EU Court Adviser: AI Training on Copyrighted Works Should Be Lawful

Training an AI model on copyrighted books, articles and other works should, as a rule, be lawful in the European Union. That is the provisional view Advocate General Maciej Szpunar set out on November 3, 2026, in case C-250/25, Like Company v Google Ireland Limited, before the Court of Justice — in an opinion on a question he told the Court it does not need to answer. The conditions: no pirate sources, respect for paywalls, bot blocks and machine-readable opt-outs, and demonstrable steps to stop the model reproducing specific works.

Szpunar advises the Court to declare the two training questions inadmissible, because nothing indicates the Hungarian articles trained Gemini, and because the right invoked does not reach training at all. He answers them anyway, "for the sake of completeness and in order to provide a proper framework for the debate," across some 95 of 239 numbered points. "I cannot stress the sensitivity of those questions enough," he writes. The stakes are broad: the TDM exception applies to every rightholder category, and the opinion says use of protected works for training "is massive, and at the heart of the development of generative AI."

Why training falls outside the dispute

Like Company Kft., which runs advertising-funded Hungarian portals including balatonkornyeke.hu, claims Google's Bard and then Gemini infringed its rights between June 13, 2023 and May 31, 2024, relying solely on the press publishers' right in Article 15(1) of the 2019 CDSM Directive. Its training evidence: Gemini could summarise a July 21, 2023 article about singer Kozsó's plan to bring dolphins to Lake Balaton.

Szpunar finds that implausible. Gemini 1.0 was released in December 2023; training "generally takes months" with datasets collected earlier, and the chatbot likely fetched the article in real time. More fundamentally, Article 15(1) covers only "online use" — training happens "on dedicated computer infrastructure, without any information society service being provided to a user." The link to chatbot answers is "simply too indirect (and, usually, diffuse)." The upshot: the press publishers' right gives news publishers no protection against training at all.

Copies at every stage — but not inside the model

The opinion divides training into four stages — corpus collection, dataset cleaning and tokenisation, the training loop, fine-tuning and deployment. Each produces copies, and Szpunar has "little doubt" they are reproductions under Article 2 of the 2001 InfoSoc Directive. Purpose is immaterial; EU law took "a purely technical approach." Nor is training "comparable to a human learning to write by reading the works of others, if only because of the industrial scale of the process." The temporary-copies exception cannot cover datasets, which have "commercial value in itself."

But a trained model is not a database. "A trained LLM is not a database," he writes; parameters reflect word use across the dataset generally. Regurgitation stems from "overfitting" and "appears incidental to the machine learning process, and not entirely avoidable." It takes "the combination of the model's parameters and of the user's prompt" to produce reproductions. This diverges from the Munich Regional Court, which ruled in GEMA v OpenAI (November 11, 2025) and GEMA v Suno (July 31, 2026) that memorised lyrics are reproductions; it aligns with the UK High Court's Getty Images v Stability AI (November 2025).

Is training text and data mining?

"I tend to think that the TDM exception does, in principle, apply," Szpunar writes. Article 4 of the CDSM Directive covers works used as data, provided works are "lawfully accessible" and rights are not reserved in machine-readable form. Lawful access excludes circumvented paywalls and bot blocks; Article 6(4) "does not grant model developers a right to hack." Robots.txt qualifies as a valid opt-out, though it "does not technically prevent scraping" and "some bots have been shown to ignore robots.txt directives." Google-Extended, released September 28, 2023, also counts; Google's claim that blocking it leaves search visibility untouched "would need to be verified."

The three-step test attaches a burden: developers "should demonstrate that they have taken the measures" — dataset diversification, deduplication, output filters — against memorisation, failing which courts should refuse the exception. On harm, LLM content "could easily replace human creators as regards texts with low originality," but not "highly original novels" or "in-depth, analytic pieces of journalism." Member states may create statutory remuneration, he adds.

Territory and next steps

On where training happens, Szpunar would not exclude Germany's middle course: EU law reaching scraping from EU servers. The AI Act, he reads, requires compliance policies "regardless of the jurisdiction" of training.

The opinion binds no one. If the Court follows the admissibility advice, judgment may say nothing on training — leaving Szpunar's analysis detailed but forceless. Takeaway: for now, the opt-out is the publisher's main lever, and the proof burden is the developer's.