EU Court Adviser Sets a Narrow Test for When an AI Answer Replaces an Article

SeaPRwire – Advocate General Maciej Szpunar advised the Court of Justice of the European Union on Tuesday to decide a publisher’s case against Google Ireland on narrow grounds, leaving the legality of AI training to another day. His opinion concerns the court’s first generative AI dispute, referred by a Budapest court hearing a claim from the Hungarian publisher Like Company.

Szpunar set the training question aside as hypothetical. Nothing in the file supported the claim that Like’s articles had been used to train Gemini, and training in any case fell outside the specific right the publisher invoked. That right comes from a 2019 European law that gives press publishers control over certain online reuse of their publications, a narrower instrument than copyright.

The instrument was written to help publishers recover their investment in journalism, and it reaches only the online use of a press publication. It does not cover the ideas inside a story, and it is not the general copyright in the text. Most of what the opinion does follows from that narrowness: a right built to protect a published article is a poor fit for a chatbot that answers questions about one.

What he examined instead was retrieval: what happens when a user asks the chatbot a question and it draws on a news article to answer.

Two routes to infringement run through that process. One is unlawful access, which includes getting past an effective paywall or a barrier against automated collection. The other is an answer that recognizably reproduces an article’s expression closely enough to replace the original, and rewording does not necessarily put a provider clear of it. The protection covers an article’s arrangement, structure and phrasing.

A further condition applies before the conduct counts as making a work available to the public. The reuse has to show up systematically in answers to related questions, which raises the bar from a single unfortunate summary to a pattern.

Szpunar attached a limit to the right itself. Invoking freedom to share and receive information, he wrote that in a democratic society no one should hold an exclusive right over information as such.

The facts that produced the test are small. In July 2023, Like reported on the Hungarian singer Kozso and his long-standing ambition to build an aquarium at Lake Balaton. Asked to summarize the coverage, the chatbot drew on the article. Like earns advertising revenue from visits to its site, and an answer delivered inside Google gives a reader one less reason to click.

At the March hearing, Like’s lawyer described that as one instance of a wider pattern of unpaid reuse that erodes publishers’ bargaining power. Google’s lawyers argued that Gemini predicts words instead of retrieving stored articles, so similarity alone did not establish copying, and said no specific reproduction had been proved.

Liability would attach where infringement was inherent in a provider’s business model, or where a provider failed to take appropriate preventive measures. Szpunar said Like’s example appeared to fall short of those conditions and left the factual assessment to the Hungarian court, which is where the case now returns once the EU judges rule.

Nicola Lucchi, a professor of comparative law at Pompeu Fabra University in Barcelona, read the outcome as a loss for the publisher. “The small Hungarian publisher in this case seems unlikely to obtain much relief,” he said.

Lucchi also disputed the treatment of displacement. He and Tim Dornis have argued that AI-generated content competing with the works used to train it can undermine their normal commercial exploitation. Szpunar treated that as ordinary competition.

Angela Mills Wade, executive director of the European Publishers Council, welcomed the rejection of the idea that freely accessible journalism carries automatic consent. She wants access restrictions and rights reservations enforced without publishers losing search visibility, and questioned whether an exception dating to 2001 can justify competing AI answers sold without a licence. “Europe’s commitment to journalism must be reflected in rights that work in practice,” she said.

Joao Pedro Quintais, an associate professor of information law at the University of Amsterdam, pointed to how demanding the replacement threshold is and to the safeguards the opinion expects of providers. On training carried out abroad, he said the advocate general doubts that downstream effects in the EU market are enough on their own, and suggests the legislature should supply a clearer answer.

Peter Mezei, a professor of law at the University of Szeged, warned against reading the opinion as a ruling on AI copyright in general, given how much narrower the press publishers’ right is. He noted Szpunar’s scepticism that reproduced text proved a model had stored copies, and his point that a training exception does not authorise later uses. Two more Gemini referrals are pending. “These are only the first steps toward clarifying the concerns at hand,” Mezei said.

Szpunar still examined training at length, and that section is the part providers can work with. Copies made for training could generally fall under Europe’s exception for automated text analysis, provided the access was lawful and rights holders had not validly opted out. Fetching an article to answer a specific question serves a different purpose, and the technical copies involved might instead fall under the exception for temporary copying, depending on how access happened and what the answer contained.

Google welcomed the part of the opinion that dealt with training. A spokesperson said the text and data analysis exemptions are vital to European competitiveness and AI development, and that Gemini learns linguistic patterns to produce original output rather than replicating content, with publisher controls that allow rightsholders to opt out.

The opinion is advice. It binds nobody, cannot be appealed, and the judgment that follows is the one the Budapest court will apply. Two other Gemini referrals are already waiting on the same judges, which means the first ruling will shape more than this dispute.

Until that lands, the practical position for a publisher weighing a claim is that the route through retrieval is open and the route through training is not, and that the second has been left to legislatures.