EU court adviser: AI chatbots may not copy how the news was written

An AI chatbot in Europe is free to tell you what happened in the news. What it may not be free to do is tell you in the words the journalist chose. That is the core of a long awaited opinion delivered on November 3, 2026 by Maciej Szpunar, First Advocate General at the Court of Justice of the European Union, in the case C-250/25 Like Company v Google. According to the court's press release, it is the first generative AI case to reach the EU's highest court, and the answer it proposes could shape how every AI search tool and chatbot handles news in the 27 member states.

The opinion is not a judgment. Advocates General advise the judges, who now start their deliberations and will rule later. But, as Reuters notes, the judges usually follow them, and this one draws a line that publishers, AI companies and anyone building on top of AI search will have to take seriously: training a model on press articles is, as a rule, allowed, while answering users in real time with recognisable chunks of those articles can cross into infringement.

The case: a Hungarian publisher against Gemini

Like Company runs several web portals in Hungary. It went to the Budapest Regional Court and argued that Google had used its articles to train the large language models behind Gemini, and that the chatbot had repeatedly reproduced those articles and made them available to the public between June 13, 2023 and May 31, 2024. Its claim rests on a relatively young right: the press publishers' right created by Article 15 of the EU copyright directive of April 17, 2019, which gives publishers control over the "online use" of their press publications by information society service providers.

The Hungarian court paused the case and sent four questions to Luxembourg. Two asked about training: does feeding press articles into an LLM count as reproduction, and if so, is it covered by the EU's text and data mining exception? The other two asked about output: when Gemini answers a user with content from a press article, does Google reproduce that article or make it available to the public?

The court's press release describes how Gemini works in a way that matters for the outcome. Its models use retrieval augmented generation, also called grounding: they collect information from the web in real time to make answers more accurate. When a user asks about the news, the chatbot often fetches, or scrapes, press articles to build its answer. That real time step, not the original training, ends up at the centre of the opinion.

Training: not the question here, and mostly allowed anyway

Szpunar advises the court to declare the two training questions inadmissible. In his view there is no evidence that Like Company's articles were used to train Gemini at all, and training an LLM is not an "online use" of press publications in the sense of Article 15. According to the Kluwer Copyright Blog analysis by João Pedro Quintais of the Institute for Information Law, the opinion states bluntly that the dispute "has nothing to do with LLM training", and that the example in front of the national court is more plausibly explained by the chatbot retrieving the article live.

Still, for completeness, the Advocate General gives his view on training, and it is good news for model developers. Using protected works to train an LLM is, as a rule, freely permitted under the text and data mining exception, provided its conditions are met. Those conditions include lawful access to the material and respect for any machine readable reservation of rights, the so called opt out. He also reasons that a trained model, as an abstract representation of statistical relationships, does not ordinarily contain a reproduction of the training material, and he urges "the utmost caution" before treating memorisation as proof that a work is stored inside the model.

That point is not academic. As the Kluwer analysis notes, if the court agrees, it could undercut part of the reasoning used by the Munich Regional Court in the GEMA cases against OpenAI and Suno, where German judges found that song lyrics had been memorised in models.

Answers: two ways a chatbot can cross the line

The admissible questions are about what the chatbot shows its users, and here Szpunar proposes a framework that could become the reference for AI news answers across Europe.

He starts from a principle that protects AI companies as much as publishers: the press publishers' right does not cover mere facts. It protects the specific form of expression, the investment that goes into producing a publication, and not a monopoly over the news itself. A chatbot that tells you a minister resigned, a company raised money or a storm hit a city is reporting facts, and facts stay free.

From there, a chatbot answer built on a retrieved press article falls under the publisher's reproduction right in two alternative situations:

• Unlawful access. The chatbot got to the article by circumventing effective technical measures, such as a paywall or a block designed to restrict access or web scraping.
• Recognisable expression. The answer does not just report facts but reuses the article's form of expression, including its phrasing, structure or arrangement, in a recognisable way, to the point that the user can entirely dispense with reading the original.

For the separate right of making available to the public, the test is stricter: it is infringed, as a rule, where the article's form of expression is systematically recognisable in the chatbot's answers to related user queries.

Who is responsible: the user or the provider?

A user types the question, so one could argue the user triggers the reproduction. The Advocate General does not accept that as an automatic defence. According to the press release, the chatbot provider is liable where such uses are inherent to its business model, or where it has failed to put in place appropriate measures to counter them credibly and effectively. The Kluwer analysis adds that, in Szpunar's view, liability should focus on the design of the service and its safeguards, not on occasional mistakes.

This is a design standard, not a strict liability rule. A provider that builds a system to summarise without copying, respects paywalls and shows its sources has a defence. A provider whose product depends on reproducing recognisable articles does not.

A twist: Google may win this round anyway

The headlines say a chatbot may infringe publishers' rights, but the opinion is more balanced than that. Applying his own framework provisionally to the example the Hungarian court described, the Advocate General suggests that the conditions for liability do not appear to be met, according to the Kluwer analysis. The article's form of expression did not seem to be recognisable in Gemini's answer, and Google appeared to have measures in place to reduce the risk of infringement. The final call belongs to the Budapest court.

So the real significance is not whether Like Company wins. It is that the EU's top court is being asked to set out, for the first time, the conditions under which an AI answer engine owes something to the press it reads. Reuters, via The Straits Times, framed it as a potential lifeline for online media that have watched traffic fall as AI summaries replace traditional search results with links.

The same fight, on both sides of the Atlantic

The timing is striking. One day earlier, on November 2, 2026, The New York Times and, in a separate suit, the Chicago Tribune, Daily News and Orlando Sentinel sued Brave in federal court in New York, alleging that its search index and API supply copyrighted articles to AI companies and that one of its features gets around paywalls. We covered that case in detail in New York Times sues Brave for selling its articles to AI companies.

The two systems are very different. US courts will look at fair use, contracts and the anti circumvention rules of the DMCA. The EU case turns on a specific related right for press publishers. Yet the questions are almost identical: did the AI system get around a paywall, and does its output replace the original article? Szpunar's two tests read like a checklist the US plaintiffs would recognise.

Why it matters

For publishers, the opinion is a partial win. It confirms that a chatbot cannot simply rewrite their stories in recognisable form or slip past their paywalls. But it also sets a high bar: the answer must let the user entirely skip the original, and a chatbot that only passes on facts is in the clear even if the publisher loses traffic. Copyright, the Kluwer analysis notes, is presented as unable to solve every problem AI creates for news, from source diversity to misinformation.

For AI companies, it is mostly reassuring on training and demanding on retrieval. Live web grounding, the feature that makes chatbots useful for current events, is exactly where the risk sits. Retrieval copies are not covered by the text and data mining exception, so providers will need to show that they respect access controls and that their answers summarise rather than reproduce.

For developers and businesses building on AI search, the lesson is practical. If your product pulls news into answers, two design choices now carry legal weight in Europe: never fetch content from behind a paywall or anti scraping block without permission, and make sure outputs condense facts rather than mirror an article's structure and phrasing. Showing and linking sources helps, because it points users back to the original instead of replacing it.

For readers, the likely effect is subtle but real. If the court follows this approach, AI answers about the news in the EU should become shorter, more factual and more heavily sourced, with fewer long paraphrases of a single article.

What happens next

The judges of the Court of Justice are now deliberating, and the judgment will come at a later date. Once it arrives, the Budapest court must decide the dispute in line with it, and the ruling will bind other national courts facing the same question. Two more Hungarian referrals on generative AI and copyright are already pending, Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26), and national cases such as GEMA against OpenAI and Suno in Germany will be read in the light of whatever Luxembourg decides.

There is also a political track. The Advocate General notes that applying EU copyright law to AI training carried out outside the EU raises hard territorial questions and suggests a legislative solution would bring more legal certainty. That lands as the European Commission prepares a targeted consultation on reviewing the EU copyright rules, so this opinion may influence lawmakers as much as judges.

Dany's take

This is the most sensible line anyone in a court has drawn on AI and news so far. Facts belong to everyone, and an AI that tells me what happened is doing what a good friend who reads the paper does. But an AI that sneaks past a paywall or serves me a slightly reworded copy of one journalist's work is not summarising, it is substituting. The hard part will be the grey zone in between, and the "entirely dispense with the original" bar is high enough that most chatbots will pass it. If you build with AI search, take the two tests seriously now rather than waiting for the final judgment.

Sources: CJEU press release No 155/26, CJEU case page C-250/25 Like Company, Kluwer Copyright Blog, Reuters via The Straits Times, CEA Digital Law.