An adviser to the European Union's highest court today said that an AI chatbot which pulls a news article from a publisher's website and reuses it to answer a user can infringe the publisher's rights - but only if the chatbot reached the article by getting around a paywall or a block on bots, or if its answer copies the article's way of telling the story so closely that a reader no longer needs the original.

Advocate General Maciej Szpunar delivered his opinion in case C-250/25, Like Company v Google Ireland Limited, a dispute that began in a Hungarian regional court over a news site, a singer and a plan to bring dolphins to Lake Balaton. According to the Court of Justice's press release No 155/26, it is "the first generative AI case to reach the Court of Justice." Szpunar acknowledges in his introduction that the opinion "significantly exceeds the customary length of opinions of Advocates General." It runs to 240 numbered points and 243 footnotes.

Its conclusions cut both ways. For publishers, it rejects Google's argument that freely accessible articles are implicitly licensed for chatbot reuse, and it says paraphrase does not, by itself, put an answer beyond the law. For Google, it calls the training questions hypothetical, says training on protected works is as a rule permitted under the EU's text and data mining exception, and finds that the one example the publisher produced does not appear to meet the test for infringement.

In Short

A Hungarian website owner sued Google, saying its Gemini chatbot copied the site's news articles, and a Hungarian court asked the EU's top court how the rules apply. Today, an adviser to that court said chatbots are free to repeat the facts in news stories, but they cross a line if they sneak past a paywall or rewrite an article so closely that you no longer need to read it. He also said the Hungarian site's own example probably does not cross that line, and that training AI on news is generally allowed unless publishers opt out. The judges still have to decide, so this is a strong signal rather than the final word.

A lake, a singer and a pool that did not exist

Like Company Kft., incorporated in Hungary, operates several web portals. One of them, balatonkornyeke.hu, publishes freely accessible articles about Lake Balaton. According to the opinion, the company "draws revenue through advertisements displayed on the portals in question." No paywall, then: a business that lives on traffic.

Before the Budapest Környéki Törvényszék (Budapest Regional Court), Like Company claims that Google, first through Bard and then through its renamed chatbot Gemini, continuously infringed its rights between 13 June 2023 and 31 May 2024. The right in question is not copyright. It is the press publishers' right created by Article 15(1) of Directive 2019/790, the Copyright in the Digital Single Market (CDSM) Directive, which gives publishers established in a member state the reproduction and making-available rights over "the online use of their press publications by information society service providers." Like Company confirmed at the hearing that it does not claim copyright in the articles themselves, according to the opinion.

The company's illustration is specific. On 21 July 2023, balatonkornyeke.hu published an article reporting that Kozsó, a Hungarian singer, wanted to introduce freshwater dolphins into an aquarium at Lake Balaton, a dream he said he had cherished for more than 20 years.

At some later date, an unknown user asked Gemini: "Can you provide a summary in Hungarian of the online press publication that appeared on balatonkornyeke.hu regarding Kozsó's plan to introduce dolphins into the lake?" Gemini answered. Its summary described a special pool to be built on the Tihany peninsula, 25 metres long, 15 metres wide and 5 metres deep, holding 1,000 cubic metres of water. It said some critics feared the dolphins would not adapt and that the construction would harm the environment.

None of that was in the article. According to the opinion, the summary "contained a series of statements finding no basis in that publication (the fact that a pool is going to be used to introduce the dolphins, the size of that pool, the reception by the public of that project, and so on) which, according to Google, are 'hallucinations' of the LLM." Szpunar calls the summary "rather approximative."

Like Company argued that the answer proved Gemini's models "knew" the article from training, and that the chatbot had reproduced it and made it available to the public.

The Hungarian court stayed the proceedings and sent four questions to Luxembourg: two on whether training a large language model is a reproduction and whether the text and data mining (TDM) exception in Article 4 of the CDSM Directive covers it, and two on whether the chatbot's answers reproduce publications or communicate them to the public.

Written observations came from Like Company, Google, the European Commission and the governments of Hungary, Denmark, Germany, Estonia, Greece, Spain, France, Italy and Poland. All but the Estonian, Italian and Polish governments attended the hearing on 10 March 2026.

Training questions: inadmissible, then answered anyway

The Commission asked the Court to throw out the whole reference as too vague about the technology. Szpunar declined: the Court "should instead guide the referring court, which is evidently lost in the AI landscape."

On training, he advises the Court to declare the second and third questions inadmissible, for two reasons.

The first is timing. The article appeared on 21 July 2023. Gemini 1.0 was released in December 2023. According to the opinion, "the training of such large LLMs generally takes months and the final dataset is usually collected well before they are deployed," which makes it "implausible" that the article was used to train the models. The likelier explanation is retrieval: the user named the domain and gave enough context for the chatbot to fetch the exact article in real time. "In sum, the case in the main proceedings has nothing to do with LLM training," the opinion says.

The second reason is more consequential for every European publisher. The press publishers' right covers only "online use." Szpunar defines that 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 a model happens on dedicated computing infrastructure, sometimes not even connected to the internet, with no service provided to anyone. The link between that upstream use and what a chatbot later tells its users "is simply too indirect (and, usually, diffuse)." The upshot is that the press publishers' right offers no protection against model training at all. Copyright held by journalists, often transferred to publishers by contract, remains available for that purpose.

Having set the questions aside, Szpunar then spends more than 90 points answering them "for the sake of completeness." "I cannot stress the sensitivity of those questions enough," he writes.

Where training happens

Gemini's models were trained essentially in the United States. Google argued that only the country hosting the main computing infrastructure, the "centre of gravity" of training, can regulate it, which Szpunar notes would leave EU rightholders without protection under EU law when their works train models abroad. Every government except Germany's argued that EU law reaches training wherever it happens once the model is offered in the EU, an approach he says would raise questions about territoriality and legal certainty. Germany proposed applying EU copyright to scraping from servers located in the Union. Szpunar "would not exclude" that, writing that "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 would prefer the legislature to settle the question.

Copies in, not copies inside

On substance, he finds "little doubt" that training involves reproductions within the meaning of Article 2 of the InfoSoc Directive: the scraping that builds the dataset, the conversion and tokenisation of texts, and the temporary copies created during the training loop. He rejects the argument that a model learns like a human reader. "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," the opinion says.

The trained model is another matter. "A trained LLM is not a database," Szpunar writes. He urges "the utmost caution" on the theory that models contain copies of works because they sometimes reproduce them when prompted, the phenomenon known as regurgitation. That theory was accepted by the Munich Regional Court in GEMA v OpenAI on 11 November 2025, in a case involving at least nine German songs, and again in GEMA v Suno on 31 July 2026. Szpunar instead cites the UK High Court's November 2025 ruling for Stability AI among the authorities finding that a model's parameters are too abstract to be copies.

An output that reproduces a work can itself infringe, case by case. But "one should not infer from the fact that an LLM may generate reproductions of works on which it was trained when prompted appropriately that such reproductions are stored inside the model." Treating models as containers of copies, he argues, would make "the mere offering of such a model" an act requiring every rightholder's consent.

The text and data mining exception

The pivotal question is whether training falls under the TDM exception in Article 4(1) of the CDSM Directive. If it does, training is as a rule free in the EU; if not, it requires consent and payment. "I tend to think that the TDM exception does, in principle, apply," Szpunar writes. The definition in Article 2(2), "any automated analytical technique aimed at analysing text and data in digital form in order to generate information," is broad enough to cover machine learning, a position taken by Google and the Estonian and French governments.

The Danish and Spanish governments argued that text and data mining means "Big Data" analysis, not model training. Szpunar finds the difference not decisive, relying above all on recital 105 and Article 53(1)(c) of the AI Act, which in his reading show the legislature considers the exception to cover training.

Greece and Spain went further, arguing that exempting training breaches the "three-step test" in international copyright treaties. Szpunar disagrees, but with conditions that carry weight. On the market for training licences, which he notes "already exists," he says the exception preserves rightholders' ability to license because of its two built-in conditions. On the broader fear that AI output will flood markets, he finds the general capability of a model "too diffuse a consequence," since copyright "does not protect an author against the creation of new content that competes with his or her works."

Model-specific substitution is different. Where a given model could routinely generate content resembling specific training material, developers "should demonstrate" that they took measures such as dataset diversification, deduplication and prompt and output filters. "Failing that, the national court adjudicating the case should refuse the application of the TDM exception," the opinion says.

The two conditions define how much control rightholders keep. First, works must be "lawfully accessible." Szpunar reads that as a duty "not to scrape obviously unlawful sources (such as pirate websites and shadow databases)" and not to circumvent effective technical measures such as paywalls or measures designed to keep out bots. Second, rightholders can opt out under Article 4(3) through a machine-readable reservation of rights. A reservation qualifies, in his view, 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 qualifies. So does Google-Extended, which Google released on 28 September 2023 and which rests on that protocol, as does metadata embedded in files. He does "not exclude" that a natural-language reservation in a website's terms and conditions could also qualify, given what language models can now read.

He also says member states are free to introduce compensation for creators whose works are used under the exception, despite recital 17 of the directive, which he reads as applying only to the separate scientific-research exception. No such scheme exists at EU level.

His assessment of who is exposed is blunt. AI-generated text "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."

What the chatbot does with a news article

The core of the opinion concerns what Szpunar calls "the real crux of the present case": the use of a chatbot to obtain news. Questions one and four, he advises, are admissible and must be answered.

When asked for information, Gemini uses retrieval-augmented generation, also called grounding. Through a Google Search application programming interface, it browses the Google Search index to find relevant pages, uses internet bots to retrieve the content, and adds that content to the prompt fed to the underlying model, which then generates its answer. The answer can include hyperlinks to the sources, though "such hyperlinks are sometimes, but not always, provided in the answers provided by the Gemini chatbot." Why, the opinion says, "is not entirely clear."

The same technology, Szpunar notes, powers AI Overviews and AI Mode, two services built into Google Search and run on Gemini models. Retrieval, he writes, gives a chatbot capabilities similar to a search engine, except that "where a traditional search engine would merely provide the user with a list of hyperlinks," the chatbot can "collect, piece together, sum up and provide that information directly, in full sentences."

Publishers' grievances are summarised in two parts: providers do not seek consent or pay before scraping articles to ground answers, and the AI-generated content "diverts internet users away from the press publishers' websites, resulting in a considerable loss of traffic and, by extension, of advertising or subscription revenue." Szpunar places this inside a larger shift. "The original business model of the Web rested on a delicate quid pro quo between search engines, such as Google, and website publishers," he writes. Search engines copied websites but sent back traffic publishers could monetise. That arrangement "is called into question" when chatbots and services such as AI Overviews provide information directly.

A footnote records the evidence dispute. Some studies, the opinion notes, indicate that most Google searches are now "zero-click," in particular because of AI Overviews. Google denies this, and the opinion quotes a May 2024 Google blog post saying that "the links included in AI Overviews get more clicks than if the page had appeared as a traditional web listing for that query." PPC Land has reported independent data: a randomised experiment by researchers at the Indian School of Business and Carnegie Mellon found outbound organic clicks fell 39.8% when AI Overviews were shown, and NewzDash data showed Web Search falling from 51.10% to 27.42% of Google mobile traffic to more than 400 news publishers between 2023 and the fourth quarter of 2025.

Two routes to a relevant reproduction

Google and the Commission argued that questions one and four were also hypothetical, since Gemini's answer was not worded like the article, making the absence of infringement obvious. Szpunar rejects that. The matter is "far from acte clair," he writes, not least because the Court has never clarified the scope of the press publishers' right. Its only ruling on Article 15(1), Meta Platforms Ireland (C-797/23) of 12 May 2026, addressed the exclusive nature of the right, not when it is infringed.

Jurisdiction is not a problem here: unlike training abroad, a chatbot offered to EU users produces "immediate and substantial effects" there, wherever the provider or its servers are located, a point the opinion supports with the Court's Anne Frank Fonds judgment of 9 July 2026 on communications received in member states. Retrieval also qualifies as "online use."

On evidence, Szpunar proposes a presumption that favours rightholders. Because chatbot providers have no transparency obligation about what they scrape, courts should presume a publication was used where it was available online and the claimant produces "a reasonable sample" of answers that resemble it or link back to it. The burden would then shift to the AI provider. A single example, he adds in a footnote, "will rarely be enough." He also suggests transparency obligations for chatbot providers on what they scrape.

The invisible copies

Retrieval copies extracts of an article and creates further digital copies as the system processes them. "Google admitted as much at the hearing," the opinion says. The TDM exception does not cover these copies, because retrieval uses a specific piece of content to answer a specific question rather than extracting patterns from a dataset. The temporary reproduction exception in Article 5(1) of the InfoSoc Directive can, since the copies are transient and have no economic value separate from the answer. But it applies only if they serve a "lawful use," which depends on two cumulative factors: how the chatbot accessed the article, and what form the final answer takes.

Access must be lawful. Where a publisher uses effective technical measures "be it by way of a paywall or a measure designed specifically to prevent access and/or scraping by internet bots," the chatbot must not have circumvented them. A chatbot that does so not only infringes, the opinion adds, but commits a separate tort under Article 6(1) of the InfoSoc Directive.

Facts are free, phrasing is not

The visible answer is assessed under the Court's ordinary infringement test, with no special rule for AI, which Szpunar says would breach the principle of technological neutrality. Two conditions apply: the publication must have been copied, and its protected elements must be "recognisable" in the result on a side-by-side comparison.

Copying is satisfied by retrieval. Google argued that because the article is fed into a prompt rather than pasted into the answer, nothing is copied. Szpunar disagrees: "there is, in such a scenario, a sufficient causal link between that publication and the answer provided."

Recognisability decides most claims. The press publishers' right, he writes, protects "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 court's press release puts it plainly: the right's "aim is to protect the investment that goes into making such a publication, not to give publishers a monopoly over the news itself."

The decisive question is substitution: whether the answer "constitutes a direct substitute for the publication, because a user, by reading that answer, can dispense with accessing the original article entirely." Verbatim copying makes that obvious, Szpunar notes, but it "rarely happens nowadays," since models rephrase what they retrieve. Google, the opinion records, is said to apply output filters that screen out passages matching the retrieved text beyond a certain length.

Paraphrase is not a safe harbour. If it were, "the press publishers' right would be too easy to circumvent, in an age where LLMs may easily paraphrase rather than replicate," a position argued by Like Company and by the French, Italian, Danish, Spanish, Polish, Hungarian and Greek governments.

Szpunar then draws a distinction with commercial consequences for different kinds of journalism. Where an answer "not only reports the same facts but also reflects a distinctive arrangement, structure and general phrasing" of an article, it may reproduce it despite rephrasing. Where the answer is "a short summary of the factual content presented differently," or a general outline built from several sources and restructured by the model, there is no reproduction. In that case, competition and lost visits do not matter legally: "the press publishers' right was never meant to protect those publishers against competition per se."

The more distinctive an article, the more recognisable it is. "Articles consisting in long, in-depth analysis are, by their nature, more recognisable than purely informative reports of events." In a purely informative report, information and expression are "indissociable," so one such article is almost indistinguishable from another. Szpunar accepts there will be a grey area, to be resolved case by case.

Making available to the public

For the separate right of making available, the bar is higher. A chatbot makes a publication available to the public only if all or part of it is "systematically perceptible (in other words, 'recognisable') to the point of direct substitution" in answers to related queries. He considers that "rather unlikely" given that chatbot answers are non-deterministic and vary with the model, the query and the conversation history, but possible "as a result of an intentional or negligent design of the RAG process." Only ordinary queries count, not "convoluted and 'abnormal' prompts designed to trick the chatbot."

Who carries the liability

Because the press publishers' right only applies to use "by information society service providers," the question of who acts is decisive. Private, non-commercial use by individuals is excluded altogether.

Szpunar accepts Google's argument that when a user directs a chatbot to do something, "primary" liability generally lies with that user. Holding providers strictly liable for everything users request would have "a dissuasive effect on the development and use of LLMs and AI chatbots."

But providers are not off the hook. The acts are attributable to the provider in two cases: where they are "inherent to the business model of the provider," or where the provider has not put in place the measures "that can be expected from a reasonably diligent operator." As an example of the first, a footnote describes "a chatbot service specifically designed to provide extensive and detailed presentations of individual news articles for its users, merely paraphrased," which would be functionally no different from the news aggregators the right was created to address.

On the second, the opinion leaves little ambiguity. "When one designs and has control over a technology as advanced as an LLM-powered online chatbot, one must generally assume the risks inherent to that technology," it says. Providers should program their retrieval bots to respect publishers' access restrictions and tailor the process so answers are "sufficiently removed from any specific publication retrieved (through output filters, and so on)." Where they do, occasional slips should not create liability. Where they do not, liability is warranted.

Such duties may also arise under the Digital Services Act, a footnote adds, where a chatbot is designated a Very Large Online Search Engine and must mitigate systemic risks such as illegal content. The Commission gave ChatGPT that designation on 31 August 2026, on a declared 159.1 million average monthly users in the EU.

Google's broadest argument was that a publisher leaving its articles freely accessible, with no paywall and no measures against scraping, as Like Company appears to have done, should be treated as implicitly consenting to their use in chatbot answers. "I do not think that is the case," Szpunar writes.

The Court accepted implied consent for hyperlinks in VG Bild-Kunst in 2021. But implied consent requires, at minimum, that rightholders were "actually and individually" informed of the use and had simple means to prevent it. Neither is true here, he argues. Some publishers may not know their work is used, and how publications are selected "is still not entirely clear."

On paywalls, the opinion says: "Publishers may well choose not to put in place such a measure because they prefer collecting advertising revenue from traffic more broadly, not because they implicitly consent to their publications being reused by third parties."

On bot-blocking, it goes further. The bots that ground chatbot answers "are not fundamentally different to those accessing those websites for the purposes of indexing in search engines. In the case of Google, it seems that the same bots serve both functions." As Like Company explained at the hearing, blocking them through robots.txt may affect how a site appears in search. A publisher may accept crawling for search to stay visible, but "it cannot reasonably follow that he or she implicitly consents to the use of those publications to ground chatbots."

Google told the court that Google-Extended allows that differentiation, and that blocking access for training and grounding through it would not affect a site's visibility on Google. "That would need to be verified," the opinion says. Even if true, findings of implicit consent "seem far-fetched" at "the current, still-early stages of AI and chatbots."

The record before the court predates two developments PPC Land has covered since. In June 2026, Google introduced a Search Console control on the day the UK Competition and Markets Authority imposed a conduct requirement on it; the setting excludes a site from AI Overviews and AI Mode and stops its content being used to ground them, without affecting conventional ranking. And on 30 October 2026, Google, Microsoft, OpenAI and Anthropic said they would honour a single robots.txt token for AI training, drawn from an IETF Internet-Draft. The opinion mentions neither.

A middle ground, with a warning attached

Szpunar presents his reading as a balance between publishers' property rights under the Charter of Fundamental Rights and chatbot providers' freedom to conduct a business. "Not all chatbots services are equivalent and to be regulated the same way," he writes.

Chatbots that retrieve content lawfully and reuse it "in an 'unrecognisable' way" should operate freely. Answers that synthesise several sources, link back and urge users to read further, he argues, may even "foster a relationship of synergy, rather than substitution." A footnote suggests the pressure could push publishers toward articles that "contain a real added value for the reader."

Services that "systematically circumvent access restrictions" or recognisably paraphrase protected content should not escape the law. Here the opinion's language sharpens: "There is a real danger that, without adequate control by, and compensation for, rightholders for such intrusive uses of their protected content, the internet becomes entirely controlled and exploited by AI-enhanced intermediaries, and internet users become accustomed, slowly but surely, to consuming slavish derivatives rather than originals."

Szpunar believes his reading "leaves those publishers in a strong bargaining position vis-à-vis chatbot providers," since chatbots need news content to answer news questions accurately, and he expects it to "incite chatbot providers to enter into adequate licensing agreements with press publishers." Some of that money already flows. Google's AI contribution pilot pays about 100 publishers whose content is used in AI Overviews, AI Mode and the Gemini app, with some smaller sites receiving under $1,000 over several months, according to figures reported by The Information.

What it means for Like Company

The guidance for the Hungarian court is short and, subject to that court's findings, unfavourable to the claimant. First, Like Company does not appear to have put in place any technical measures restricting access to or scraping of its articles. Second, the form of expression of the dolphin article is not apparent in Gemini's answer. A side-by-side comparison shows "essentially" the same facts in different terms, with a single sentence from the article, its first, appearing in the answer. Third, any acts of reproduction or making available through Gemini do not appear inherent to Google's business model. Fourth, "Google seems to have put in place appropriate measures to mitigate the risk of such acts occurring."

The article itself, Szpunar notes, appears to be "a purely informative report of events," using "commonplace arrangement, structure and sentences, much of which appears, in addition, to be borrowed from other publications." He also flags, without resolving it, whether portals like Like Company's qualify as press publications at all, observing that the company and its sites "seem far removed from the publishers of 'traditional' newspapers that the legislature had in mind." That is for the national court to decide.

Why this matters beyond Budapest

For publishers that depend on advertising, the opinion sets out where the leverage lies. Content behind a paywall or a working bot block is protected against retrieval that circumvents those measures. Content left open, the model Like Company relies on to sell advertising, is protected only against answers that recognisably reproduce it, and the protection is strongest for distinctive, analytical journalism, weakest for short factual reports.

That aligns closely with the commercial choices PPC Land has tracked. Google has been under a formal European Commission antitrust investigation since 9 December 2025 into whether it uses publishers' content for AI Overviews and AI Mode without compensation or a way to opt out without losing Search visibility, the same coupling of search and grounding crawls that Szpunar identifies. On 29 October 2026, five French publisher groups lodged a further complaint with the Commission.

Courts elsewhere are also separating training from answering. US judges found training on books to be fair use in Bartz v Anthropic in June 2025, and Meta won a similar ruling days later. Newer cases target retrieval: CNN's May 2026 complaint against Perplexity is built on copying at indexing and output, and on 29 October the Wikimedia Foundation asked a federal judge to find ChatGPT's web retrieval likely not fair use where it ignores licence terms. Szpunar cites Perplexity's multiple US cases and a suit brought in Paris by the Alliance de la presse d'information générale and 53 publishers against Brave as examples of the same grievance.

For advertisers, the outcome bears on where news audiences, and the inventory around them, end up. A ruling that tracks this opinion would leave chatbot providers free to summarise facts from open websites, while making systematic paraphrase of individual articles and any circumvention of access controls legally risky. It would also leave much of the evidentiary work to publishers, who would need samples of recurring, recognisable answers.

What happens next

The opinion is not binding. According to the court's press release, the role of Advocates General is to propose "in complete independence, a legal solution" to the cases assigned to them, and "the Judges of the Court are now beginning their deliberations in this case. Judgment will be given at a later date." The Court of Justice does not decide the Hungarian dispute itself; the Budapest court will apply the eventual ruling, which will bind other national courts facing similar questions.

The AI Act does not apply to this dispute, which concerns conduct between June 2023 and May 2024, according to the opinion. Szpunar still leans on it and on the General-Purpose AI Code of Practice published in July 2025, whose signatories commit to crawlers that follow robots.txt, not to circumvent effective technical measures, and to safeguards against infringing outputs.

Szpunar's own conclusion is modest. "The press publishers' right (or copyright more generally) will obviously not provide a solution to all the issues arising from the use of online chatbot services," he writes. "Nevertheless, I believe the interpretation suggested in this Opinion would be a step in the right direction."

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 web portals including balatonkornyeke.hu, and Google Ireland Limited, which provides the Gemini chatbot in the EU. Nine governments and the European Commission submitted observations. Press publishers, AI chatbot providers and the advertisers whose spending follows both are affected.

What: A non-binding opinion advising the Court to declare inadmissible the questions on AI training, while stating that training is in principle covered by the EU text and data mining exception, subject to lawful access, opt-outs and safeguards. On chatbot answers built through retrieval, it proposes that a relevant reproduction of a press publication occurs if the chatbot accessed it unlawfully, for instance by circumventing a paywall or bot block, or if the article's form of expression is recognisable to the point that a reader can skip the original. Making available to the public requires systematic recognisability. Providers are liable where such acts are inherent to their business model or where they lack reasonable safeguards. Google's implied consent argument is rejected, and facts reported in articles remain free to use.

When: The opinion was delivered today, November 3, 2026. The hearing took place on March 10, 2026. Like Company's claim covers June 13, 2023 to May 31, 2024. 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. The interpretation would apply across the EU wherever chatbots are offered to users.

Why: Publishers argue that chatbots and AI search features scrape their articles without consent or payment and divert the traffic that funds them through advertising and subscriptions. The Court has never defined when the press publishers' right created in 2019 is infringed, and this is the first generative AI case to reach it.