As a small introductory note, those who developed entire conspiracy theories following the ‘disappearance’ of the Opinion from the Curia Calendar (the release was originally scheduled 3 September) can now find peace knowing that, in fact, the case did not die while sitting in Luxembourg.
While the Opinion is lengthy (240 paragraphs), in essence the AG proposes that the CJEU declares inadmissible the questions regarding the training of the LLM in question (Google’s Gemini) and instead answers the questions concerning whether the display of responses by a chatbot may infringe the rights of press publishers.
In any event, the AG himself warned against reading too much into his or the Court’s future responses to the referred questions, given that the referral “concerns only a specific subset” of the ‘revolution’ brought about by genAI. The AG himself declared that his view would be only provisional, thus reserving the right to further shape it in future cases. Let’s not in fact forget that Like Company is the first but not the only referral currently pending before the CJEU on AI and copyright, the others being (for now) Manchester United Fanatics Klub, C-806/26 and Mobilguru, C-829/26.
The above said, let’s see in greater detail how the AG developed his analysis.
The press publishers’ right and the hypothetical character of the training questions
The AG began by detailing the ‘convoluted nature and scope’ of the press publishers’ right under Article 15 of the DSM Directive, before turning to explaining why the questions relating to the training of Gemini would be hypothetical and should be therefore declared inadmissible by the CJEU.
In this context, there are two key points that must be flagged: the first concerns the scope of the press publishers’ right, the second relates to where the training of an LLM takes place.
Scope of the press publishers’ right: online uses
Regarding the former, the AG pointed out that Article 15 only concerns ‘online uses’ of a press publication. In his view, LLM training is not a form of online use, thus rendering the right inapplicable.
While this position is in itself intriguing and deserving of closer scrutiny (with the AG noting that neither the referring court nor the interested parties paid sufficient attention to it), it should be noted that – unlike the press publishers’ right – copyright is not subject to such a limitation. Yet, the background national litigation does not appear to have any copyright angle, but only reliance – on the side of Like Company – on the related right for press publishers.
Place of training
Turning to the latter, the AG noted – correctly – that LLM training engages the EU right of reproduction, although it is necessary to distinguish between different stages thereof.
Importantly, the EU legislature takes a technical as opposed to a functional approach: whether the protected content is used as such or as data does not make a difference. The AG also developed important remarks regarding the treatment of memorization and regurgitation.
In all of this, a preliminary question must be tackled first: if the training takes place outside of the EU, can the EU right of reproduction be even enforced against non-EU AI developers for reproductions done outside the EU?
That, as readers know, is a thorny question which does not touch upon the principle of lex loci protectionis, but rather the understanding of territoriality.
The AG conceded that Google’s position – that is: that the applicable law is that of the place where the main computer infrastructure is located – would offer the advantage of simplicity. However, it would also ‘have a significant downside: EU artistic and literary property law could not offer rightholders in the European Union any protection when their works and other subject matter are used for the purposes of model training outside the Union.’
Nor did the AG consider the proposal to focus on the effects of the training on the EU to deem EU law applicable entirely convincing either, although he recalled case law of the Court itself (including referrals where he acted as appointed AG!) that paved the way for such an interpretation. That is so because, in the case of a trained LLM, the effects in question would be indirect, not direct as in those previous instances. He thus conceded that only applying such an approach to certain acts could make sense, albeit noting that ‘[s]uch a solution is not perfect’.
The AG ultimately called for legislative clarification on this point. As a coincidence, it should be noted that today is also the final day to submit responses to the European Commission’s Public Consultation, which include these very aspects too.
Applicability of exceptions
This part of the Opinion also tackles the applicability of exceptions to reproductions done for training purposes, noting that the text and data mining (TDM) provisions may not necessarily be the only applicable exceptions – at least to certain reproductions. In this context, the AG specifically referred to temporary copies under Article 5(1) of the InfoSoc Directive.
That said, he focused on Article 4(1) of the DSM Directive, noting that it is in principle applicable to LLM training and that it is historically untrue that AI was not considered a potential application of TDM back when the DSM Directive was finalized and adopted.
Consideration of the three-step test would not affect such a conclusion. Regarding the three-step test analysis, I limit myself to noting that this part of the Opinion alone deserves rivers of ink.
In all of this, the AG highlighted how the application of Article 4 requires lawful access to the protected content in question and that no rights reservation has been exercised by the concerned rightholder. Both aspects are tackled in detail by the AG and they also deserve (very) careful attention since they go to the heart of some of the most contentious issues, including how the rights reservation needs to be done to be enforceable.
The AG concluded his take on this part by raising a policy argument too: he considered that Article 4 could become a compensated limitation. This, again, is something that the also features in the European Commission’s Public Consultation.
Whether a chatbot’s response may infringe the press publishers’ right
Coming to the questions that he considered the CJEU should address, the AG noted at the outset that a chatbot – like AI Overviews and AI Mode on Google Search – does not only operate on the basis of an LLM but also relies on RAG technology.
If press publications are used by a chatbot when providing answers, then this is an online use that is potentially relevant under Article 15 of the DSM Directive. Importantly, in this context, no territoriality issues (as instead in the consideration of training activities) would subsist: the answers are provided directly to EU-based users.
Also key:
- It is not a given that such uses of press publications would be covered by the right’s exclusions, including the one relating to ‘mere facts’ and those concerning ‘very short extracts’ and ‘individual words’.
- A general argument cannot be convincingly made that press publishers have implicitly consented to the use of their press publications by simply placing them online.
- Article 4 of the DSM Directive does not apply (though other exceptions in principle might, e.g., Article 5(1) of the InfoSoc Directive) given that the use of a press publication by a chatbot is not for the purpose of TDM.
As to the user-facing part, the AG considered that the display of an answer by a chatbot might, depending on the circumstances, engage the right of reproduction and making available to the public. Again correctly, the AG recalled that the press publishers’ right is informed by an investment logic, so the relevant threshold to enforceability is, accordingly, investment-based. That said, the AG also noted that ‘[a]n act of making available to the public of the press publication at issue occur if the form of expression in all or part of that publication is systematically recognisable.’ As previously discussed [IPKat here], it may be questionable whether recognizability is the correct or even a workable criterion.
It is in any event the case that the chatbot provider would be responsible for its own acts. That does not mean that no responsibility whatsoever would subsist regarding users’ own acts, but the test to apply would be that developed by the CJEU itself (notably YouTube, C-682/18 and C-683/18 [IPKat here]).
The AG concluded by considering that the proposed interpretation would ensure a fair – yet ‘delicate’ – balance of different rights and interests.
Comment
This is an Opinion that cannot be called otherwise than landmark – it is mandatory reading for anyone interested in AI and copyright.
Being a landmark analysis, there is a lot to unpack too, which cannot be covered in the space of a rapid-response blog post. It is also remarkable that some of the ‘heaviest’ remarks and conclusions offered relate to questions that, according to AG Szpunar, the Court should not even answer. Should the CJEU indeed declare such questions inadmissible, it is in any event expected that it will need to weigh on the same or similar issues sooner or later, even in the context of the other pending AI/copyright referrals or new referrals from national courts.
Let’s now wait for the judgment. Stay tuned!
[Originally published on The IPKat on 3 November 2026]