CJEU Advocate General Outlines Copyright Limits for AI News Summaries

Advocate General Maciej Szpunar of the Court of Justice of the European Union’s (CJEU) issued an opinion on Tuesday in the first case on generative AI to reach the Court. The opinion ultimately held that “the retrieval and reuse, in real time, of online press publications by AI chatbots in order to generate answers to user queries may, in certain circumstances, infringe the rights of press publishers.” The opinion is nonbinding on the CJEU.

The case was brought by Like Company, a Hungarian web portals operator, against The Alphabet Group, specifically, Google’s Gemini chatbot, in the Budapest Regional Court (Hungary). The Hungarian publisher claimed that Google used its news articles in Gemini without permission or payment. One of Like Company’s portals publishes articles about the Lake Balaton region, which were freely accessible, and Like Company earned revenue from advertising displayed on the portals.

Like Company argued it held press-publisher rights in its articles under Hungary’s law implementing Article 15 of the EU Copyright Directive. The company alleged that Google infringed those rights from June 13, 2023, through May 31, 2024, first through its Bard chatbot and later through Gemini.

The suit alleged the articles had been used without authorization both to train Gemini’s underlying language models and then through Gemini’s retrieval of articles in real time through web searches when answering user questions, using a process known as retrieval-augmented generation, or RAG, in which the articles were reproduced or made available to users.

In one example, in July 2023, the publisher posted an article about Hungarian singer Kozsó’s reported plan to bring freshwater dolphins to Lake Balaton. Later, a user asked Gemini in Hungarian to summarize the article and identified the publisher’s website as the source. Gemini responded with a summary that repeated the basic premise but also supplied details the article apparently did not contain, including proposed dimensions for a dolphin pool, claims about public reaction and possible environmental effects. The question raised was whether copying during retrieval, or an answer that retained recognizable expression from the article, could fall within publishers’ exclusive rights.

The Hungarian court asked the CJEU to issue a preliminary ruling on how EU law applies to chatbot use of press publications, including whether model training involves reproduction and whether an AI-generated answer can amount to reproduction or making the article available to the public. The Advocate General considered the training questions hypothetical because the record did not establish that Like Company’s articles had been used to train Gemini. The real dispute was around Gemini’s use of articles through RAG to answer user queries, said the opinion, although Szpunar offered a provisional view on training. In his analysis, training involves reproductions of works, but the EU’s text and data mining (TDM) exception can, in principle, cover those copies, including for commercial model development. That exception applies when material is lawfully accessible and rights holders have not expressly reserved its use for text and data mining. Szpunar added that developers relying on the exception should take steps to reduce the risk that their models generate outputs closely resembling specific works in their training data. National courts should consider whether developers have taken such steps when deciding whether the exception applies.

With respect to RAG, the central dispute, the opinion said that process is akin to online use of press publications under Article 15 of the EU Copyright Directive, which gives publishers rights over online use of their publications by information society service providers.

Szpunar proposed a test focused on what the chatbot retrieves and what it returns. A chatbot’s technical copying of an article may infringe publishers’ rights if it accessed the article unlawfully, such as by bypassing effective measures restricting access or scraping. The opinion explained:

“…[W]here the press publisher concerned has put in place effective technical measures (within the meaning of Article 6(3) of the InfoSoc Directive) to restrict access to that publication (be it by way of a paywall or a measure designed specifically to prevent access and/or scraping by internet bots), the chatbot  should not have circumvented those measures.”

An answer may also amount to reproduction if the article’s expression remains recognizable to the point that users can read the answer instead of visiting the original. “[T]he answer provided by the LLM-powered online chatbot will be regarded as a reproduction (within the meaning of Article 2 of the InfoSoc Directive) of the press publication reused through the RAG process where the protected elements that make up that publication are recognisable in that answer, through a side-by-side comparison,” wrote Szpunar.

A chatbot that extracts facts and presents them in a substantially different form does not, on that basis alone, reproduce the press publication, clarified the opinion. The publishers’ right protects investment in publications; it does not give publishers control over the underlying facts or competition. Whether an answer crosses the line would depend on the circumstances: “Indeed, the press publishers’ right does not subsist in mere facts,” the opinion said. “The specific form of expression of those facts in that article must be recognisable.”

A chatbot could also make an article available to the public if its answers to related queries systematically reproduce recognizable parts of the article in a way that lets users dispense with the original. A single output would not necessarily establish that kind of systematic access; publishers would need to support the claim with examples.

The opinion also addressed when conduct through a chatbot should be attributed to its provider. A user’s request may be the immediate trigger, but that does not automatically absolve the provider. Szpunar said provider responsibility may arise if infringing reuse is built into the service’s business model or if the provider fails to adopt measures reasonably expected of a diligent operator to prevent infringement. Such measures could include respecting publishers’ access restrictions and filtering outputs that reproduce articles too closely.

Szpunar rejected Google’s argument that leaving articles freely available online amounts to implied consent for chatbot use. Publishers may keep content open to attract readers and advertising, and blocking bots may also affect visibility in search engines. “Unless the exclusive rights are to be deprived of all their effectiveness, only in certain, quite specific circumstances can it be reasonably assumed that, because a rightholder has not prevented, through technical means, a certain use of his or her protected material by a third party, he or she has implicitly consented to that use,” Szpunar wrote.

Applying the proposed framework, Szpunar said the Hungarian court should determine whether Google or the chatbot crossed these lines. He noted that Like Company apparently had not installed effective measures to block access or scraping; the example answer did not clearly reproduce the article’s expression; and Google appeared to have mitigation measures in place. The national court must now make the final findings.

Image Source: Deposit Photos
Author: Zerbor
Image ID: 76616205