Maciej Szpunar, an Advocate General at the Court of Justice of the European Union, gave his opinion on Tuesday, Nov. 3, in what the court calls the first generative AI case to reach it. Advocates general independently propose to the court's judges how to decide a case. A Hungarian company, Like Company, which runs several websites, sued Google before the Budapest Regional Court in Hungary, and that court asked the EU court how EU law applies. Whether Like Company counts as a news publisher under the law is for the Hungarian court to decide.
The case is about what Gemini does when someone asks it about the news. The Court of Justice says the chatbot can look up web pages as a question comes in, often copying news articles, and use them to write its answer. Szpunar notes that the same method is behind AI Overviews and AI Mode in Google Search. Like Company's example is a July 2023 article on one of its sites about a Hungarian singer's dream of bringing dolphins to Lake Balaton, which it says Gemini summed up for a user.
EU law gives news publishers a right over the online use of their articles. Szpunar says that right protects the way an article is written, not the facts it reports, because it is not meant to give publishers a monopoly over the news. In his view, a chatbot's use of an article may break that right in either of two cases.
The first is when the chatbot got the article unlawfully, for example by getting around a paywall or an effective block on bots that copy web pages automatically. The second is when the answer reuses the way all or part of the article is written, even when reworded, so recognisably that the reader no longer needs to open the original at all. Repeating the facts it reports is not enough. If the article shows up that way again and again in answers to related questions, he says, the chatbot is also making the article available to the public, a second right publishers hold. The company behind the chatbot can be held responsible only when that copying is built into its business model, or when it has not taken the steps a careful operator would take to stop it reliably.
On training, Szpunar advised the court not to answer the two questions about it, because they don't arise in this case. Nothing shows Like Company's articles were used to train Gemini, he says, and the publishers' right covers only uses of articles as part of an online service, which training is not. He added that using protected work to train AI is, as a rule, allowed under an EU exception for text and data mining, the automated analysis of large amounts of text and data. That exception does not cover material the trainer could not lawfully reach, or work whose owners have opted out in a form that software can read. In his view, developers must also show they have taken steps to keep their model from turning out close copies of what it was trained on, or lose the exception.
The opinion is not binding. The court's judges have begun their deliberations and will rule at a later date. The Hungarian court will then decide the case itself, following the EU court's answer. Szpunar left that to the Hungarian court, but said it does not appear that Gemini crossed either line here. Like Company had not blocked access to its articles or automated copying of them, and the dolphin answer did not seem to reuse the way the article was written recognisably. Such copying also did not appear to be built into Google's business model, and Google seemed to have taken steps to limit the risk of it.