EU court adviser says AI training on news is generally lawful, but Gemini's real-time answers may infringe publishers' rights

• System at issue: Google's Gemini chatbot, formerly Bard. The opinion does not name a model version.
• Example output: Asked about a Like Company article on a Hungarian singer's plan to bring dolphins to Lake Balaton, Gemini produced a detailed summary. The publisher says that summary contained substantial original content (Legal 500).
• Product scope: At the hearing, Like Company's complaint also covered AI Overviews and AI Mode (Digital Policy Alert).

The opinion separates two pipelines:

• Training: As a rule, using works to train an LLM is permitted under the TDM exception. Separately, the two training questions should be declared inadmissible because there is "no evidence" Like Company's articles trained Gemini (Quintarelli).
• Retrieval-augmented generation (RAG): The publisher's making-available right is infringed when an article's form of expression is "systematically recognisable" in answers (Quintarelli).
• Engineering levers: One AI-generated analysis says the opinion also covers robots.txt and Google-Extended opt-outs, the temporary-copy exception, and deduplication and output filtering. No other reviewed source confirms this (Substack).

• Training is generally lawful, but no training occurred here: The AG addresses training in principle while saying the training questions are hypothetical. Both points are reported (Quintarelli; MLex).
• When infringement may occur: Relevant factors are unlawful access, such as getting around a paywall, and reuse of an article's form to the point that users skip the original (Il Sole 24 Ore).
• Provider liability: A provider is liable when the use is "inherent to the business model" or it lacks appropriate measures (Quintarelli).
• Sourcing: The quotations above come from secondary sources that quote the opinion. Google's response does not appear in the reviewed sources.

• Legal basis: Article 15 of the 2019 DSM Directive created a related right for press publishers. It excludes individual words and very short extracts (EU Law Live; Substack).
• Why it matters: Earlier fights over this right involved search snippets. This is the first time the CJEU has been asked to apply it, and EU copyright law more broadly, to generative AI (Legal 500).

• Not binding: Advocate General opinions do not bind the court, although the judges often follow them (TokenPost).
• Narrow questions: The European Copyright Society's warning about narrow, "factually murky" questions limits how far the ruling may generalise (World Trademark Review).
• Too strict for chatbots? Quintarelli argues that the "inherent to the business model" test could overregulate general-purpose assistants, since answering questions on every topic is their core function (Quintarelli).
• Too weak for publishers? Other critics say the substitution threshold is too high and that market harm without literal copying goes unaddressed (Substack).

• Where the risk sits: Real-time retrieval, not pretraining, is now the main EU exposure for answer engines, AI search summaries and news-grounded assistants.
• Mitigations implied by the opinion:
• respect paywalls and access barriers;
• limit verbatim or close paraphrase of articles;
• track whether answers systematically substitute for publisher pages.
• Publishers: They gain leverage in licensing talks over retrieval. Claims based on training look weaker.
• Timing: The ruling is not expected until 2027 (Legal 500), so the opinion is guidance only for now.