An AI-generated answer can help readers navigate the news. It can also remove their reason to visit the publisher’s website. But does that commercial substitution, often resulting in lost traffic, necessarily amount to infringement? According to Advocate General (AG) Szpunar, the answer is no: what matters is how the chatbot accesses the publication and whether it reproduces its protected expression.
The dispute behind the AG’s Opinion concerns Gemini’s use of articles published by a Hungarian news business. In Like Company v. Google Ireland Limited (C-250/25), the CJEU is considering its first generative AI case. AG Szpunar’s Opinion, delivered on 3 November 2026, examines press publishers’ related rights, not a claim based on authors’ copyright. The judges have yet to decide the case.
Training a model is not the same as retrieving an article
A large language model learns patterns from training data. Retrieval-augmented generation (RAG), by contrast, retrieves external material to answer questions. An article can therefore inform an answer without having been used for training.
In his Opinion, the AG highlights that there is a factual gap in the training debate: the evidence does not sufficiently establish that Like Company’s articles formed part of the model’s training data. The AG therefore sees no need for the Court to decide that issue here. In his analysis, model training falls outside of the press publishers’ right at issue (namely Article 15 of the Digital Single Market (DSM) Directive), which is limited to online use of press publications by online service providers. Authors’ copyrights still needs to be considered separately, and the AG’s observations on training are therefore expressly cautious, urging the Court not to rule on training in this case.
Where training takes place also matters. The AG does not endorse a general rule extending EU rights to all overseas training simply because the model later enters the EU market. He does not exclude a narrower connection where content is scraped from EU-based servers, but calls for legislative clarification. The territorial link is clearer for chatbot answers provided to users within the EU, since such use produces immediate and substantial effects on the territory of Member States, justifying an EU law regulation of the matter.
For AI developers, the practical message would be to check the conditions, and not assume that every training copy needs a licence (assuming that the CJEU follows the AG). The EU text and data mining (TDM) exception may allow training on copyrighted material, provided access is lawful, rights have not been appropriately reserved and the other conditions are met. Developers should also test whether the model reproduces training material too closely. The AG expects developers relying on the exception to demonstrate measures limiting such outputs and would otherwise refuse the exception. This is proposed conditional route, not a blanket exemption for training.
For rightsholders, an effective TDM opt-out is equally important. Although there is some doubt as to whether scraping technologies can reliably understand the reservation when accessing the content, the practical lesson is to make reservations clear, technically detectable and consistent across authorised copies online, rather than assume that any general notice will suffice.
Facts remain free, but paraphrasing is not a safe harbour
The press publishers’ right protects an article’s expression, including its arrangement, structure and phrasing, not the underlying facts themselves. Individual words and very short extracts are excluded. The AG asks whether protected expression remains recognisable beyond that threshold, with such a substantial similarity that the answer becomes a direct substitute for the article and makes accessing it unnecessary. Paraphrasing is not enough if the original expression survives. Equally, answering a user’s factual question is not, in itself, automatically a reproduction of the article. The question remains what expression, rather than information alone, has been reused.
Consider two hypothetical news assistants: one combines facts from several sources into a new overview; another follows an investigation’s distinctive narrative, replacing sentences with synonyms. Under the proposed approach, those are materially different uses.
The important limit is that lost traffic alone is not enough. An answer may compete with an article while reproducing only its facts. The AG does not treat the publishers’ right as protection against competition generally.
Lawful access and lawful outputs are separate checks
The less visible copying matters too. A chatbot may copy an article while retrieving it, even if its final answer uses only facts. Those copies need their own legal basis. For the particular retrieval process examined, the AG considers the TDM exception inappropriate, but accepts that the separate exception for temporary copies may apply, provided all conditions are met.
Among other conditions, those copies must support a lawful use. Circumventing an effective paywall or anti-scraping measure can prevent reliance on the exception. Lawful access does not, however, excuse an answer that infringes the publisher’s rights: access and outputs require separate checks. Similarly, free access does not necessarily imply consent to chatbot reuse, especially where blocking crawlers may harm search visibility (for instance, where indexing and chatbot retrieval share infrastructure). Remaining discoverable does not mean accepting every downstream use.
Responsibility turns on design, safeguards and evidence
The provider’s role is a separate question. The AG does not propose automatic provider liability whenever a user prompts an infringing answer. But pointing to the user is not enough if reproducing protected content is facilitated by the service’s business model or reasonably expected safeguards are missing. In practice, providers should check both how their systems access articles and how closely their answers reproduce them. The AG’s proposed safeguards address both.
A chatbot could also make a publication available to the public, although the AG sets a different evidential hurdle. In his opinion, a publication could only be regarded as being made available to the public if all or part of the publication is systematically a recognisable expression, to the point were it constitutes a direct substitution. Although the AG does not exclude this possibility, he highlights the significant degree of randomness involved in the provision of chatbot answers. He nevertheless considers generative delivery as specific technical means, different from ordinary website access.
For publishers, the practical starting point is to record what the chatbot accessed and what it returned. For retrieval and reuse, evidence of access and a reasonable sample of similar or linking answers would shift the burden to the provider to rebut the claim. A claim that the service makes protected content available to the public needs more: answers to related queries must systematically reproduce recognisable expression and directly replace the articles. One unusual answer is not enough, and testing should document repeated ordinary queries.
Importantly, subject to the national court’s findings, the AG doubts that the facts of the case amount to infringement: effective access restrictions and recognisable expression appear absent, the relevant acts do not appear inherent in Google’s business model, and Google seems to have adopted appropriate safeguards.
The last word on the relationship between copyright, related rights and AI has yet to be written . Despite various public bodies and institutions publishing studies and guidelines, the legal framework remains far from crystalized. While the CJEU typically follows an AG’s Opinion, there is also no guarantee that it will do so in this case.
If you have any questions related to AI and intellectual property rights, do not hesitate to contact the Stibbe experts mentioned below.
Authors:
- Thibau Duquin, Associate at Stibbe
- Anya Murphy, Associate at Stibbe
