AI Chatbot Answers and Press Copyright: The EU Court Adviser's Test
TL;DR: On 3 November 2026, Advocate General Maciej Szpunar gave his opinion in Case C-250/25, Like Company v Google, the first generative AI case to reach the Court of Justice of the EU. He suggests that training a large language model on press content is, as a rule, covered by the EU text and data mining exception, but that a chatbot which fetches a press article in real time and reuses it in its answer may infringe the press publisher's right. The test has two parts: how the article was accessed, and whether the answer reuses the article's form of expression so closely that the reader no longer needs the original. The opinion is not binding, but any company running an AI assistant with live web retrieval should check its setup against this test now, before the judgment arrives.
Key facts
- The case: Like Company, a Hungarian operator of web portals, sued Google Ireland before the Budapest Regional Court over Gemini. It claims Gemini reproduced and made its articles available to the public between 13 June 2023 and 31 May 2024 (Court of Justice press release No 155/26).
- The referral: the Budapest court sent four questions to Luxembourg in April 2025, two on training and two on chatbot answers (The Next Web).
- Training: the Advocate General proposes that the two training questions be declared inadmissible, because there is no evidence the articles were used for training and training is not an "online use" of press publications. For completeness, he adds that training an LLM on protected material is, as a rule, freely permitted under the text and data mining exception.
- Answers: retrieval and reuse of press publications in real time to answer user queries "may, in certain circumstances" infringe the press publisher's right, according to the press release.
- Status: the opinion does not bind the Court. The judges have started deliberating and judgment will be given at a later date. Case documents are listed on the CURIA case page.
What the Advocate General actually proposed
The opinion separates two activities that are often discussed as one. Training happens once, offline, and turns large bodies of text into model weights. Grounding, also called retrieval-augmented generation (RAG), happens every time a user asks a question: the system fetches live web pages and uses them to build the answer.
On training, the message for model providers is broadly reassuring, but it is a side remark. The Advocate General asks the Court not to answer those questions at all. If the Court agrees, there will be no binding ruling on training from this case.
On answers, the opinion is more specific. The press publisher's right, introduced by Article 15 of the 2019 copyright directive, protects the investment behind a publication. It does not give publishers a monopoly on the news itself. Mere facts stay free; the protected element is the specific form of expression.
Facts versus form of expression: where the line sits
According to the press release, whether a chatbot answer falls under the publisher's reproduction right depends on two factors. A third point decides who is liable.
| Question | Lower risk under the proposed test | Higher risk under the proposed test |
|---|---|---|
| How was the article reached? | Openly accessible page, no technical barrier bypassed | Paywall or anti-scraping measure circumvented |
| What does the answer contain? | The facts reported, in new wording | The article's wording or structure, recognisably reused |
| Does the reader still need the original? | Yes, the answer points to the source | No, the answer fully replaces reading the article |
| Is it a pattern? | Occasional overlap | The form of expression is systematically recognisable across related queries |
| Who answers for it? | Provider with credible, effective counter-measures | Reuse inherent to the business model, or no appropriate measures in place |
The last row matters most for companies. The Advocate General says the chatbot provider is liable where such reuse is inherent to its business model or where it has failed to put in place appropriate measures to counter it "credibly and effectively". That is an organisational test, not only a technical one.
What changes for companies using AI with live web search
Most companies in Bulgaria and the wider EU are not building Gemini. Many, however, now run assistants that search the web: customer chatbots that answer questions about a market, internal research agents, sales tools that summarise news about prospects, or daily media digests built with automation platforms. The opinion addresses the provider of an online chatbot. A company that offers such an assistant to customers is, in ordinary terms, the provider of that assistant, so the same questions are worth asking internally even though the opinion does not say how far the reasoning extends.
The opinion also does not create new obligations under the EU AI Act. Copyright duties there sit mainly with providers of general-purpose AI models, who must have a policy to comply with EU copyright law and respect rights reservations made under the text and data mining rules (AI Act, Article 53(1)(c)). Grounding behaviour in your own product is a separate question, and this case is where it is being tested.
Who should act now
Owners of customer-facing assistants with web retrieval. Check whether answers about news topics quote or closely paraphrase specific articles, and whether they link to the source. A design that summarises facts and sends the reader to the original fits the lower-risk column.
Teams building internal news digests or research agents. Confirm how the tool reaches content. Shared logins to paid publications, scraping through proxies to get around blocks, or tools that strip paywalls fall into the first risk factor, regardless of how the output looks.
Procurement and legal. Ask AI vendors how their grounding handles press content: whether they respect technical access restrictions, how they limit verbatim reuse, and what they commit to contractually if a publisher complains.
AI governance owners. "Appropriate measures" will need evidence. A short written policy, output limits for news content, logging of retrieved sources and a takedown route for publisher complaints are the kind of controls that can be shown later.
Publishers and content-led brands. The opinion confirms that facts are free, while distinctive analysis and wording are what the right protects. Content that offers original structure and interpretation keeps its value even when AI answers summarise the underlying news.
What happens next
The Court of Justice will deliver its judgment later; no date has been announced. As The Next Web notes, Advocate General opinions are not binding, and Google has argued that Gemini does not memorise articles and that exceptions for temporary copies and text and data mining apply. The Court may follow the opinion, narrow it or take a different route. Once the judgment is given, the Budapest court will decide the actual dispute.
Until then, the practical position is simple. Training and answering are being treated as different acts with different risks, and the risk now sits with the answer your assistant gives. If your company runs AI tools that read the web on its behalf, a review of retrieval sources, output rules and vendor terms fits naturally into an existing AI governance process, and the same checks can be built into new AI automation workflows from day one.
