The much-awaited Opinion by the Advocate General in the landmark CJEU case on Like Company v Google has now been released, and it’s a quite something, filled with very interesting analysis. We will undoubtedly spend the next weeks unpacking a very rich text that is considerably longer than usual, as AG Szpunar points out in the opinion, but I wanted to write a quick reaction to a few parts of the opinion that caught my eye. This is by no means a comprehensive breakdown, there are already very good analyses out there, but I want to highlight a few issues that I think are very important going forward.
But first a short introduction to the case if you haven’t been following it, and you may be wondering about the dolphins. In July 2023, Like Company, a Hungarian publisher that runs several regional web portals, published a short article reporting that the singer Kozsó still hoped to introduce freshwater dolphins into Lake Balaton. Some time later, someone from the publisher asked Google’s chatbot, to summarise that article in Hungarian, and it duly obliged, adding a few details of its own along the way. Like Company sued Google Ireland before the Budapest Regional Court, arguing that its press publishers’ right under Article 15 of the DSM Directive had been infringed, both through the use of its articles to train Gemini and through the chatbot’s answers. In April 2025 the court referred four questions to the CJEU, asking whether chatbot outputs reproduce and make available press publications, whether LLM training involves reproduction, and whether any such reproduction is covered by the text and data mining exception. This made Like Company v Google the first generative AI case to reach the Court of Justice.
AI training
One of the first issues handled by AG Szpunar was the fact that of AI training, as it was an important part of two of the referred questions, and here he found that there was no evidence that the article in question was in the training data for the model, as the dates do not match. This is something that we found in our article on the subject, the knowledge cutoff point for the Gemini model in question was February 2023, while the article was published in July of that year. The article could not have been in the training data, and AG Szpunar agreed. This meant that two of the questions dealing with AI training should be dismissed.
The fact that the training questions should be dismissed did not stop the AG from providing an opinion on the underlying issue, knowing quite well that this will be a contentious question that will certainly make its way to the CJEU again. One of the main takeaways for me is that he is clear that AI training falls squarely under the existing TDM exceptions in Arts 3 and 4 of the DSM Directive. The discussion is centred on Art 4 as it is the most contentious, but I think that this covers both. Here are the highlights of the discussion in paras 97–109 (edits mine):
“98. Whether that exception is applicable to the reproductions of works and other subject matter carried out in relation to LLM training is a seminal issue in EU artistic and literary property law. […]
99. I tend to think that the TDM exception does, in principle, apply to the reproductions of protected works and other subject matter carried out for the purposes of LLM training. This stems, in my view, from the wording, context and purpose of that exception (1). Furthermore, contrary to what is put forward by some of the governments who submitted observations, it seems to me that such an interpretation complies with the ‘three-step test’. […]
108. […] the argument that the legislature did not envisage LLM training at the time of the adoption of the CDSM Directive is not entirely accurate. To give one example, a study commissioned, at the time, by the JURI committee of the European Parliament on text and data mining expressly ‘flagged’ AI and machine learning, indicating that that would acquire ‘increasing centrality’ and that, ‘although classical TDM and machine learning have different utility, it should not be overlooked that both use the same key algorithms to discover patterns in data’.
109. Thirdly, and in any event, as Google and the German, Estonian, French and Hungarian Governments state, whatever the EU legislature had in mind at the time of the adoption of the CDSM Directive, it has clearly expressed its intent in the AI Act in the meantime. It is clear from recital 105 and Article 53(1)(c) of the AI Act that, in the EU legislature’s view, the TDM exception applies to the use of protected works and other subject matter for the purposes of LLM training”
A lot of energy has been spent on pushing the narrative that TDM does not cover AI training, an opinion that I have strongly disagreed with at any opportunity I can get. I have to admit that I have been surprised that this idea endured this long, the DSM wording is broad on purpose, it is evidently a technology-neutral definition. Any doubt that remained should had been destroyed by the AI Act, and yet it prevailed. Perhaps we can now let it rest? Sure, the exception still has to meet the three-step-test, but national European courts have been finding that it does. AI training still has to respect reservation of rights.
Memorisation
Another of the most important issues in the Opinion is memorisation, and specifically whether a model contains copies of the works that it was trained on. Needless to say, this matters a great deal for copyright. If models are copies, then every model in existence is potentially an infringing copy of every one of the billions of works in its training data, which would be a rather awkward conclusion for an entire industry. The reality is, as I have argued countless times is that models are not copies. To put it simply, models are sets of parameters which encode statistical relationships between tokens, these are used to predict what is likely to come next. Some of those relationships can be strong enough to reproduce a passage when prompted in the right way, but that is not the same as the work sitting inside the model waiting to be retrieved.
The reproduction of works in the training data is known as memorisation, and there is incontrovertible proof that it can occur, models can indeed remember things that were in their training data verbatim. But this doesn’t always occur, and AI companies work hard to minimise it. The reality is that we do not know exactly when something has been memorised. The discussion on memorisation for me is one of the best parts of the Opinion (paras 84-92). Here is a quotation edited by me with the main highlights:
“[A] trained LLM is not a database. It is not a repository of (parts of) works or other subject matter that the model accesses and ‘collates’ to produce an output. […] Rather […] the model integrates information extracted from that dataset, namely individual tokens (numerically represented) and the statistical patterns and relationships connecting them. […] Such abstraction representation of words and language patterns could hardly qualify as reproductions of the works and other subject matter in question: copyright and the related right do not subsist at that level of abstraction. […]
[O]ne should not confuse what seems to [be] a glitch with the overall functioning of an LLM. An LLM is normally designed not to ‘memorise’ and ‘recite’ the texts on which it was trained, but to ‘generalise’ […]. [O]ne should not infer from the fact that an LLM may generate reproductions of works on which it was trained when prompted appropriately that such reproductions are stored inside the model. […] [I]t is more a complex recalling of the interrelationships between words, which resembles, to a certain extent, what happens in the human brain. […]
“[P]rompting an LLM is not like opening a .pdf file with a PDF reader. […] [I]t is the combination of the model’s parameters and of the user’s prompt that may lead to the model generating reproductions of its training dataset. Thus, those parameters, seen in isolation, are merely preparatory material for the potential generation of such reproductions by the LLM. […] [R]eproductions of protected works and other subject matter generated by the model as outputs are, in themselves, actionable.”
Paragraph 92 is particularly interesting, it points out something I have repeatedly said about memorisation: as you cannot possibly know what will be memorised, the proof is in the outputs. Wait. Is AG Szpunar a Technollama reader? [waves].
Territoriality
As I have written before, one of the things that worried me most from the first hearing was the issue of territoriality, and in particular the suggestion that training and output could be treated as one continuous chain of events. I’m happy to report that the Opinion keeps them firmly separate.
The first thing to note is that the AG does not think training falls under the press publishers’ right at all. Article 15 only covers “online use”, which he defines as a use that “occurs as an integral part of the provision of an information society service via the internet” (para 55). Training does not fit, because it “occurs as part of the development of a piece of AI technology on dedicated computer infrastructure, without any information society service being provided to a user via the internet” (para 56). Even when the model ends up in a chatbot, the link between training and what the chatbot later tells its users is “simply too indirect (and, usually, diffuse)” (para 57).
Nevertheless, he goes on to discuss the wider question of whether EU copyright covers training that takes place abroad, as in this case, training happened in the US. Applicable law is not the problem, it is evident that Hungarian law applies under Rome II because that is where protection is sought. The question is whether that law can “confer protection against acts of use of works and other subject matter carried out for the purposes of training an LLM in a third country” (para 62). Google argued that only the country where the main computing infrastructure sits should regulate training. The AG concedes this offers legal certainty, but notes it would leave EU rightholders without any protection for training outside the Union (para 65). Most of the governments argued the opposite, namely that EU law should apply whenever the model is later made available in the internal market, because that is where the harm is felt. The AG was not persuaded, extending EU law to training “anywhere in the world, on the basis of those effects alone” would raise problems with territoriality and legal certainty (para 67), and any harm is “merely the indirect (and usually diffuse) consequence of the acts carried out during training” (para 69).
He is more sympathetic to the German Government’s middle ground, under which EU law would catch the scraping of works from servers located in the EU, wherever the training later happens. He would not exclude this interpretation (para 71), and suggests that works “should not be transferred outside the European Union for training purposes where the rightholders oppose such a use”. He admits the solution is not perfect (para 72), since server location is often unknown in the cloud. His final word is that “it would be preferable if the legislature expressly addressed that issue in a clear manner” (para 73).
I really like this approach, minus the server issue. If there is something that we have learned from years of Internet jurisdiction case law is that server location tests are extremely problematic, particularly as the AG says, because there is a substantial issue with distribution. I would also add to this that as of today, there is a disproportionate concentration of training compute power in the US. European training is still negligible.
Outputs are a different story. A chatbot provided to users within the EU can be regulated “irrespective of the place where the chatbot provider is established or of the location of the servers from which it operates” (para 165), because there the effects are “immediate and substantial” (para 166). In other words, the input/output distinction still matters, and as I keep saying, the proof is in the output. Inputs are hard to reach territorially, outputs are not. That is broadly what the UK court found in Getty v Stability AI, and it is what I’ve been arguing for some time.
Bits and pieces
I won’t even bother covering the press publishers’ rights and communication to the public, and RAG. Colleagues have already been doing a good job on that, and there will be time to tackle those issues here in the future (or whenever I write a post-mortem to the final decision). There are a couple of things I wanted to point out.
While I largely agree with the Opinion, there are a couple of things I would like to push back on. The AG finds that training does involve reproduction, taking a technical rather than functional approach and rejecting the argument that using works as data is akin to a human learning to write by reading. In his view, that comparison fails “if only because of the industrial scale of the process” (para 83). It’s a bit ironic that a few paragraphs later, he describes what happens inside the model as resembling “to a certain extent, what happens in the human brain” (para 90). So the human analogy works for the model, but not for training. I also dislike human analogies when it comes to AI for the most part, but sometimes they can be very useful. We use them all the time!
The three-step test analysis is also very interesting. AG Szpunar rejects the market dilution argument as too diffuse to count, but adds a condition, and that is that developers relying on the TDM exception should demonstrate that they have taken measures to mitigate the risk of outputs that closely resemble their training data. If they cannot, the national court “should refuse the application of the TDM exception” (para 133). In effect, good practice under the AI Act’s Code of Practice becomes a condition of the exception. You can read that as sensible housekeeping or as a backdoor, depending on where you sit.
Concluding
I thoroughly enjoyed this Opinion, even if it sits at an astounding 240 paragraphs and 243 footnotes. I even made a couple of footnotes, which is a first, but I digress. I have been vocal in my view that this case is a bit bizarre in the sense that the originating article is quite short, and the alleged infringing output does not reproduce any part of it. Add to that the fact that the article was not on the training data, and you can see why one would very likely consider that Google will likely win the case in the national court. This despite some press clippings out there that contain headlines such as “Setback for Google”. Did we read the same Opinion?
I have made my thoughts about this case clear. When I first reported on the lawsuit and the referral, I was shocked that most people had not even bothered to visit the article and the website. If they had, they would probably share my dismissal.
But I am glad that the case continued, as we got a fantastically rich Opinion by AG Szpunar, one that such a complex subject deserves. I do not agree with everything, but the text clearly gets the technology right, and it is evident that a lot of care has gone into getting the details right.
On the meantime, I’ll just point out that dolphins in Lake Balaton remain unlikely. So, it seems, do copies inside the model.
0 Comments