EU Court Adviser Says Gemini Can Infringe Publisher Rights When AI Answers Replace the Original

The first generative-AI case to reach the Court of Justice of the European Union has produced a warning that matters far beyond one publisher and one chatbot.

On 3 November, Advocate General Maciej Szpunar said an AI chatbot can infringe press-publisher rights when it retrieves a news article from the web in real time, reuses protected expression from that article and gives the user an answer that can substitute for reading the original publication.

The case, C-250/25 Like Company v Google Ireland, concerns Google Gemini and a Hungarian publisher.

The opinion is not a judgment. The Court is still considering the case and is not bound by the Advocate General’s analysis.

Even so, the reasoning is important for publishers and search teams because it focuses on a problem that sits at the centre of AI search: what happens when an answer engine uses a publisher’s current reporting to satisfy the user’s need before the user visits the source.

The case is about retrieval and reuse, not simply AI training

The legal dispute reaches across several AI issues, but the most immediate part concerns what happens at answer time.

Gemini can use retrieval-augmented generation to find information on the live web and build a response from current sources. The Advocate General examined whether that process can amount to reproduction and making available under EU press-publisher rights.

His answer was not that every AI summary is automatically unlawful.

The opinion says the circumstances matter.

The CJEU’s summary says infringement may arise where the chatbot’s retrieval and reuse of a press publication crosses the boundaries of what EU copyright law permits. The underlying opinion also separates this real-time use from the more general question of whether publisher material was used to train an LLM.

That distinction is useful.

Training asks what material went into building a model.

Retrieval asks what the system does with a publisher’s article when a user asks a question today.

For publishers watching AI search, the second issue is often the more visible one because it can affect referral behaviour immediately.

The substitution problem is the part marketers should notice

The commercial tension is straightforward.

A publisher invests in reporting, editing and distribution. An AI system retrieves that reporting, turns it into a concise answer and presents the result inside its own interface.

If the user has enough information after reading the AI answer, the publisher may get the citation but lose the visit.

That does not mean every lost click is a copyright infringement. Copyright law does not operate as a traffic-guarantee system.

But the Advocate General’s reasoning makes the degree of reuse important. The question becomes more serious when the AI output reproduces the publisher’s protected form of expression closely enough that the user can consume the substance without needing the original.

For search and publishing teams, that creates a distinction between two very different kinds of AI visibility.

One is an answer that identifies the source and sends the user onward for the reporting.

The other is an answer that uses the reporting as raw material and leaves little reason to click.

The CJEU has not yet decided where the legal boundary sits. The opinion shows that the boundary itself is now a live EU-law question.

Access controls could become more important

The case also puts attention on how the AI system obtains the publication.

Publishers already use paywalls, robots directives, crawler controls and other technical restrictions to manage access. Those controls do not all have the same legal effect, and they should not be treated as interchangeable.

But the Advocate General’s analysis suggests that the manner of access can matter when assessing whether real-time use is lawful.

That gives publishers another reason to document how their content is exposed to AI crawlers and retrieval systems.

A practical audit should include which pages are publicly available, which require authentication, which bots are allowed, whether licensing arrangements exist, and whether the publisher can reproduce the AI response that used the content.

That last point is easy to overlook.

If a publisher later needs to show that an AI answer reused its reporting in a recognisable way, screenshots, timestamps, query wording and source-page versions become useful evidence.

NEMO has already seen a related European tension around publisher compensation in our Poland and Google publisher-pay analysis. The legal mechanism is different, but the commercial question is similar: who captures value when a platform intermediates access to journalism?

The opinion does not make AI citations illegal

There is a risk of overstating this development.

The Advocate General did not say Gemini is broadly unlawful, that AI search citations are prohibited, or that publishers automatically deserve payment every time an LLM uses information from a news page.

The opinion deals with specific EU rights and specific forms of retrieval and reuse.

It is also only an opinion.

A final judgment could accept the reasoning, narrow it or take a different approach.

That is why publishers should resist treating the case as a finished licensing rule.

For SEO and AI-search teams, there is an equally important restraint: visibility in an AI answer is not evidence that the answer infringes copyright.

The factual and legal analysis depends on what was accessed, what was reproduced and how much of the original expression appears in the generated answer.

What publishers should measure before the judgment arrives

The useful work now is evidence collection.

Publishers should identify queries where their reporting is repeatedly surfaced by Gemini and other answer engines, then compare the generated answer with the source article.

Record whether the answer merely references the reporting or reproduces distinctive wording, structure or reporting detail.

Track the referral result separately.

A source can be heavily cited while receiving little traffic. That is a commercial signal even before it becomes a legal question.

For paywalled publishers, test whether AI systems can surface details that are supposed to sit behind access controls.

For open publishers, compare which article types are most easily substituted by a generated answer. Commodity explainers may behave differently from investigations, original datasets, interviews or tools that require the reader to visit the page.

This is also relevant to content strategy. If an article can be completely satisfied by a two-paragraph AI summary, the publisher has both a distribution problem and a product problem.

The judgment will matter more than the headline

The significance of Case C-250/25 is not that Europe has already ruled against Gemini.

It has not.

The significance is that the EU’s highest court is now considering whether live retrieval and reuse by an AI chatbot can cross the line from information access into infringement of press-publisher rights.

That puts AI answer design, publisher controls and source substitution into the same legal conversation.

For publishers, the next step is not to wait for the judgment before looking at the data.

The evidence worth collecting now is simple: which articles are being retrieved, what the AI reproduces, whether the user still needs the source, and how much referral traffic survives after the answer is generated.