EU court adviser says Gemini's news answers may infringe publishers' rights

The reference came from Hungary. Like Company runs a group of web portals. It sued Google in the Budapest Regional Court. Like Company argued Gemini reproduced its articles and made them available to the public between 13 June 2023 and 31 May 2024.

EU law gives press publishers a copyright-related right over online use of their publications. The claim rests on that right.

For news publishers, the opinion is the first word from Europe's highest court. It asks whether a chatbot's answers, not only the training of the model behind it, count as a copyright-relevant use. It is advice to the court. The court has yet to rule.

Advocate General Maciej Szpunar split the four preliminary questions in two. On the pair about training, he advised the court to declare them inadmissible. There is no evidence that Like Company's articles were used to train Gemini. Training an LLM is not a form of "online use" of press publications. That made the questions hypothetical. For completeness, he added a further point. Using protected works to train a model is, as a rule, freely permitted under the text and data mining exception in EU law.

On the answers, he went the other way.

Szpunar said an AI chatbot retrieves and reuses online press publications in real time to generate answers to user queries. That conduct "may, in certain circumstances, infringe the rights of press publishers". The opinion sets out how Gemini uses retrieval-augmented generation, or grounding, to collect information from the web as it replies. It notes that chatbots often scrape press articles when asked about the news. The answer itself can be the infringing act.

In the United States, the biggest consolidated AI copyright case is on a parallel track. In re OpenAI Copyright Infringement Litigation gathers more than a dozen suits. These include cases the Authors Guild and news organizations such as the New York Times filed. The court received 51 amicus briefs by the 30 October deadline, Publishing Perspectives reported on 3 November. Of these, 26 support the plaintiffs, 18 support OpenAI and eight are neutral. A final round of briefs is due on 6 November. The court will set oral arguments before Judge Sidney H. Stein after that.

Both sides filed cross-motions for summary judgment last month. They rehearse arguments both sides have made for two years. The plaintiffs say OpenAI infringed by copying works to train its models, including from pirated sources. They say its outputs can infringe too. OpenAI's lawyers counter that copying for training is fair use. They say the outputs do not infringe.

The goals of copyright are undermined when fair use is applied to authorize the uncompensated appropriation of the creative labor of thousands, if not millions, of copyright owners where licensing is practical, beneficial, and commercially reasonable.

The Association of American Publishers, CCC, IBPA and STM Publishers filed that brief.

The Justice Department has already picked a side. On 4 September it filed a statement of interest. It urged the court to find that training on copyrighted works is fair use. In October, Judge Stein struck a study from the evidence. The study suggested AI books were displacing human-written ones. OpenAI's lawyers argued its author had not disclosed a funding conflict. The author denies the conflict. The plaintiffs did not oppose the motion.

Music rights holders are running their own cases in parallel.

Music rights holders sued Anthropic and Suno over the training of their music models. The claims total $338 million, according to a report on SaaS Rise. The plaintiffs include Elvis Presley's estate, Reservoir Media and Round Hill. They allege the companies used automated scraping tools. They also allege the companies stripped metadata to hide where the material came from.

The Elvis case came on 23 October. It covers 274 Presley compositions. It seeks about $41.1 million in statutory damages. It alleges the model copied protected lyrics. It also alleges the model produced derivative works based on the songs.

Both companies assert a fair-use defence. They say they will defend the claims.

OpenAI, Anthropic and Suno all argue that copying for training is fair use. They also argue that their outputs do not infringe. The EU opinion complicates that position for outputs. That is at least true where a chatbot pulls a publisher's article in real time to answer a question. Nothing in it settles the US cases. The Court of Justice has not said when it will rule.