A dolphin, a singer and a pool that was not in the article

Like Company runs several Hungarian news portals. One, balatonkornyeke.hu, publishes freely accessible articles about Lake Balaton and earns its money from advertising on the pages. On July 21, 2023 it reported that Kozsó, a Hungarian singer, wanted to introduce freshwater dolphins into an aquarium at the lake.

An unknown user later asked Gemini for a summary, in Hungarian, of that article. Gemini complied, and described a special pool on the Tihany peninsula, 25 meters long, 15 wide and 5 deep, holding 1,000 cubic meters of water. The invented figures do not even agree with each other: a pool measuring 25 by 15 by 5 meters would hold 1,875 cubic meters, not the 1,000 Gemini stated. None of that was in the article. The opinion calls those statements hallucinations and the summary "rather approximative."

Like Company treated the answer as proof that Gemini knew its articles from training and had reproduced them. It sued Google Ireland in the Budapest Regional Court under Article 15, the press publishers' right. The Hungarian court paused the case and sent four questions to Luxembourg: two about training, two about answers. The hearing was on March 10, 2026.

Training is the question the opinion declines to answer

Szpunar advises declaring the training questions inadmissible. The dolphin article appeared on July 21, 2023; Gemini 1.0 was released in December 2023, and training a large model generally takes months, with the dataset collected well before deployment. It is "implausible" that this article trained the models. The likelier explanation is retrieval: the user named the domain and gave the chatbot enough context to fetch the article in real time. "In sum, the case in the main proceedings has nothing to do with LLM training," the opinion says.

Article 15 also covers only "online use" of press publications as part of an information-society service provided to an individual user. Training happens upstream, outside that individual service interaction. On Szpunar's interpretation, the press publishers' related right under Article 15 therefore does not extend to LLM training itself. That does not eliminate other copyright claims or the conditions attached to the text-and-data-mining exception.

He answers anyway, for completeness. Training does involve reproductions, in the scraping and tokenization that build a dataset, but a trained model is not a database or a container of copies. The Article 4 text-and-data-mining exception, in his view, generally covers training, subject to lawful access and the absence of an effective machine-readable reservation of rights. But he adds another safeguard: where a particular model could routinely generate content resembling specific training material, its developer should demonstrate measures such as dataset diversification, deduplication and prompt and output filters. If those safeguards are missing, the national court should refuse the exception.

One related claim gets a caveat. Google told the Court that Google-Extended lets a publisher block training and grounding without affecting its visibility in search. That, the opinion says, "would need to be verified."

When Gemini answers a question about current events, it searches an index, retrieves pages with bots, adds their content to the prompt and generates an answer, sometimes with links and sometimes without. The same machinery powers AI Overviews and AI Mode. A traditional search engine, Szpunar writes, would provide a list of hyperlinks. The chatbot can "collect, piece together, sum up and provide that information directly, in full sentences."

Retrieval produces copies, and the mining exception does not cover them, because retrieval uses one specific article to answer one specific question. Whether those copies are lawful turns on two things: how the chatbot got the article, and what the answer looks like.

Access first. If a publisher protects an article with 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.

The answer second. The right protects the way an article is "arranged, structured and phrased," not the facts in it. An answer infringes when its expression is recognizably the article's, to the point where 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 would settle it, but models rephrase, so paraphrase gets no automatic pass. A short factual summary, expressed independently, would generally fall outside the test. A synthesis drawing on several sources is also less likely to substitute for any one publication. Neither format is an automatic safe harbor: the question remains whether recognizable expression from a particular article has been reused to the point that the reader can dispense with that original. Lost traffic alone is not the test: "the press publishers' right was never meant to protect those publishers against competition per se."

The final question is who carries responsibility. For the purposes of the press publishers' right, Szpunar says the relevant acts must be attributable to the chatbot provider. That is the case where the reuse is inherent to the provider's business model, or where the provider has failed to take the measures expected of a reasonably diligent operator to prevent it. The fact that a user initiated the query does not, by itself, settle that question. And an open website is not implied consent. A publisher may stay open to earn advertising revenue from visits. That is not a license.

Applied to the case that brought it, the claim looks thin

Szpunar leaves the final assessment to the Hungarian court, but his guidance points one way. Like Company does not appear to have put any technical measure in place. The dolphin article's expression is not apparent in an answer that carried over the same facts in different terms, with one sentence, the article's first, appearing in the answer. Any reproduction does not appear inherent to Google's business model, and Google seems to have put appropriate measures in place.

The article itself, he notes, appears to be "a purely informative report of events," in a "commonplace arrangement, structure and sentences." That category is the hardest to protect under his test, because in a purely informative report fact and expression are close to inseparable. Distinctive, in-depth analysis is easier to recognize in an answer, and easier to protect.

There is another threshold question. Szpunar notes that Like Company and its portals seem far removed from the traditional newspapers the legislature had in mind when creating the press publishers' right. He does not conclude that the company is ineligible, but leaves it to the Hungarian court to determine whether its websites qualify as press publications under Article 15. Even before the infringement test is applied, the claimant's status remains to be established.

Read whole, this is narrower than the headline version. Publishers gain a substitution test for answers and a clear statement that an open page is not consent. They do not gain a right over facts, short summaries, multi-source syntheses or, under Article 15, training. The European Publishers Council welcomed the recognition that indexing is not authorization for substituting answers. Google welcomed the training passage. Both reactions are accurate, about different halves of the same 240 points.

The judges are not bound by any of it. Reuters reports that the Court follows its Advocates General in a majority of cases, but the judges are now beginning their deliberations and have not announced a judgment date. The Hungarian court will ultimately decide the underlying dispute. Until then, the working test is the two routes above, plus Szpunar's expectation that the opinion leaves publishers "in a strong bargaining position" and should push chatbot providers toward licensing. That position is strongest, on his own logic, for work distinctive enough to be recognized in an answer and protected well enough that a bot has to break a lock to read it.

Sources

  1. [1] Court of Justice of the European Union, Press Release No 155/26, "Opinion of Advocate General Szpunar in Case C-250/25, Like Company", Luxembourg, November 3, 2026 (primary source, read first-hand in full at Gate 1 — the first generative AI case to reach the Court of Justice; Like Company claims Google repeatedly reproduced and made available its articles through Bard and Gemini between June 13, 2023 and May 31, 2024; the two training questions are hypothetical and inadmissible, there being no evidence its articles trained Gemini and training not being "online use", though for completeness training is, as a rule, freely permitted under the text-and-data-mining exception; whether a chatbot's answers fall under the reproduction right depends on two factors — unlawful access, in particular circumvention of effective technical measures such as a paywall, and reuse of the form of expression in a recognizable manner to the extent the user is entirely dispensed from accessing the original; the making-available right is, as a rule, infringed where the form of expression is systematically recognizable in answers to related queries; the provider is liable where the acts are inherent to its business model or it failed to put in place appropriate measures; the opinion is not binding, the judges are now beginning their deliberations, and judgment will be given at a later date)Read source
  2. [2] PPC Land, "Google's Gemini can infringe press rights in 2 ways, EU court adviser says", November 3, 2026 (read first-hand in full at Gate 1 — detailed read of the 240-point opinion: the dolphin example and the hallucinated pool figures; the inadmissibility reasoning and the retrieval explanation; the completeness passage on training, including the safeguards a developer should demonstrate; the making-available distinction; the publisher-status question Szpunar flags without resolving; the four considerations pointing against Like Company's own claim; and the opinion's note that Google's claim for Google-Extended — that blocking training and grounding through it would not affect a site's search visibility — "would need to be verified")Read source
  3. [3] Reuters, Inti Landauro and Foo Yun Chee, "Setback for Google as EU court adviser warns AI chatbot may breach copyright rules", November 3, 2026 (read first-hand in full at Gate 1 via KTWB syndication — the core holding in Szpunar's words; copyright protects "only the specific form of expression", not the "mere facts"; training "freely permitted under the text and data mining exception"; no evidence Like Company's articles trained Gemini; the Court follows the majority of its advisers' non-binding opinions; Google spokesperson response; European Publishers Council reaction)Read source
  4. [4] Agence Europe bulletin, "CJEU Advocate General says chatbots can infringe press publishers' rights", November 3, 2026 (corroboration of the inadmissibility recommendation on the training questions and of the press release's core formulation)Read source
  5. [5] European Publishers Council statement, November 3, 2026, via Great Reporter (read first-hand at Gate 1 — EPC welcomed the recognition that making journalism freely accessible does not mean consenting to its use by AI chatbots, that simply rewording an article does not necessarily remove the need for a publisher's authorization, and that the making-available right is, as a rule, infringed where protected expression is systematically recognizable in answers to related queries; Executive Director Angela Mills Wade quoted)Read source
  6. [6] IPKat, "AG Szpunar's Opinion in Like Company is here", November 3, 2026 (corroboration that this is the first CJEU referral on copyright and AI, that the opinion runs to 240 paragraphs, and that it proposes inadmissibility on the training questions while answering the output questions)Read source
  7. [7] Editorial note on the claim period (M9 disclosure): the opinion itself is not internally consistent on dates — point 15 and the Court's press release give the claim period as June 13, 2023 to May 31, 2024, while point 49 refers to June 2023 to February 2024, the shorter period that also appears in wire summaries of the opinion. This article uses the full period given in the press release and does not silently swap it.Read source