When an AI answer replaces the article

A proposed boundary, before the judgment

On 3 November, First Advocate General Maciej Szpunar advised the Court of Justice that fetching and reusing online news can, in some circumstances, infringe press publishers’ rights. His opinion is non-binding; judgment comes later.

The case began with Hungarian publisher Like Company’s claims against Google Ireland over Gemini. The court’s announcement explains the issue: whether retrieving a publication and using it to answer a question amounts to reproduction or making it available to the public.

The facts can travel. The article’s expression is the question.

The full opinion distinguishes reporting the same facts from recognisably reproducing how an article presents them. Its proposed test asks whether the answer retains that expression so substantially that reading the original becomes unnecessary. Changing the words alone would not settle the question: distinctive arrangement and structure can matter too.

There is a separate access question. Circumventing effective restrictions on access or scraping can make the technical copying involved in retrieval legally relevant. For making a publication available to the public, Szpunar looks for systematic recognisability in answers. Provider responsibility also depends on service design and appropriate safeguards.

Then comes a revealing qualification. In paragraph 239, Szpunar says the conditions appear not to be met in Like Company’s own example, subject to the Hungarian court’s findings. Among his reasons: the article’s expression was not apparently recognisable in the answer, and Google appeared to have mitigation measures.

Learning from text is a separate step

Training adjusts a model using examples; retrieval fetches material to help answer a particular question. Confusing those steps makes a copyright headline sound much broader than the dispute.

In his detailed reading, University of Amsterdam scholar João Pedro Quintais explains why Szpunar recommends leaving the training questions unanswered: the case does not establish that these articles trained Gemini. The adviser nevertheless discusses an exception for text and data mining—automated analysis of material—subject to conditions including lawful access and respect for valid reservations of rights.

Quintais highlights a further limit: this dispute concerns the specific right granted to press publishers. It cannot resolve every question about authors’ rights and AI. His analysis identifies substitution and the safeguards expected of a diligent provider as questions to watch if the court adopts the approach.

A lost click is not the whole legal test

The European Copyright Society’s earlier submission urged caution about building broad rules from this referral’s uncertain technological description. It separates publishers’ rights from authors’ rights and warns that a licensing economy could strengthen large intermediaries without resolving creator consent or remuneration. Quintais was one of its five drafters.

For publishers and marketers watching discovery move into answers, the distinction is practical: an answer may compete for attention without recognisably reproducing an article. The opinion does not turn every diverted visit into an infringement. Nor does it show how much traffic or revenue any publisher has lost.

The shared AI training preference proposals concern another part of this relationship. This case asks what happens when a service fetches a publication and presents an answer. The court’s judgment will determine which parts of the adviser’s approach become the EU’s interpretation of those rights.