In-depth view on how EU copyright law applies to gen-AI offered by CJEU adviser

While advocate general Maciej Szpunar said the training of large language models (LLMs) on copyright works would generally constitute copyright infringement, he said that activity can fall within the text and data mining (TDM) exception written into EU copyright law.

Szpunar said whether the TDM could be relied upon for LLM training purposes should be assessed on a case-by-case basis and further clarified that AI developers would definitely need a licence to use copyright works in that training if the relevant rights holders explicitly exercised their rights to opt out their works from being used for TDM in a machine‑readable format.

Szpunar addressed questions of how EU copyright law applies to AI training in a new non-binding opinion issued for the Court of Justice of the EU (CJEU). He did so despite considering that the matters he expressed a view on are out of scope from what the CJEU has been asked to determine in the underlying case before it.

The CJEU case centres on a Hungarian publishers’ claim to press publisher rights in the EU and its assertion that those rights have been infringed by Google in respect of the output produced by its AI chatbot, Gemini.

The press publisher’s rights are a right EU-based publishers of press publications enjoy that supplement the rights they have under EU copyright law, providing them with the exclusive right to authorise or prohibit the reproduction and online communication to the public of their copyright works for commercial purposes, subject to certain limitations and an exception that enables others to reproduce and extract from lawfully accessible works and other subject matter for the purposes of TDM – including in the commercial context.

To help it rule on the claims brought by Like Company Kft. against Google, a Hungarian court has asked the CJEU to help it interpret the application of EU law in the gen-AI context.

Two of the questions posed to the CJEU concern use of press publications to train an LLM. Szpunar said the CJEU should “leave aside” those questions, describing them as “hypothetical” to the underlying case. This is because Szpunar considered there to be nothing to support the assumption that Like Company’s articles were used to train the Gemini LLMs. He also said LLM training does not constitute an online use of a press publication, meaning Like Company could not assert press publisher rights in respect of that activity.

However, to assist the court in its adjudication of the case, Szpunar examined the hypothetical matters in detail, setting out his own provisional views on a raft of issues that go to the heart of whether, and the extent to which, EU copyright law is engaged by the way gen-AI systems operate.

Territorial scope of EU copyright law

Among other things, Szpunar considered whether, and if so when, EU copyright law can be said to apply to training of LLMs that takes place outside of the EU. He acknowledged different views on the issue.

Google, for instance, argues that only the laws of the country where the main computer infrastructure used during LLM training is located should regulate the extent to which use of copyright works in that LLM training constitutes copyright infringement. In contrast, a number of governments of EU countries advocate for an interpretation of EU copyright law that would confer protection where an LLM model trained outside of the EU is, after deployment, then made available in the EU. Szpunar said that “it would be preferable if the legislature expressly addressed” the issue of the territorial scope of EU copyright protection “in a clear manner”.

Whether LLMs store or memorise data they are trained on

Szpunar said using copyright protected works for the purposes of LLM training would, in his view, involve acts of reproduction that could engage the rights of copyright owners. However, in going on to consider whether the trained LLMs reproduce those works in the models themselves, Szpunar expressed scepticism, challenging the idea that in cases where the copyright works are replicated in their entirety in the output an LLM generates it is because the LLM stores or memorises the data on which it has been trained.

His views on the issue are at odds with those of the Regional Court of Munich, which endorsed the concept of memorisation in respect of how gen-AI systems operate in recent rulings in which it found AI providers liable for copyright infringement.

“An LLM is normally designed not to ‘memorise’ and ‘recite’ the texts on which it was trained, but to ‘generalise’, that is to say generate new text on the basis of the statistical patterns and relationships it ‘learned’ during training,” Szpunar said.

“According to the available literature on the matter, ‘memorisation’ occurs when the model, during the automated machine learning process, unintendedly ‘learns’ not only abstract language patterns, but also patterns which are specific to a given piece of text in its training dataset. This seems to be the result of a phenomenon called ‘overfitting’: when a piece of text is duplicated many times in the training dataset (because, for instance, it consists in the lyrics of a famous song, scraped multiple times from the World Wide Web) and, accordingly, the model is overly exposed to the text in question, it may ‘learn’, within its parameters, the specific sequence of tokens corresponding to that text. When the beginning of that text is prompted by a user, the model may recite the rest because it will appear to be the probable, statistical answer in the light of its parameters. Such ‘memorisation’ appears incidental to the machine learning process, and not entirely avoidable,” he said.

“That being clarified, it seems to me that, on the one hand, where an LLM generates, upon a user prompt, a text which is identical, or near-identical, to (part of) a protected work on which it was trained, as the case may be because the model has ‘memorised’ and ‘regurgitates’ it, such an output may, in itself, constitute a reproduction of that work and, thus, infringe the exclusive right of the author concerned. That should be determined on a case-by-case basis,” the advocate general added.

Application of the TDM exception

Having considered use of copyright works for the purposes of LLM training to involve the making of reproductions, Szpunar then examined whether that activity would fall under the TDM exception. His view is that the TDM exception does apply to such reproductions “in principle”, citing support for this in the wording of EU copyright law and in international conventions that provide the basis for law makers to apply copyright exceptions.

Szpunar said the regulation of LLM training involves a balancing of rights and interests and that it is his view that EU law makers were “at liberty to tip the scales more in favour of innovation and, through the TDM exception, relieve model developers from most of the burden associated with the ‘clearing’ of copyright and related rights”. 

However, Szpunar said that there are circumstances in which AI providers may not be able to rely on the TDM exception when using copyright works to train LLMs without a licence. This would include “where such use … would unreasonably prejudice” the legitimate interests of rights holders, he said.

Szpunar said a case-by-case assessment would be necessary to determine if the TDM exception could be relied upon and cited factors that would be relevant to such an assessment. For example, he implied that LLM developers would need to meet transparency obligations arising under the EU AI Act, in relation to the works they train their models on, to rely on the TDM exception. He also said that an “important element” was the extent to which rights holders are compensated for use falling subject to an exception. In this regard, he said there is nothing to stop EU countries legislating to require companies that rely on the TDM exception to compensate creators of the content they make use of under that exception.

The press publishers’ rights

Having examined the hypothetical matters, Szpunar turned his attention to addressing the questions the Hungarian court posed that go to the heart of the dispute between Like Company and Google.

Szpunar confirmed that the press publisher’s right only applies to the way facts are expressed in press publications and not to the facts themselves.

While he considered that LLM training does not constitute an online use of a press publication, he drew a distinction between that training process and the process by which chatbots retrieve information in real-time from the internet in response to user prompts. Szpunar said the ‘Retrieval-Augmented Generation’ (RAG) technology used may involve the making of technical reproductions of press publications but that those reproductions can fall within the temporary reproduction exception under EU copyright law.

In relation to chatbot output featuring reproductions of press publishers’ content, Szpunar said that “online chatbots which lawfully retrieve content online and reuse it in an ‘unrecognisable’ way to provide information on a given topic should be able to operate freely”.

If the answers displayed are “recognisable”, or if the chatbot has circumvented a paywall to retrieve the content, those are factors that could result in the chatbot provider being liable for infringing press publishers’ rights, according to Szpunar. In such circumstances, however, providers would have a defence if they could show they had put in “appropriate measures that can be expected from a reasonably diligent operator in its situation in order credibly and effectively to counter the carrying out of such acts by way of the chatbot”, he said.

What happens next?

The CJEU judges can deviate from the opinions issued by advocates general to the court. Those opinions are designed to help the judges in their analysis of the relevant law and reach their own interpretations of it. The judges often follow the opinions, but not always. The ruling in this case is not expected until some time in the middle of 2027.