Gemini and copyright: Advocate general says training is generally allowed, but the answers could be a problem

An advocate general at the EU Court of Justice has given his view on how copyright should apply to AI chatbots. On November 3, 2026, Maciej Szpunar delivered his opinion in the case Like Company v. Google Ireland (C-250/25). It is the first case before the EU Court of Justice on generative AI and copyright.

The opinion is not binding. The Court of Justice will issue its own ruling and is not obliged to follow the advocate general's assessment. The oral hearing was held on March 10, 2026.

The case concerns Google's Gemini

The case was brought by the Hungarian press publisher Like Company. It concerns the period from June 2023, when the publisher claims that Google's chatbot Gemini reproduced and made parts of the publisher's articles available without consent. The end of the period is unclear, because the sources differ: Crypto Briefing gives February 2024, while Substack gives May 2024.

The advocate general distinguishes between two phases: when the model is trained, and when it generates answers.

Training: generally allowed, according to the advocate general

According to Szpunar, training AI models generally falls under the text and data mining exception, provided the developer meets the conditions. That requires lawful access and respect for rightsholders' effective reservations, for example a robots.txt file or metadata. These are technical markers that let a website opt out of having its content used.

The advocate general does, however, propose that the questions about training be dismissed as inadmissible. His discussion of training is therefore preliminary.

Circumventing technical protection measures undermines the protection under the exception, he says. And the fact that something is publicly available on the web does not automatically permit its use in AI.

He also sees the model's parameters and stored documents as two different things. But he acknowledges that systems that can consistently produce recognizable passages may raise other copyright questions.

Answers: where it can go wrong

When the chatbot retrieves content, it is called retrieval-augmented generation. This can create temporary copies, which according to the advocate general are subject to different rules than training.

According to Crypto Briefing, Szpunar says that when a chatbot retrieves an article from a press publisher's website, reuses it for an answer and delivers the answer to the user, the publisher's right may in certain circumstances be infringed.

To assess an infringement, he proposes two questions: Is the protected form of expression still recognizable? And does the AI answer replace the need to visit the original publication?

Copyright does not protect the facts that media report, only the specific form of expression, that is, the way the article is worded.

The sources do not indicate how Google or Like Company has responded to the opinion.