EU Advocate General issues first detailed framework Opinion on AI and copyright
Please confirm you are human to start the audio.
AI chatbots that look up news articles online and summarise them for users do not infringe publishers’ rights as long as they access the content lawfully and their answers do not allow users to dispense with the original article. This was the Opinion delivered by EU Advocate General (AG) Maciej Szpunar on 3 November 2026 in Like Company v Google Ireland (C-250/25), the first case before the EU Court of Justice (CJEU) on AI and copyright.
The Opinion is not binding, but it sets out the first detailed framework for applying EU copyright rules to AI chatbots.
The dispute: Like Company vs Google Ireland
Like Company, a Hungarian publisher of online news portals, sued Google Ireland after Google’s Gemini chatbot summarised some of its articles at a user’s request. The Budapest Regional Court asked the CJEU four questions, covering both the training of the AI models behind Gemini and the chatbot’s answers. CMS reported on the referral in our Legal Update of 2 September 2025.
A case about chatbot answers, not AI training
In the AG’s view, the case “has nothing to do with” AI training since the article was published on 21 July 2023, only months before the first version of Gemini was released in December 2023, and was thus unlikely to have been used for training. Gemini most likely read the article live when the user made the request. This technique, known as retrieval-augmented generation (RAG), allows a chatbot to search the internet, retrieve content from webpages and use it to formulate its answer.
In addition, in the AG’s view, the use of a press publication for AI training does not qualify as “online use” of the publication. The AG recommended that the CJEU should not answer the questions on training.
“For the sake of completeness”, however, the AG stated, in his view, training involves reproduction of protected content, which is permitted under the EU exception for text and data mining (TDM), provided of course that the content was lawfully accessible and the rights holder has not opted out in a machine-readable way. Developers should be able to show that they took steps to prevent their models from producing outputs that closely copy training material. If these steps are not taken, courts should refuse to apply the exception.
The AG also rejected the idea that a trained model contains copies of the works it was trained on, even if it can occasionally reproduce them when prompted.
Since Gemini was trained mainly in the US, the AG noted that applying EU law to training carried out abroad raises difficult territorial questions best resolved by the EU legislature.
When does a chatbot answer cross the line?
For the questions he considered relevant, the AG proposed a test based on how the chatbot obtained the content and what its answer looks like. EU rules apply whenever the chatbot is offered to users in the EU, wherever the provider or its servers are located.
First, the chatbot must access the content lawfully: getting around a paywall or measures designed to block bots makes its use of the article unlawful. Second, the answer must not replace the original. The rights of press publishers protect the way an article is written, including its structure and wording, but not the facts it reports. A chatbot may therefore pass on the facts, give a short summary in its own words or combine information from several sources. If, however, the answer follows the article so closely that the reader can dispense with the original entirely, it counts as a copy, even if reworded. A chatbot would only “make an article available to the public” if the article were recognisable to the point of direct substitution.
Primary responsibility generally lies with the user who directs the chatbot. The provider is also liable where copying is built into its business model or where it fails to take the precautions expected of a reasonably diligent operator, such as programming its bots to respect access restrictions and filtering out answers too close to the source.
Applying these principles to the case, the AG indicated that, subject to the Hungarian court’s assessment, the conditions for liability do not appear to be met. Like Company had not restricted access to its articles, the article does not appear to be recognisable in Gemini’s answer, and Google seems to have put appropriate safeguards in place.
What this means
The Opinion’s main significance lies in this framework for chatbot answers, which could shape future disputes between right holders and AI providers. The AG presents it as a middle ground. Chatbots that retrieve content lawfully, rephrase it and point users to their sources should be able to operate freely while services that bypass access restrictions or provide recognisable paraphrases should not. He also suggests that this approach may encourage licensing agreements between chatbot providers and publishers.
For press publishers, technical access restrictions and clear, machine-readable opt-outs are key to keeping control over their content, and enforcing their rights will require systematic examples of infringing answers. AI providers should ensure that their bots respect access restrictions, use output filters and document these safeguards since their diligence will be central to any liability assessment.
What’s next
The judgment of the CJEU is still to be delivered, which will also be relevant to two other Hungarian referrals on generative AI and copyright pending before the CJEU: Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26). CMS will publish a Legal Update when the judgment is issued. Previous CMS articles on the case might be found here.
For more information on AI and copyright compliance, contact your CMS client partner or the CMS experts.
This article was co-authored by Péter Virányi.