EU adviser shifts AI copyright risk from training to answers

A top EU court adviser reads model training as covered by the data-mining exception in principle and moves liability to answers that restate publishers.
In the reading of the EU's top court adviser, training a language model on copyrighted European text is lawful in principle, and the legal exposure sits in what the model says back. That is the core of the 240-paragraph opinion Advocate General Maciej Szpunar delivered on 3 November in Like Company v Google Ireland, the first generative AI copyright case before the Court of Justice of the EU, referred by a Budapest court after a Hungarian publisher alleged that Gemini reproduced its articles in answers.
The opinion moves the copyright bill from the model to the product. Szpunar asks the Court to declare the training questions inadmissible, since the case never established that Gemini trained on the publisher's articles. Even so, he sets out his view that training falls in principle under the text-and-data-mining exception in Article 4 of the 2019 copyright directive, the rule that lets anyone mine lawfully accessible content unless the rightsholder opts out in machine-readable form. For model makers that signals the largest tail risk in European litigation is shrinking. For companies whose product fetches a third-party article and restates it, the risk rises. The adviser treats that retrieval step as a use that can infringe the press publishers' right when the publisher's form of expression is systematically recognisable in the answers.
The conditions are where the cost sits. Training qualifies only on lawfully accessed material, which rules out paywall workarounds, circumvention of technical blocks and pirate libraries. Opt-outs such as robots.txt and Google-Extended count. Where a model can readily reproduce identifiable works, developers without deduplication and output filters lose the protection of the exception. On outputs, providers are liable when infringement is built into the business model or reasonable safeguards are missing, and Szpunar rejected the argument that an open website implies consent. Facts stay free. The pattern repeats the US fair-use rulings that moved risk to data sourcing: the act of training wins, the shortcut gets a price.
For a team building agents, research assistants or summarisation tools on top of news and professional content, unit economics now carry a line for licensing or for filtering that keeps the publisher's wording out of the answer. Opt-out logs, provenance records and output checks become evidence an enterprise buyer asks for in diligence, the same way Europe's AI Act turned compliance into a product deadline. Publishers gain leverage, since their claim rests on what the answer reproduces rather than on proving training use. Fanaura looks for AI-native companies whose value comes from their own data and workflows rather than from restating other people's reporting, the distinction behind our venture thesis, and this opinion widens the gap between the two.
The objection is that the opinion binds no one, and on training it asks the Court to say nothing. The Grand Chamber could take that route and leave training open, or narrow the exception by treating safeguards as a condition few developers can prove. The Court usually follows its advocates general and will rule in the coming months, and its judgment sets the reading for every national court in the bloc. Whether it keeps the duty on outputs settles the question.
In Europe, the right to learn from the web is close to settled. The right to repeat it is about to carry a price.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building AI products on your own data and workflows for regulated European buyers? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.