Europe's Top Court Adviser Drew the AI Copyright Line at Retrieval, Not Training

On November 3, Advocate General Maciej Szpunar delivered his opinion in Case C-250/25, Like Company v Google Ireland. It is the first generative AI copyright dispute to reach the Court of Justice of the European Union, and the opinion amounts to the first authoritative legal map of where liability sits.
The short version: training a model on freely accessible press articles is, in principle, covered by the EU's text and data mining exception. A chatbot that fetches an article in real time and reproduces it so completely that the reader no longer needs to visit the publisher's site is a different matter.
How a dolphin story got to Luxembourg
Like Company is a Hungarian publisher. One of its sites, balatonkornyeke.hu, reported in July 2023 that a well-known singer wanted to install dolphins in an aquarium next to Lake Balaton. A user asked Google's chatbot, then called Bard, for a summary. The bot produced a detailed answer drawing on the article.
For a publisher whose revenue tracks visitor numbers, an answer that delivers the article's content without sending anyone to the site is an existential threat, not a citation problem. Like Company sued Google Ireland in Hungary. The Budapest Regional Court referred four questions to the CJEU, asking whether a chatbot answer that partially matches a press article makes that article available to the public, whether training is reproduction, whether the TDM exception applies, and whether output constitutes reproduction by the provider.
The adviser declined to answer the training questions
Szpunar's first move is procedural. He advises the Court to declare the questions about LLM training inadmissible, for two reasons.
The first is evidence. There is no proof that Like Company's article was used to train Gemini. The article appeared in July 2023, months before the model launched, and training a model of that scale takes far longer. The AG calls it implausible that the article was in the training set. Gemini most likely found it on the internet at the moment the user asked.
The second is scope. The press publishers' right under Article 15 of the 2019 digital single market directive protects uses in an online service offered to users. Training happens behind the scenes, before any service exists, so the right does not reach it at all.
He gives his preliminary view anyway, because the question will come back. Copying texts to build and run a training dataset counts as copying under copyright law. Training is in principle covered by the TDM exception in Article 4, provided the developer accessed the content lawfully, respected machine-readable opt-outs, and took measures to prevent the model from producing near-copies of specific works.
That is the most favourable sentence any EU authority has written about AI training, and it arrived in a case where the adviser recommended not answering the question.
The model is not a copy
The opinion then does something more useful than most commentary on this topic. It separates the trained model from what the model says.
Szpunar writes that a trained language model is an abstract representation of words and language patterns rather than a database, and copyright does not subsist at that level of abstraction. The model can accidentally memorise a text that appeared many times in training, such as song lyrics, and that cannot always be prevented. If the model then reproduces that text in an answer, the output can infringe. The model itself still does not contain a copy.
He notes the practical consequence of the alternative. If offering a model required permission from every rightsholder whose work appeared in training, developing AI would become close to impossible.
Where the line actually falls
The retrieval analysis is the part product teams should read twice. A chatbot answering from a press publication may infringe the publisher's right in two situations.
It accesses the article unlawfully, for example by circumventing a paywall or other technical barrier. Or its answer reproduces the article's expression and structure so recognisably that the user no longer needs to read the original.
Liability then depends on design. Where that substituting effect is built into the business model, or follows from the absence of safeguards, the service provider is responsible.
Read together with the training section, the doctrine has a clean shape. Remembering is permitted. Re-serving someone else's work in place of the source is not. The dividing line is whether the reader still needs the publisher, not whether a machine read the article.
Why this lands harder than a fair use ruling
American litigation has spent two years arguing about fair use, which is a fact-intensive defence that a well-funded defendant can litigate for years. The EU question is narrower and more operational: what did your product do when a user asked?
For anyone building retrieval-augmented products, that reframes the compliance work. It becomes a feature-by-feature review rather than a training-data problem you solve once, and it applies to image and video search as much as to text.
A visual search tool that indexes a stock library and returns near-identical frames instead of sending the user to the licensor is running the same risk structure as a news chatbot that summarises an article. So is an assistant that pulls a specific photograph from a publisher's site and answers questions about it without a link.
The response from both sides
Google's position is that European TDM exceptions are vital for competitiveness, and that it welcomes the adviser's recognition that EU law permits training on publicly available information. The company added that Gemini creates original work by learning language patterns rather than copying, and that publishers have controls and an opt-out.
The European Publishers Council focused on the other half, arguing that the important distinction is between permitted search indexing and licensed AI answers that replace original reporting.
Both are reading the same opinion correctly. The training half favours the model developers. The retrieval half favours the publishers, and it is the half that describes how most AI products actually work today.
What happens next
The opinion is advisory. The Court follows its advocate general in most cases but not all, and the judgment is expected in 2027.
In the meantime, three things are worth doing rather than waiting. Confirm that training data came from lawful access and that opt-outs were honoured. Review any feature that retrieves and summarises source material to check whether the answer could stand in for the source. And write down the safeguards, because the opinion ties liability to their absence.
The uncomfortable part for the industry is that the test is experiential rather than technical. You cannot answer it with a licence audit or a robots.txt file. You have to ask whether a reader, given your output, would still bother clicking through. If the answer is no, the design is the problem.
Related articles
METR Spent 11 Weeks Inside Anthropic. The Number Is 1.9, and Most of the Evidence Is Blacked Out.
A policy threshold with a number in it, an auditor who says it may have been crossed, and evidence nobody outside can check.
One Percent of Firms Pay 80% of the AI Bill. That Cohort Just Stopped Growing.
A $600 million revenue cut, a 26.8% price decline, and a cohort that slipped below $10,000 per employee.
Claude Can Now Read Your Bank Statements. Anthropic Deliberately Gave It No Hands.
The read-only design is the product, and unlinking does not retract what is already in a chat.
Adobe Stopped Asking Marketers to Open Adobe
For a decade the strategy was own the interface. Adobe just bet the opposite.