EU Court Adviser Says AI Training Is Legal — but Gemini's Answers May Still Infringe Copyright

The Case

In Case C-250/25, Like Company v Google Ireland, Advocate General Maciej Szpunar has issued an opinion that draws a sharp line between two questions that are often conflated in the AI copyright debate: whether training a model on copyrighted material is lawful, and whether the model's outputs infringe.

The case began when a Budapest court asked the Court of Justice of the European Union two questions about training models on press articles. Szpunar advised the Court to declare both questions inadmissible, because the case before the Hungarian court had nothing to do with training at all.

Training: "As a Rule, Freely Permitted"

Despite recommending inadmissibility, Szpunar went further — "for the sake of completeness" — and offered a substantive view on training itself. In his assessment, training is "as a rule, freely permitted" under the EU's text and data mining exception, set out in Article 4 of Directive (EU) 2019/790.

The text and data mining exception allows the reproduction of lawfully accessible works for the purpose of mining them, unless the rightsholder has expressly reserved their rights. Szpunar's reading treats AI training as falling within that framework — a position that, if the Court adopts it, would give developers in Europe a significant measure of legal certainty about the training stage of model development.

It is important to be precise about what happened here. An Advocate General's opinion is advisory. It does not bind the Court, which will issue its own judgment at a date that is not yet known. But opinions from the Advocate General are influential, and the Court frequently follows them.

The Output Problem

Szpunar's reassurance about training came with a pointed warning about the other end of the pipeline: what a model says. He warned that Gemini's answers could infringe copyright in two distinct ways — by reproducing protected excerpts from source material, and by generating summaries that substitute for the original articles.

The distinction matters. A model that has been lawfully trained may still produce outputs that reproduce a protected work or act as a market substitute for it. Training legality and output legality are separate questions, governed by different rules, and a clean answer to the first does not settle the second.

The Facts of the Dispute

The article at the centre of the case appeared on 21 July 2023. Gemini 1.0 shipped in December 2023. Because models are trained months before their release, the timeline does not show that the article was actually in the training set — a point that underlines why the training questions were, in Szpunar's view, beside the point of the case before the Hungarian court.

The dispute, in other words, is really about outputs: about what Gemini produced when asked about the article, and whether those outputs reproduced or replaced the protected work.

Why This Matters for AI Companies

The opinion, if followed, would give AI developers a clearer footing on training under EU law — but it would also sharpen the focus on outputs. For companies building and deploying large language models, the practical implications are clear:

• Training data policy still matters. The text and data mining exception applies to lawfully accessible works and is subject to rights reservations. How a company sources and documents its training data remains a live compliance question.
• Output controls are the next frontier. If outputs that reproduce protected excerpts or substitute for original articles can infringe, then guardrails on generation — limiting verbatim reproduction, avoiding summary outputs that replace the source — become legal risk management, not just product polish.
• The press relationship is central. The case concerns press articles, and publishers have been among the most active rightsholders in AI copyright disputes. How courts treat summarisation and excerpting will shape the commercial relationship between AI companies and the press.

What to Watch

The Court's judgment is still to come, with no date set. When it arrives, the key question will be whether the Court follows Szpunar on both halves of the opinion — the permissive view of training and the warning on outputs — or only one. Either way, the opinion already frames the debate: training is one question, output is another, and the second may prove the harder one for AI companies to answer.

Angeled.com connects founders with investors.