatbots, Szpunar notes that Google's search indexing and chatbot grounding appear to use the same bots, and records Google's statement at the hearing that blocking access through Google-Extended for training and grounding would not affect a site's visibility on Google. "That would need to be verified," he writes.

Other methods also count: other protocols, metadata embedded in files and, as recital 18 of the directive envisages, a website's terms and conditions. On that last point, he does "not exclude" that a reservation written in natural language could qualify, "which does not seem that far-fetched given the current natural language processing capabilities of LLMs." He would prefer a small number of standard formats and expects the processes under way to produce them "will, in all probability, succeed." The opinion does not mention that on October 30, 2026, Google, Microsoft, OpenAI and Anthropic said they would honour a single robots.txt token for AI training, drawn from an IETF Internet-Draft that has no formal standing yet. Google said it would treat that training opt-out as disallowing Google-Extended. The Commission's review of the Copyright Directive, meanwhile, is at the consultation stage.

Rightholders' difficulty in checking whether opt-outs are respected is addressed through transparency. Under Article 53(1)(d) of the AI Act, providers of general-purpose models must publish "a sufficiently detailed summary about the content used for training," following a template from the AI Office, "irrespective of their location." Szpunar calls that obligation "a crucial complement to the TDM exception." Not everyone shares his confidence in it: in August 2025, a coalition of European creators' organisations rejected the training data disclosure template along with the Code of Practice.

Compensation is left to member states

No EU-wide scheme pays creators for training under the exception. Szpunar suggests that "the significant effects that generative AI may have on the demand for human creations could potentially justify putting in place a form of statutory remuneration," similar to private copying levies, and sees "nothing" to stop member states creating one now.

Recital 17 of the directive, which says member states should not provide compensation for uses under the TDM exceptions, does not stand in the way in his reading. Its reasoning concerns the scientific research exception, where harm is "minimal," while the harm caused by training under Article 4 "may be considerable."

What it means for publishers and AI developers

For news publishers, the opinion's message on training is mixed. The right created for them in 2019 does not apply. Journalists' copyright does, and so does the exception, which makes training on their open content lawful by default unless they reserve their rights in a form machines can read. A paywall or bot block also takes content out of the exception's reach. For an advertising-funded site with no paywall, such as Like Company's, the opt-out is the main tool.

For developers, the opinion is broadly permissive, and on whether a model contains copies it sits closer to the UK judgment for Stability AI than to the Munich rulings. But it attaches duties. They must avoid pirate libraries, honour access controls and opt-outs, keep datasets only as long as training requires and, if the issue reaches a court, be able to show what they did to prevent memorisation. In the United States, judges reached a broadly similar split in 2025: Bartz v Anthropic found training on books to be fair use but rejected the defence for a pirated library, and Meta won summary judgment days later in Kadrey v Meta, a case Szpunar cites on the market effects of AI-generated content.

The opinion's acknowledgement that a licensing market "already exists" matches what publishers have disclosed. The New York Times has a licensing agreement with Amazon, USA Today Co. has joined Microsoft's Publisher Content Marketplace, and Google's AI contribution pilot pays about 100 publishers, some under $1,000, according to figures reported by The Information. Under Szpunar's reading, the opt-out is what gives such negotiations their footing in the EU: without one, training on open content needs no licence at all.

Other developments PPC Land has tracked bear on the same mechanics. The European Commission has had a formal antitrust investigation open since December 9, 2025 into Google's use of publishers' content for AI Overviews and AI Mode and of YouTube content for training AI models. In June 2026, Google added a Search Console control that removes a site from AI Overviews and AI Mode; Google's documentation for it says it does not affect AI training, which remains governed by Google-Extended. Whether blocking Google-Extended leaves search visibility untouched is exactly the claim the opinion says "would need to be verified."

What happens next

The opinion binds no one. The judges are now deliberating and will deliver judgment at a later date. If they follow Szpunar's advice on admissibility, the judgment may say nothing about training, leaving his analysis as the most detailed statement on the subject from within the Court but without legal force. The Budapest court will then apply whatever the Court rules on the chatbot questions.

Szpunar himself frames the training discussion as an attempt "to draw attention to their complexity, and why the Court should proceed with caution on the matter," rather than a final answer. On the hardest question, where training takes place, he passes the decision back to lawmakers: "Some refined version of the solution suggested by the German Government, or indeed an entirely new solution, which complies with the limits of the principle of territoriality, could be put forward in a future instrument."

Timeline

Summary

Who: Advocate General Maciej Szpunar of the Court of Justice of the European Union, in a dispute between Like Company Kft., a Hungarian operator of advertising-funded news portals, and Google Ireland Limited, provider of the Gemini chatbot. Google, the European Commission and the governments of Hungary, Denmark, Germany, Estonia, Greece, Spain, France, Italy and Poland submitted observations. AI developers, publishers, authors and other rightholders across the EU are affected by the reasoning.

What: A non-binding opinion advising the Court to declare the questions on AI model training inadmissible, followed by a provisional analysis of them. Szpunar finds that training involves reproductions, that a trained model does not itself contain copies, and that the text and data mining exception covers training in principle, on three conditions: works must be lawfully accessible, rightholders must not have opted out in a machine-readable way, and developers must show they took measures against memorisation. He also finds that the press publishers' right does not cover training, that EU law may reach the scraping of works from EU servers for training abroad, and that member states may introduce compensation for creators.

When: The opinion was delivered on November 3, 2026. The hearing took place on March 10, 2026. Judgment will follow at a date not yet set.

Where: The Court of Justice in Luxembourg, on a reference from the Budapest Környéki Törvényszék (Budapest Regional Court) in Hungary. Gemini's models were trained essentially in the United States.

Why: Whether training on protected works is free under the text and data mining exception or requires consent and payment is one of the most contested questions in EU copyright law, with lawsuits pending across Europe and the United States. This is the first generative AI case to reach the Court of Justice, and the Advocate General's reading, even if the Court does not adopt it, sets out how the exception, the AI Act and opt-out tools such as robots.txt and Google-Extended fit together.