EU court adviser says Gemini’s reuse of press articles may infringe publishers’ rights

• Real-time retrieval and reuse of press articles can infringe publishers’ rights, an EU court adviser says.
• Reporting facts remains distinct from reproducing a publication’s protected expression.
• The opinion is nonbinding; the adviser recommends dismissing the model-training questions as hypothetical.

Google’s Gemini generative AI chatbot may infringe press publishers’ rights when it retrieves and reuses online articles to answer users, an adviser to the European Union’s highest court said on November 3. The nonbinding opinion gives publishers a potential basis to challenge answers that reproduce their journalism, while preserving the distinction between protected expression and freely reportable facts.

Gemini interfaces pictured in Google’s February 2024 announcement renaming Bard. Image: Google. Source: Google — Bard becomes Gemini.

A dispute over answers that replace articles

Maciej Szpunar, an advocate general at the Court of Justice of the European Union, delivered the nonbinding opinion in Case C-250/25, Like Company. He has served as an advocate general since 2013 and previously represented Poland’s government in EU court proceedings, according to his official biography. His role here is to propose an independent legal solution for the judges to consider.

Like Company, a Hungarian operator of web portals, sued Google Ireland, a member of the Alphabet group, in the Budapest Regional Court. It alleges that Google used its articles to train Gemini’s models and repeatedly reproduced and made them available through the chatbot between June 13, 2023, and May 31, 2024. Those are allegations, rather than findings of infringement. The Hungarian court referred four questions about EU law to the Court of Justice.

Gemini converses through text or speech using large language models, neural networks trained on large bodies of text. Google’s chatbot was previously called Bard; the company announced the Gemini name on February 8, 2024, during the period covered by the allegations.

The court’s account describes Gemini’s use of retrieval-augmented generation, also called grounding: fetching web information in real time to improve answers. That process can involve retrieving online journalism to answer news questions. It differs from the earlier processing of material to train a model. The court calls this its first generative AI case.

Like Company invokes the publishers’ related right under Article 15 of the EU’s 2019 Digital Single Market copyright directive. That provision gives EU-established press publishers rights over online use of their publications by service providers. Its express limits exclude hyperlinks, individual words and very short extracts, as well as private or noncommercial uses by individuals. It also leaves authors’ separate rights intact.

The distinction matters commercially. A chatbot can use a publisher’s reporting to satisfy a reader’s request before that reader reaches the publisher’s website. But under Szpunar’s proposed approach, losing a potential visit does not by itself establish infringement: the protected subject is the publication’s particular expression, rather than the facts it contains.

Access, expression and provider responsibility

Szpunar identifies two factors relevant to the publisher’s reproduction right: whether the chatbot accessed an article unlawfully, including by circumventing effective paywalls or other measures restricting access or scraping; and whether its answer merely reports facts or recognisably reuses the article’s expression so extensively that the user need no longer consult the original.

He also says the publisher’s right of making material available to the public is generally infringed when that expression is systematically recognisable in answers to related queries to that same substitutive extent. The court’s statement therefore does not describe all web scraping, all summaries or all factual answers as unlawful.

Responsibility can reach the chatbot provider where the infringing uses are inherent in its business model or it has failed to implement credible, effective measures against them, the adviser says. That makes both the answers and the provider’s safeguards relevant to the proposed liability test. Google did not immediately respond to a request for comment, according to a report on the opinion.

Training is a separate question

Szpunar recommends declaring the two training questions inadmissible because there is no evidence that Like Company’s articles trained Gemini’s models and because model training is not an online use of press publications. He therefore considers those questions hypothetical. For completeness, he says training with copyright-protected material is, as a rule, permitted under the EU text-and-data-mining exception.

That is a qualified position, rather than blanket permission to train on anything online. Article 4 of the directive provides for mining lawfully accessible material, on condition that rights holders have not expressly reserved the use in an appropriate manner, such as machine-readable means for publicly available online content. Neither the opinion nor those general conditions establish what material Google actually used.

A commercial opening with limits

The case sits alongside an existing debate about publishers’ power to authorise and charge for online use. In a separate May 12 judgment, the Court of Justice examined online service provider Meta Platforms Ireland’s challenge to Italy’s press-remuneration system and a decision by its communications regulator, AGCOM. The court said EU countries may provide for fair remuneration when publishers authorise use of their publications. Publishers must also remain free to refuse permission or grant it without charge; payment cannot be required where a provider does not use the publications.

That judgment also addressed bargaining power. The court said providers possess the data needed to assess the economic value of use, leaving publishers at a negotiating disadvantage. It found obligations to negotiate, disclose relevant data and maintain content visibility during negotiations permissible, subject to national-court checks on the framework. The visibility requirement serves to prevent pressure on publishers during bargaining.

That precedent concerns remuneration for authorised use. The Gemini reference asks the logically prior question of which chatbot uses fall within publishers’ exclusive rights. The new opinion itself awards no compensation and sets no licensing price.

Commercial licensing already coexists with litigation. The report on the opinion disclosed that News Corp, owner of The Wall Street Journal, has an agreement to supply content on Google platforms. That disclosure does not establish authorisation for the uses disputed by Like Company.

There is evidence of the audience stakes, though it does not measure Gemini. A Pew Research Center analysis used tracked browsing by 900 US adults during March 1–31, 2025. Its dataset contained 68,879 unique Google searches; about 18% produced an AI summary when researchers reconstructed the results. Traditional search-result links drew clicks in 8% of visits associated with a summary, versus 15% without one. Links within summaries drew clicks in just 1% of visits to pages with a summary.

Researchers collected the reconstructed results on April 7–17, so summaries could have changed since the original visits. The study covers Google Search across topics, rather than Gemini or only news. Query types also differed: longer searches and questions were more likely to produce summaries. The comparison does not establish that AI caused the click-rate difference or quantify Like Company’s losses.

Pew observed traditional search-result clicks in 8% of visits with AI summaries and 15% without. The study covers Google Search, not Gemini, and does not establish causation. Results were reconstructed April 7–17, 2025. Chart: Magica; data: Pew Research Center. Source: Pew Research Center.

Judges still have to draw the line

The Court of Justice’s judges will now deliberate and deliver a judgment later. The EU court interprets EU law; the Hungarian court must resolve the underlying dispute under that interpretation. The adviser’s opinion binds neither court as a judgment would.

The next decision is which legal test the EU judges adopt. Applying it to Like Company’s allegations will then require the Hungarian court to assess the disputed conduct. Under Szpunar’s proposed approach, evidence about access restrictions, recognisable reuse in answers and the provider’s measures against infringement would matter. A fall in referrals alone would not answer those questions, and this opinion is no verdict against Google.