During the First World War the International News Service, barred from British and French cable facilities, took Associated Press stories from bulletin boards and early East Coast editions and wired them to its own subscribers. The Supreme Court decided International News Service v. Associated Press on 23 December 1918. Justice Pitney's majority opinion called the news itself "publici juris; it is the history of the day," and still let AP stop a competitor from taking its reports while they had commercial value. Justice Brandeis dissented, arguing that a legislature, with room to weigh the public interest in information, should draw that line. A century later the European Union has a statute on the subject, and its highest court is being asked how that statute applies when the reader of the news is a chatbot.
The case is C-250/25, Like Company v Google Ireland, the first generative AI case to reach the Court of Justice of the European Union. On 3 November 2026 Advocate General Maciej Szpunar delivered his opinion. This post takes one question from it: at what point does a chatbot answer built from a news article become an infringement of the publisher's right?
Practical takeaway. On the Advocate General's reading, a chatbot that retrieves a press article in real time infringes the publisher's right if it got past an effective access or anti-scraping measure, or if its answer reproduces the article's expression so recognisably that the user can skip the original. An answer that reports the same facts in different words does not reproduce the publication, even when it diverts readers; whether the retrieval itself was lawful is a separate question. The provider is responsible where such use is built into its business model or where it lacks credible measures to stop it.
What Like Company asked the Budapest Regional Court and the Court of Justice
Like Company runs several Hungarian web portals. According to the Court's press release, it claims that Google used its articles to train the Gemini models and then reproduced them and made them available through the Gemini chatbot between 13 June 2023 and 31 May 2024. Its claim rests on the related right that Article 15 of Directive (EU) 2019/790 grants press publishers over the "online use" of their publications by information society service providers. That right lasts two years, excludes hyperlinking, and does not cover individual words or very short extracts.
The Budapest court stayed its proceedings and sent four questions to Luxembourg. Two concern training: whether training a model reproduces the publications, and whether the text-and-data-mining exception in Article 4 of the same directive covers it. Two concern output: whether showing press content in a chatbot answer is a communication to the public, and whether it is a reproduction by the chatbot provider.
Why Szpunar would set aside the two questions on training Gemini
The Advocate General treats the training questions as hypothetical. Nothing in the file shows that Like Company's articles were used to train Gemini, and the example before the national court is more plausibly explained by the chatbot fetching the article live. He adds a legal reason: training a model is not an "online use" of press publications within Article 15, so it falls outside the publishers' right that Like Company invokes.
He answers the training questions in the alternative all the same. In the account of João Pedro Quintais of the University of Amsterdam's Institute for Information Law, writing in the Kluwer Copyright Blog on 3 November, the opinion holds that training generally involves reproductions under copyright law, that the trained model itself does not ordinarily store a copy of its training material, and that such reproductions are as a rule covered by the Article 4 exception if its conditions are met, including lawful access and respect for any reservation of rights. Quintais reports that the Advocate General calls the inference that a model which can regurgitate a work must contain it "far-fetched", a point that could weaken part of the Munich Regional Court's reasoning in GEMA v OpenAI and GEMA v Suno.
Which two tests decide whether a retrieval answer reproduces a press article
Retrieval-augmented generation, which Google calls grounding, is where the opinion does its real work. The chatbot fetches specific pages to improve one answer. That act, the Advocate General concludes, can be an online use under Article 15, and it falls outside the text-and-data-mining exception because it retrieves particular content to improve an answer and extracts no patterns from it. The temporary copies it creates may still be lawful under the temporary-reproduction exception of the 2001 InfoSoc Directive.
Two alternative routes lead to infringement. The first is unlawful access: the chatbot reached the article by circumventing an effective technical measure, such as a paywall or a block on web scraping. The second concerns the answer itself. It reproduces the publication when the article's form of expression, its arrangement, structure and phrasing, is recognisable even in paraphrase, to the point that a reader of the answer can dispense entirely with the original. For the right of making available to the public, the Advocate General asks whether that recognisability occurs systematically across answers to related queries.
The opinion draws its line between facts and expression. That line sits closer to copyright's ordinary boundary than to the 1918 majority, which protected AP against a competitor's appropriation of commercially valuable news even when INS rewrote it. On the Advocate General's test, an answer that reports what the article reported, in its own form of expression, does not reproduce the publication, even if the reader never visits the publisher's site; the lawfulness of the retrieval copies is assessed separately. The press release puts the policy plainly: the right protects the investment in the publication and gives publishers no monopoly over the news.
When Google, and not the user, answers for what Gemini writes
Most chatbot answers start with a user's question. In the Advocate General's view that fact does not by itself remove the provider's responsibility, and he would attribute the acts to the provider where they are inherent in its business model, or where it has failed to put in place measures that a reasonably diligent operator would use to counter them credibly and effectively. Quintais reads this as liability for the design of the service and its safeguards, with room for occasional mistakes.
Applied provisionally to the example in the file, the test does not appear to be met. As Quintais summarizes it, the article's form of expression did not appear recognisable in the Gemini answer, and Google appeared to have measures in place to reduce the risk. The final assessment stays with the Budapest court, which must also verify whether Like Company qualifies as a press publisher at all.
What the opinion changes for publishers, AI providers and readers in Europe
If the Court follows the opinion, the commercial pressure in Europe moves from training data to retrieval. A publisher's strongest claim would concern the live answer that replaces its article, and a licence would be worth most for real-time access to paywalled or scrape-protected content. Training would sit under Article 4 and its machine-readable opt-out, which general-purpose model providers already have to respect: Article 53(1)(c) of the AI Act requires them to keep a policy for complying with EU copyright law, including reservations of rights under Article 4(3). That obligation has applied since 2 August 2025; under Article 111(3), providers of models placed on the market before that date have until 2 August 2027 to comply.
Two further Hungarian references are pending, Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26), and the judgment in Like Company will shape both. The contrast with the United States is sharp. There the Third Circuit's decision of 29 September 2026 held that training a non-generative legal search tool on Westlaw headnotes was not fair use. The EU opinion and the US ruling sit on different statutes and different facts, and a company selling one AI product on both sides of the Atlantic now meets two different legal tests for the same training data.
The human stake is the economics of reporting. Local and regional newsrooms pay for the reporting that chatbots retrieve, and the Advocate General acknowledges that copyright cannot answer every concern about source diversity, misinformation and readers' reliance on AI-generated news. His test protects a publisher against substitution, and leaves the subtler loss, a reader who learns the facts and never sees the masthead, outside the right.
How Quentir Reads It
Quentir reads the opinion as a design specification for retrieval products. The two tests translate into engineering questions a provider can answer from its own systems: whether the crawler honors access controls and scraping blocks, and whether answers on recurring news topics carry an outlet's phrasing and structure closely enough to stand in for it. The "reasonably diligent operator" standard ties liability to those safeguards, which makes their documentation part of the legal position. The strongest uncertainty is the threshold itself. "Dispense entirely" is a high bar, and the Court may lower it, keep it, or tie it more closely to the 2019 directive's text.
The same chatbot answers also carry duties under the AI Act's transparency rules. Quentir's Signature Brief 2026.3, AI Act Article 50 Disclosure Readiness, sets out what Article 50 requires providers and deployers to disclose about AI-generated content and when, including the trigger for AI-generated text published to inform the public on matters of public interest, with an internal-use license for the whole organization. This post stays with copyright. Quentir will follow the judgment in C-250/25 and the two Hungarian references that wait on it.
Sources: Court of Justice of the European Union, Case C-250/25, Like Company v Google Ireland, case documents and Opinion of Advocate General Szpunar, delivered 3 November 2026; Court of Justice press release on the Opinion, Luxembourg, 3 November 2026, as reproduced by Insight EU Monitoring; João Pedro Quintais, First Reflections on AG Szpunar's Opinion in Like Company, Kluwer Copyright Blog, 3 November 2026; Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market, Articles 2(2), 4 and 15; Regulation (EU) 2024/1689 (AI Act), Article 53(1)(c); International News Service v. Associated Press, 248 U.S. 215, decided 23 December 1918. Links checked 4 November 2026.
Published intelligence, built to inform your own decisions. Published: November 4, 2026.