The Like Company v Google case we are all waiting for: Advocate General Szpunar’s opinion in the first CJEU case on generative AI and copyright

On 3 November 2026, Advocate General Szpunar delivered his opinion in Like Company v Google Ireland (C-250/25). His core message is that the case concerns what a chatbot fetches from the internet in real time, not what its model was trained on. He advises the Court to declare the questions on LLM training inadmissible, while indicating that, in his view, the text and data mining exception (TDM exception) in principle covers such training. 

On the real issue, a chatbot reusing news articles in real time, he proposes a middle ground. A chatbot would only infringe the press publishers’ right if it accesses an article unlawfully or if its answer mirrors the article’s wording and structure so closely that the user no longer needs to read the original. 

Background 

The first case on generative AI and copyright to reach the Court of Justice began with a story about dolphins. Like Company, a Hungarian news publisher, reported on balatonkornyeke.hu that the well-known singer Kozsó wanted to put dolphins in an aquarium next to Lake Balaton. When a user asked Google’s chatbot Gemini (then Bard) for a summary of that article, Gemini gave a detailed answer drawing on information from Like Company’s news outlets. For a publisher that earns advertising revenue in proportion to its number of visitors, a chatbot that serves up the content of an article without sending readers to the publisher’s site is a direct threat. Like Company therefore sued Google Ireland in Hungary for reproducing and making available its press publications without consent, and the Hungarian court referred four questions to the Court of Justice.

Questions In essence, the Budapest court asked the CJEU: 

1. Does a chatbot answer displaying text partially identical to a press article make that article available to the public (Article 15(1) DSM Directive and Article 3(2) InfoSoc Directive), and does it matter that the chatbot merely predicts the next word? 

2. Is training an LLM-based chatbot a reproduction (Article 15(1) DSM Directive and Article 2 InfoSoc Directive)? 

3. If so, is it covered by the TDM exception in Article 4 DSM Directive? 

4. If a chatbot displays all or part of an article in response to a user prompt, is that a reproduction by the chatbot provider? 

The Advocate General’s opinion 

LLM training (questions 2 and 3) 

The Advocate General (AG) first explains that LLMs and chatbots are two different things. The questions are therefore imprecisely worded and should be read as referring to the “process of training an LLM”. 

The AG considers that the Court should not answer the questions on training, for two reasons. 

First, there is no evidence that Google used Like Company’s articles to train Gemini. The dolphin article appeared in July 2023, only months before Gemini was launched, and training such a model takes far longer. He notes that it thus seems “implausible” that the article was used for the training of the model. Gemini most likely found the article on the internet at the moment the user asked about it (see the explanation of RAG below). 

Second, the special right that protects press publishers only covers uses of their articles in an online service offered to users. Training happens behind the scenes, before any service exists, so that right does not apply to training at all. He nevertheless shares his preliminary view on training. Copying texts to build and run a training dataset counts as copying under copyright law. 

Training is in principle covered by the TDM exception (Article 4 DSM Directive), provided the developer accessed the content lawfully, respected machine-readable opt-outs and took measures to prevent the model from producing near-copies of specific works. 

Whether EU law also covers training carried out outside the EU is, in his view, a question the EU legislator should resolve. 

The Trained Model and Outputs 

The trained model itself is a different matter. The AG stresses that a trained LLM is not a database but an abstract representation of words and language patterns, which can hardly qualify as a reproduction of the works used, since copyright does not subsist at that level of abstraction. He acknowledges that a model sometimes “memorises” a text by accident while it learns, for instance song lyrics that appear many times in the training data, and that this cannot always be prevented. 

If the model then reproduces that text in an answer, the output can be infringing but this does not mean that the model itself contains a copy. In other words, infringement lies in what the model produces, not in the model as such. Holding otherwise would mean that companies need the rightsholders’ permission just to offer a model, which would make developing AI very difficult. Rightsholders still keep some protection. They can refuse or license the use of their works for training, and act against infringing answers. The AG expressly states that this is his view “at this stage”. He thereby departs from the Munich Regional Court, which held in GEMA v OpenAI that a model that “memorises” song lyrics does contain reproductions of them. That judgment is under appeal before the Munich Higher Regional Court. 

Chatbots looking up news in real time (questions 1 and 4) 

According to the AG, this is the real heart of the case. When answering a question, a chatbot can search the internet, read relevant articles and base its answer on them. This technique is called retrieval-augmented generation, or RAG. RAG is an online use of press articles that the press publishers’ right can cover, and EU law applies as soon as the chatbot is offered to users in the EU. Because chatbot providers are not required to disclose which content they retrieve, the Advocate General suggests easing the burden of proof for publishers. Courts could presume that a chatbot used an article where the article is accessible online and the publisher produces a reasonable sample of answers that are similar to it or link back to the webpage hosting it. The provider would then have to prove the contrary (point 163). This presumption only concerns the use of the article, not whether that use infringes. 

The AG then answers five questions. 

1. Are the copies made “behind the scenes” allowed? 

When a chatbot fetches an article, it makes temporary copies of it on its systems. The text and data mining exception does not help here, because the chatbot is looking for one specific article rather than analysing large amounts of text. For such short-lived technical copies, another exception can apply according to the AG, namely the “temporary reproduction exception” (Article 5(1) InfoSoc Directive). It applies only if two conditions are met: 

(i) the chatbot accessed the article lawfully, without bypassing paywalls or blocks on bots, and 

(ii) the answer it gives the user does not itself infringe the publisher’s right. 

2. When does the answer infringe the publisher’s right under article 15 DSM Directive? 

The Court’s usual two-step test applies. 

(i) First, the article must have been copied to produce the answer. That is always the case with RAG, because the article is fed to the AI model before it writes its answer. 

(ii) Second, the article must still be recognisable in the answer. The publisher’s right protects the way an article is written (its structure, wording and presentation), not the facts it reports. Rewording therefore does not automatically make an answer safe. The question is whether the article remains so recognisable that the reader no longer needs to read the original. A short summary of the facts, or an answer combining several sources, does not infringe. 

3. Does the chatbot make articles available to the public? 

Only if it reproduces the article recognisably systematically when users ask related questions. Chatbot answers vary each time, so this will be rare unless the system is built, deliberately or carelessly, to work that way. 

4. Who is responsible? 

The publisher’s right only applies to the service provider, not to private users. In principle, the user who asked the question is responsible for what he or she asks the chatbot to do. The provider is also responsible if reusing articles is an inherent part of its business model, or if it fails to take reasonable precautions, such as respecting access restrictions and filtering answers that come too close to the original. 

5. Did publishers implicitly agree? 

No. Leaving articles freely available online does not amount to consent. Publishers were never clearly told about this use, and measures that publishers can put in place to restrict access are not simple. Choosing not to set up a paywall is an important business decision. Blocking the bots Google uses for its chatbot also seems to remove the website from Google Search, because Google uses the same bots for both. 

Application to the case 

It is for the Hungarian court to decide once the Court of Justice has ruled, but the AG doubts that Google infringed Like Company’s rights. Not blocking bots does not mean that Like Company consented, but it does mean that Gemini accessed the article lawfully. Gemini’s answer contained the same facts but only one sentence of the original. 

Conclusion 

The opinion draws a clear line between what a chatbot reads and what it writes. For press publishers, their IP right offers no protection against AI training as this cannot be considered an ‘online use’ of their articles. Against training, they can rely on machine-readable opt-outs from the TDM exception. Against chatbots retrieving articles in real time, they will need effective technical measures restricting access or scraping, and evidence that they content was used by the chatbot. The AG proposes to ease that burden of proof. For AI companies, real-time retrieval is no free pass. Lawful access is a condition for relying on the temporary copies exception, and filtering answers that come too close to the original is part of what a diligent provider must do to avoid liability. 

The opinion is not binding, and the Court may still choose to rule on training. The judgment will show whether the Court follows it. 

Timelex is actively monitoring case law on LLM training and generative AI. If you have any questions related to AI, don’t hesitate to contact us.