An AI system’s permission to learn from material does not necessarily answer whether it may retrieve and reproduce that material for a user.
That distinction drives Advocate General Maciej Szpunar’s November 3 opinion in Like Company v Google Ireland, C‑250/25. The opinion proposes an interpretation of EU law. It is not a judgment or a finding that Google infringed publishers’ rights.
The Hungarian dispute concerns press articles allegedly used through Google’s chatbot. Szpunar recommended leaving the training questions unanswered because the record did not establish that the publisher’s articles trained the models. His alternative discussion of the text-and-data-mining exception should not be presented as blanket permission for AI training.
His proposed analysis focuses instead on real-time retrieval and generated answers. It examines access restrictions and whether an answer recognizably reuses protected expression sufficiently to replace consulting the original article. News facts themselves are not protected by the publishers’ right.
For counsel, the useful discipline is to separate three activities: acquiring training material, retrieving a publication when answering a query, and delivering the resulting answer.
A licence or defence relevant to one stage should not simply be assumed to resolve the others. Evidence should likewise identify what the system actually did.
The judges must now decide which interpretation to adopt. Until then, the opinion supplies a proposed analytical framework - not a universal answer to AI copyright disputes.
Sources: CJEU press release; Official opinion record, ECLI:EU:C:2026:919; Reuters.
Verification note: Cleared for the limited account above. Avoid reducing the opinion to “training is legal” or treating its different infringement analyses as a simple interchangeable “paywall or copying” test.
Related coverage: Publishers Sue Brave Over AI Content Supply and Alleged Paywall Bypass.