On November 3, an adviser to the EU's top court issued an opinion: Google's Gemini chatbot may violate copyright rules when it uses publisher articles to answer users' questions.

The case pits Hungarian site operator Like Company against Google. The underlying question is blunt — does using news content to generate answers infringe?

The distinction the opinion rests on

Advocate General Maciej Szpunar built the argument on a classic line: copyright protects not the facts a publication reports, but the specific form of expression.

Under that logic, restating facts is one thing; reusing the expression is another, and the latter falls inside the rights holder's scope.

Two aggravating elements were cited: the chatbot accessed publications by circumventing paywalls or other technical barriers, and it reused the content to generate answers delivered to users.

Where training stands

Notably, the advocate general also said that using protected content to train large models is the kind of use permitted under the text and data mining exception.

The opinion thus separates training from output: training is treated relatively permissively; output turns on whether it substitutes for the original.

That split matters commercially. If training itself were infringing, the cost of model development in Europe would rise sharply.

Why publishers are pushing

The backdrop is a collapse in referral traffic. As AI summaries replace search results, clicks to media sites fall. Users get answers; publishers lose visits.

The European Publishers Council's position is that indexing and licensed AI answers must be distinguished: the former brings traffic, the latter removes it.

Positions on both sides

Google emphasizes that the EU's text and data mining exception is essential to competitiveness and AI development; that Gemini creates original work by learning language patterns rather than copying; and that publisher controls give rights holders clear opt-out.

Publishers want payment. A ruling favoring Like Company could force platforms to pay for such use — a realistic survival path for smaller online media.

Timeline

The court usually follows its adviser's opinion, with a final ruling expected in the coming months.

This is an EU proceeding and does not bind other jurisdictions, but it will serve as a reference point in global AI copyright negotiations.

What it means elsewhere

Rules are still forming in other markets, but the logic rhymes: input (training) and output (generation) are assessed separately; market substitution is the central test; rights reservations and technical measures will carry more weight.

Practical steps for content product teams:

Keep source and licensing records for training data.
Set citation boundaries for news content; avoid long verbatim reproduction.
Offer a rights-holder opt-out with an auditable record.

The real subject of this case is not one damages award. It is how much the "AI answer" format owes the people who produce the content it answers from. In an era where answers replace links, every market will have to answer that question once.