In short
On 3 November 2026 Advocate General Maciej Szpunar gave his opinion in Case C-250/25, Like Company v Google, the first generative AI case to reach the EU's top court. He said a chatbot answer built from a news article can infringe the publisher's rights in two situations: when the bot got around a paywall or anti-scraping block, or when its answer reuses the article's wording so closely that nobody needs to read the original.
- Facts stay free; only the specific form of expression is protected
- Training AI models is, as a rule, allowed under the EU text and data mining exception
- The opinion is not binding; the Court of Justice will rule later
The case started in Budapest. Like Company, which runs Hungarian news portals, sued Google over its Bard and then Gemini chatbot, arguing that between June 2023 and May 2024 the bot reproduced its articles and showed them to users. The Hungarian court sent four questions to Luxembourg, and on Monday the court's adviser answered them. The Court of Justice press release sets out the reasoning, and RTÉ reported that Google said Gemini learns language patterns to create original output and gives publishers extensive controls.
Gemini, like most chatbots that answer questions about current events, uses retrieval-augmented generation: it fetches pages from the web in real time and writes an answer from them. That is the same technique behind reading assistants, book recommendation bots and the companion apps some writers now build for their fans. Small presses, literary magazines and author collectives that want to build retrieval-augmented AI features of their own now have a clearer picture of where a European court might draw the line.
The right at stake belongs to press publishers, not to novelists, so this is not a ruling about books. But the logic, about what counts as copying when a machine summarises, will be quoted in every argument about AI and writing that follows.
What did the adviser say?
Szpunar started from the principle that the press publisher's right, created by Article 15 of the 2019 Copyright in the Digital Single Market Directive, protects the investment in a publication, not the news itself. A chatbot that tells you a minister resigned is reporting a fact. A chatbot that hands you the reporter's paragraphs with a few words swapped is doing something else.
He also addressed responsibility. In his view the chatbot provider is liable when such use is built into its business model, or when it has not put in place appropriate measures to stop it credibly and effectively. That puts the burden on the companies that design these systems, not only on the users who type the questions.
What are the two tests?
- How the text was obtained. If the bot accessed an article unlawfully, in particular by circumventing technical measures such as a paywall or a block on web scraping, its use falls under the publisher's reproduction right.
- How much of the expression survives. If the answer only reports facts, there is no problem. If it reuses the article's form of expression in a recognisable way, so that the reader no longer needs the original, it may infringe. Where that happens systematically across related queries, the right of making available to the public is also infringed.
What about AI training?
The training questions went the other way. Szpunar advised the court to declare them inadmissible, since there was no evidence that Like Company's articles had been used to train Gemini and training is not an “online use” of press publications. For completeness he added that using protected works to train a large language model is, as a rule, freely permitted under the text and data mining exception, which leaves rights holders with the opt-out tools the directive already gives them.
For authors that is a sobering line. It echoes the debate we covered in Australia's proposed AI opt-out for books: the fight is moving from whether models may read your work to what they are allowed to say with it afterwards.
What does it mean for writers?
First, protection follows expression. Distinctive voice, structure and phrasing are what a court can recognise in a machine's answer; a plain list of facts is not. That is one more argument for writing with a voice nobody can mistake.
Second, technical barriers count. If you publish essays, serials or reader guides on your own author website, a paywall or a clear block on AI crawlers is not just a gesture. In the adviser's reading, getting around it changes the legal position of whoever does.
Third, if you build AI tools around your own books, design them to point readers back to the source rather than replace it. A book character chatbot that quotes sparingly and links to the chapter is both safer and better for sales. The Court of Justice usually follows its advisers, but not always, and a judgment is expected in the coming months. Until then this opinion is the clearest guide Europe has.
Independent reader guides, science notes and writing craft.
