First Reflections on AG Szpunar’s Opinion in Like Company

On 3 November 2026, Advocate General (AG) Szpunar delivered his much-awaited Opinion in Like Company (C-250/25), the first generative AI and copyright case to reach the Court of Justice of the European Union (CJEU). Two further cases concerning generative AI and copyright are also pending before the CJEU following referrals by Hungarian courts: Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26).

The Like Company case concerns the use of press publications by Google’s Gemini chatbot and raises questions about both large language model (LLM) training and the generation of chatbot responses.

The Opinion proposes a framework for assessing the use of press publications by AI chatbots. At the same time, it urges the Court not to answer the questions on training: in the AG’s view, those questions are inadmissible.

The Opinion is quite long, extensively sourced and, whether one agrees with it or not, thoughtful. The AG recognises the complexity of the case and does a good job of systematically addressing a poorly drafted referral by the Hungarian court. In this blog post, I will mostly try to set out its main conclusions and some brief reflections, rather than provide a detailed critical analysis. To make my own views clear, I will consider the AG’s proposed answers alongside European Copyright Society’s (ECS) Opinion on the reference, of which I was a member of the drafting committee.

The case and the questions referred

Like Company, a Hungarian press publisher, brought proceedings against Google concerning the alleged use of its articles to train Gemini and the subsequent reproduction and making available of their content through the chatbot. For readers who have time and want a fun read, I recommend checking Like Company’s claim in paras 14 to 19: it involves bringing dolphins to Lake Balaton, a welcome departure from all those hotel room cases...

The national court asked four questions. In simple terms, these concern whether:

1. displaying content from a press publication in a chatbot response amounts to making it available to the public;
2. LLM training involves acts of reproduction;
3. the text and data mining (TDM) exception in Article 4 of the Copyright in the Digital Single Market Directive (CDSMD) applies to those reproductions; and
4. displaying content from a press publication in a chatbot response amounts to reproduction by the chatbot provider.

A central feature of the Opinion is its distinction between training an LLM and using a trained model to generate an answer. The latter can involve retrieval-augmented generation (RAG) or ‘grounding’: the chatbot retrieves online material and uses it as context when generating a response.

Two preliminary remarks before we dive in.

First, I have chosen to analyse the questions in their original order. However, readers should note that the AG considers questions 1 and 4 admissible and questions 2 and 3 inadmissible. You might therefore want to read them in that order.

Second, the AG rejects the Commission’s submission that the reference is inadmissible as a whole. Although the request leaves Like Company’s status as a press publisher insufficiently explained and describes the relevant technology imprecisely, the AG considers that Article 15 CDSMD could apply and that the Court has enough information to provide a useful answer. The referring court must ultimately verify the publisher’s status and the relevant facts. In the spirit of cooperation under Article 267 TFEU, the AG argues that the CJEU should therefore address the relevant questions and guide the referring court, “which is evidently lost in the AI landscape” (paras 26–35; the “burn” is in para 35).

Question 1: Making press publications available through chatbot responses

The AG recommends that the Court address the first and fourth questions together. In his view, a chatbot can make a press publication available to the public in the relevant EU territory where its form of expression is systematically recognisable in responses to related user queries, to such an extent that users can dispense entirely with accessing the original publication (paras 210-216 and 240).

The distinction between facts and expression plays a central role here. The press publishers’ right in Article 15 CDSMD does not grant publishers a monopoly over the facts reported in their publications. Instead, it protects the form in which those facts are presented. The fact that a chatbot generates responses through a predictive process does not per se exclude the application of the right.

The ECS likewise suggested considering questions 1 and 4 together and emphasised that the Court should carefully distinguish the technologies and acts at issue. It did not, however, propose a comparable test for systematic recognisability or substitution. Instead, it urged the Court to remain within the limits of the press publishers’ right and to avoid drawing broad conclusions from an uncertain factual record.

Question 2: Does LLM training involve reproduction?

The AG considers the second question “hypothetical and, thus, inadmissible”. First, there is no evidence that Like Company’s articles were used to train the Gemini models. As the AG emphatically states: “the case in the main proceedings has nothing to do with LLM training” (para 50). In his assessment, the example before the national court is more plausibly explained by the chatbot retrieving the article in real time through RAG. Secondly, he considers that LLM training is not an “online use” of press publications within the meaning of Article 15 CDSMD and therefore falls outside the press publishers’ right relied on in the proceedings (paras 46-56). In my view, this is a consequential clarification of what the AG calls “the limited scope of the press publishers’ right”.

This conclusion does not place AI training outside copyright law altogether. In his alternative analysis, the AG considers that training generally involves reproductions of protected material under Article 2 of the InfoSoc Directive. He distinguishes those training copies from the trained model itself, whose abstract representation of statistical relationships does not ordinarily constitute a reproduction of the training material.

The AG also discusses the complications raised by memorisation and regurgitation (paras 74-92). A separate blog post would be needed to do justice to this analysis. Here, it is sufficient to say that he advises “the utmost caution” (para 87), noting that “‘memorisation’ appears incidental to the machine learning process, and not entirely avoidable” (para 88). In his view, the fact that an LLM can generate a reproduction of a work on which it was trained does not mean that a reproduction of that work is stored inside the model. He considers that inference “far-fetched” (para 90; see also paras 91-92). If the CJEU agrees, this could call into question an important part of the Munich Regional Court’s reasoning in its judgments in GEMA v OpenAI and GEMA v Suno, which the AG discusses in this context (footnote 80).

There is a significant point of contact with the ECS Opinion. The ECS also questioned whether training reproductions fall within Article 15, given that the provision protects press publishers against the online use of their publications by information society service providers. It was, however, more conditional: it argued that training copies could fall within the press publishers’ right only if they qualified as reproductions for such online use. The AG goes further by concluding that LLM training falls outside that requirement.

Both the AG’s Opinion and the ECS Opinion also caution against treating a dispute based on the press publishers’ related right as an opportunity to resolve all questions concerning authors’ rights and AI training.

Question 3: Does the TDM exception cover training?

The AG also recommends declaring the third question inadmissible. Nevertheless, he addresses it for completeness. In his view, reproductions of protected material for LLM training are, as a rule, covered by the TDM exception in Article 4 CDSMD, provided its conditions are satisfied, including lawful access and the absence of an applicable reservation of rights (often referred to as the “opt-out”).

The AG reads the definition of TDM in Article 2(2) CDSMD broadly enough to encompass the automated extraction of linguistic patterns and relationships used to train an LLM. The subsequent use of those patterns to generate text does not, he states, remove the analysis from the definition of TDM (paras 97-109).

This is broadly consistent with the ECS Opinions on Like Company and on Copyright and Generative AI (and my own previous scholarship, e.g. here), which also consider that Article 4 can apply to qualifying reproductions made for AI model training, subject to the exception’s conditions. In its Like Company Opinion, the ECS emphasised that this analysis should be confined to training and distinguished from RAG that does not form part of the model’s learning process.

Finally, the AG embarks on a lengthy analysis of the three-step test, concluding that exempting the use of protected material for LLM training under the TDM exception does not, as a rule, violate the conditions of the test. He notes, however, that such training might run afoul of the test in “exceptional situations”, which need to be determined on a case-by-case basis (paras 110-152).

Question 4: Does a chatbot provider reproduce a press publication?

For the fourth question, the AG proposes an analytical framework that distinguishes the technical acts involved in RAG from the content of the response presented to the user.

The AG states that, unlike LLM training, retrieving publications to generate chatbot responses can constitute an “online use” under Article 15 CDSMD. He considers that RAG involves technical reproductions which are not covered by the Article 4 TDM exception, because the process retrieves specific content to improve an answer rather than extracting patterns or other information through data analysis. Still, those copies may benefit from the temporary reproduction exception in Article 5(1) InfoSoc Directive. In the absence of authorisation, the AG considers that this depends on whether the chatbot lawfully accessed the publication and whether the resulting answer falls within the press publisher’s exclusive rights (paras 153-190).

In his proposed answer, relevant acts of reproduction occur where a chatbot retrieves and reuses a press publication and either:

• accesses the publication unlawfully, namely by circumventing effective technical measures restricting access or scraping (e.g. those under Article 6 InfoSoc Directive); or
• produces an answer in which the publication’s form of expression is recognisable to such an extent that the user can dispense entirely with accessing the original (paras 229 and 240).

These are alternative routes. The first concerns the lawfulness of the use enabled by temporary technical copies made during RAG, and thus the application of the temporary reproduction exception in Article 5(1) InfoSoc Directive. The second concerns whether the response itself reproduces the protected publication. Accordingly, lawful access to an article does not necessarily make its subsequent reproduction in a chatbot answer lawful (on the concept of lawful access, see also our cited ECS Opinion on the topic).

To assess whether an answer reproduces a press publication, the AG applies what he calls the Court’s general, technology neutral test: the publication must have been copied, and its protected elements must remain recognisable in the output. RAG establishes the necessary causal link, even without direct copying and pasting. Recognisability requires more than the repetition of facts or very short extracts: the publication’s form of expression (including its arrangement, structure and phrasing) must be recognisable, even if paraphrased. For the AG, the decisive question is whether the answer is a direct substitute for the original, allowing the user to dispense entirely with accessing it. An answer that reports the same facts without a recognisable form of expression does not reproduce the publication, even if it diverts users from the publisher’s website (paras 193-209).

The AG also addresses attribution of the conduct in question. He proposes that the relevant acts may be regarded as carried out by the chatbot provider where they are inherent in its business model, or where it has failed to implement measures expected of a reasonably diligent operator to counter them credibly and effectively. The fact that an individual user initiated the query does not automatically exclude the provider’s responsibility (paras 217-222 and 240).

The AG presents this approach as a balance between press publishers’ interests, chatbot providers’ freedom to innovate and users’ freedom to receive information. Chatbots should have room to lawfully retrieve content and reuse it in an unrecognisable form, potentially directing users to their sources. Conversely, services that circumvent effective access restrictions or recognisably reproduce protected expression may require authorisation. While acknowledging a grey area and the practical difficulty of ensuring lawful retrieval, the AG considers that provider liability should focus on the design of the service and appropriate safeguards, rather than occasional mistakes. He also recognises that copyright cannot resolve wider concerns about source diversity, misinformation and users’ reliance on AI-generated news (paras 233–238). This is a broader theme in the opinion, which at different stages stresses that intellectual property cannot provide a complete solution to the challenges posed by AI in this area.

The ECS proposed a negative answer to question 4 on the facts as presented in the reference, taking account of the limited scope of the press publishers’ right. It nevertheless acknowledged that publishers may have legitimate complaints against AI services that reproduce or communicate editorial content.

The difference between the two approaches may be less pronounced in the result than in the reasoning. The AG offers a more detailed framework under which RAG uses can engage Article 15. But, applying that framework provisionally to the example before the national court, he suggests that the conditions for liability do not appear to be met. Among other things, the article’s form of expression did not appear to be recognisable in the response, and Google appeared to have implemented measures to mitigate the risk of infringement. The final assessment remains for the referring court (para 239).

Brief note on the issue of territoriality

The AG addresses the territorial reach of EU copyright law in relation to AI model training carried out outside the EU. He distinguishes the law applicable to an infringement claim from the question whether that law protects against acts with a sufficient connection to the territory concerned. In his view, making a model available on the EU market after training does not, per se, provide an unproblematic basis for applying EU copyright law to training acts carried out abroad: any harmful effects in the EU are “merely the indirect (and usually diffuse) consequence of the acts carried out during training”.

That said, he does not exclude applying EU law to the scraping of protected material from servers located in the EU, even where subsequent training takes place elsewhere (echoing Peukert’s position). Given the complexity of the issue and the need for legal certainty, the AG considers a legislative solution that complies with the “limits of the principle of territoriality” preferable (paras 60–73; I have previously examined the topic in this paper, also cited in the Opinion). This is perhaps something to consider in the Commission’s upcoming targeted consultation on the review of the copyright acquis.

Final thoughts

The Opinion’s most important contribution may reside in its treatment of retrieval and outputs, rather than training. Although the AG advises the Court not to answer the training questions, he sets out an approach to the interaction between temporary technical copies, recognisable expression, substitution for the original publication and provider responsibility. He also concludes clearly that AI model training is mostly covered by the concept of TDM and the relevant exceptions, provided their conditions are met.

Several elements will merit close attention if the Court follows this approach. Regarding the admissible questions, these include the high threshold implied by a user being able to dispense entirely with the original publication, the role of systematic recognisability in making available, and the measures expected of a reasonably diligent chatbot provider.

For the ECS, the case called for caution because of the narrow right invoked and the uncertainty surrounding the technology. The AG shares some of that caution, particularly in distinguishing training from RAG and in his provisional assessment of the facts. At the same time, he proposes a more developed framework for future disputes. Whether the Court adopts it, and how it defines its limits, remains to be seen.

Finally, the Court’s judgment will have significant implications for the pending referrals in Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26), as well as for national cases such as those brought by GEMA against OpenAI and Suno in Germany.

Funding acknowledgement: João Pedro Quintais’s research was supported by the Dutch Research Council (NWO) through the Vidi project Generative AI Content Moderation: Regulation for Fundamental Rights (file number VI.Vidi.241R.016; grant DOI: 10.61686/HTKXT49425)