An adviser to Europe's top court has drawn a line through generative AI: training a model on press articles is allowed, but a chatbot answer that reuses an article in recognisable form can infringe the publisher's rights. Advocate General Maciej Szpunar gave that view on November 3, 2026, in Case C-250/25, Like Company v Google Ireland, the first Court of Justice of the EU case on generative AI and copyright. The News (Pakistan) quotes the court's statement, and wire coverage ran the same day.
The opinion is not binding. But the Court's judges often follow their Advocate General, so it is the clearest signal yet of how EU law may treat chatbots such as Gemini. This post explains what he said, what he left open, and what it means for builders and publishers. It extends our coverage of the AI lawsuit landscape and of Google's CNN deal and publisher pilot.
TL;DR: the opinion in one table
| Question | Answer | Source |
|---|---|---|
| Who gave it? | Advocate General Maciej Szpunar, EU Court of Justice | Court statement, via press |
| When? | November 3, 2026 | Press reports |
| Case? | C-250/25, Like Company v Google Ireland, referred by a Budapest court | CJEU case information, law-firm notes |
| Training on press text? | "Freely permitted under the text and data mining exception" | Szpunar, quoted by The News |
| Chatbot answers? | The publisher's right "may be infringed in certain circumstances" | Szpunar, quoted |
| When exactly? | Access gained by getting around paywalls or technical barriers, or reuse of the "form of expression" in a "recognisable manner" | Szpunar, quoted |
| Facts? | Not protected; "only the specific form of expression" is | Szpunar, quoted |
| Binding? | No. The Court decides later | Press reports |
What exactly did Szpunar say?
Per the court statement as quoted by The News, Szpunar split the problem in two stages.
Training. The use of copyright-protected material to train large language models is "freely permitted under the text and data mining exception." That is the EU exception that lets machines analyze lawfully accessible text for patterns. This is the part that AI developers wanted.
Answering. When a chatbot "retrieves a publication from a press publisher's website, reuses that publication to generate an answer, and provides that answer to the user, the press publisher's right may be infringed in certain circumstances." He named two. First, where the chatbot gets access by circumventing paywalls or other technical barriers. Second, where it reuses "the form of expression of that publication in a recognisable manner."
He added that copyright does not protect the "mere facts" in a news story, only the specific form of expression. So a chatbot that states what happened in its own words is on firmer ground than one that returns a near-copy of the article.
What is the case about?
Like Company, a Hungarian publisher, says Gemini produced answers that echoed significant parts of its articles. The Budapest court asked the Court of Justice in April 2025 how EU copyright and press publisher rules apply. Law-firm summaries of the hearing on March 10, 2026 report Google's position: Gemini does not store or retrieve copies of the articles. That is Google's argument as summarized by third parties, not a finding.
According to press coverage of the opinion, a ruling against Google could mean fees for use of publisher content, which publishers hope would help outlets whose traffic has dropped as AI summaries replace search results with links.
Why the training versus answering split matters
For two years, the loudest copyright fights about AI have bundled two questions: was the model trained on protected work, and does its output reproduce it. U.S. cases such as the ones in our NYT fair-use brief coverage and the Wikimedia amicus on training versus retrieval have been wrestling with the same split. The EU adviser has now put it in a clear sentence: training is allowed, retrieval and reuse is where liability sits.
That has a practical consequence. Retrieval-augmented products, which fetch live pages and compose answers, carry more legal exposure than a model that answers from its weights. Search-style features such as Google's AI Mode are exactly this retrieval pattern.
Two EU rules behind the opinion
The opinion rests on two parts of EU copyright law, and they pull in different directions.
The first is the text and data mining exception. It lets machines analyze lawfully accessible works to find patterns, and it is the basis for Szpunar's view that training is allowed. Rightsholders can usually opt out of the commercial version of this exception in a machine-readable way, which is why opt-out signals such as robots.txt matter to both sides.
The second is the press publishers' right, created by the 2019 EU copyright directive. It gives news publishers their own right over online use of their press publications. The Hungarian court asked whether a chatbot answer counts as such a use. Szpunar's answer was a conditional yes: it can, when the answer circumvents barriers or reuses the article's form of expression recognisably.
A short timeline of the case
- April 3, 2025: the Budapest District Court refers Like Company v Google Ireland to the Court of Justice, raising the first preliminary questions on generative AI and copyright, per law-firm summaries.
- March 10, 2026: the oral hearing takes place before the Court of Justice, per law-firm summaries.
- After a rescheduling: the opinion date moved from an earlier slot (September 3, 2026, per Penningtons) to November 3, 2026.
- November 3, 2026: Szpunar delivers the opinion.
- Next: the Court's judgment, then the Hungarian court applies it to the facts.
How does this compare with the U.S. debate?
| Point | EU adviser's view (Nov 3) | U.S. debate |
|---|---|---|
| Training on protected text | Freely permitted under the TDM exception | Argued under fair use; cases still running |
| Output that echoes articles | May infringe the press publisher's right | Argued as infringement or as fair use, case by case |
| Facts in news | Not protected, only the form of expression | Facts also unprotected; expression is protected |
| Legal weight today | Non-binding opinion before a final EU ruling | Mixed district rulings and appeals |
The comparison shows why this opinion matters outside Europe. A court-level voice has now said, in plain terms, that the legal risk sits in what the chatbot hands back to the user and how it got there, not in the training run. Companies that ship in both markets will need to design the answer layer for the stricter reading. That means citing and linking rather than copying, and it means being able to show how a page was fetched.
What it does not say
- It is not a ruling. The Court of Justice still has to decide, normally some months after the opinion.
- It does not say Gemini infringed. The language is conditional: "may," "in certain circumstances." The Hungarian court must apply the facts.
- It does not ban retrieval. Summaries of facts, short quotations within legal limits and licensed uses remain possible.
- The reporting varies. Some headlines say Gemini "may breach" the law. The wording from the court is narrower. We have not read the full opinion text, and this post relies on the quoted court statement and press coverage.
- The "business model" framing needs a source check. One summary says providers could owe damages where infringing output is inherent to their business model or safeguards are missing. We could not trace that wording to the quoted statement, so treat it as press interpretation.
What builders should do now
- Do not bypass paywalls. Fetching gated pages with credentials, tricks or scrapers is the clearest risk in the opinion.
- Prefer paraphrase over reproduction. Return facts in your own words with a link to the source. Avoid long verbatim passages of articles.
- Honor opt-outs. Respect robots.txt and emerging signals, such as the shared robots.txt AI training signal work OpenAI said it would honor.
- Log what you fetch. If a publisher asks, you need to show what was retrieved and what was returned.
- Watch licensing. Deals like Google's with CNN suggest paid access is becoming the safe route for premium news.
What publishers should do now
Keep technical barriers and machine-readable opt-outs in place, since the opinion treats circumventing them as a trigger. Track how your articles surface in chatbot answers, with dated examples. And note that the training half of the opinion cuts against publishers: if the Court follows it, a training-data claim in the EU is hard, while an output claim has a route. The same pattern appears in the New York Times' recent suit against a search company over scraping and paywalls, covered in our report on the Brave complaint.
What happens next
The Court's judges will deliberate and issue a judgment, typically several months after an opinion. Commentators note the judges often follow the Advocate General but can differ. The final judgment will bind the referring Hungarian court and guide courts across the EU's 27 member states, and it will interact with the EU AI Act's copyright duties for general-purpose model providers. We will update this post when the full opinion text or the judgment is available.
Our read: the training safe harbor is good news for model developers, and the answer-stage liability is good news for publishers. The real fight will be over what counts as a "recognisable" reuse. Facts in this post are accurate as of November 3, 2026.
Related reading
- AI lawsuit tracker: 248 cases against 76 companies
- CNN and Google multiyear AI deal
- OpenAI fair use amicus excerpts in the NYT case
- Wikimedia amicus: training vs retrieval
- NYT sues Brave over AI scraping
- Google AI Mode ads and search growth
- Shared robots.txt AI training signal
- Sources: The News, EU court adviser report; Pennington's Law, Like Company v Google