Italian Law No. 132 of September 23, 2025 (which came into force on October 10) - the “AI law” - contains provisions to define the regulatory framework chosen by the national legislator on artificial intelligence. It is undoubtedly “the first comprehensive law of a Member State on AI” (as stated in the explanatory report), although it should not be forgotten that the European Union issued the AI Act (EU Regulation 2024/1689) in June of last year.
In this context, it seems not entirely unfounded to ask whether the national provisions are fully compatible with European legislation, which remains hierarchically superior. Or whether, alternatively, they can be considered compliant only under a restrictive interpretation.
A reinterpretation of domestic rules is also prompted by the recent decision in the GEMA/OpenAI case (Munich Tribunal, November 11, 2025), one of the first to qualify, in light of the legislation introduced in European countries, activities for training AIGen.
At the national level, for example, with regard to copyright, the definition of protected works has been amended to include works of human creativity “even when created with the aid of artificial intelligence tools, provided that they are the result of the author’s intellectual work” (Article 1 of the Italian Copyright Law).
Anyway, the amendment does not appear to introduce any substantial change.
It is undisputed that the requirement of human creativity is not negated by the mere fact that the author uses auxiliary tools: just as there has never been any doubt that a figurative work obtained using a chisel, a literary work written with the support of a grammar correction program, or a photograph reworked in post-production are protectable, it is equally clear that the distinction between ‘intellectual work’ and ‘non-creative material’ does not lie in the complete absence of any aid, but in the extent of creativity that remains with the artist in using it.
On this issue, we can think of the extensive debate, even overseas, in relation to techniques such as dripping, of which Jackson Pollock was a master.
For these reasons, the amendment to Article 1 of the Italian Copyright Law does not seem suitable for automatically bringing the output generated by generative AI systems within the scope of protection.
The new exception on TDM expressly referring to AI raises further doubts, an issue also addressed by the aforementioned ruling of the Munich Tribunal of November 11, 2025.
The ‘AI law’ in Italy has introduced a new provision into copyright law concerning exceptions for text and data mining (TDM) and generative Artificial Intelligence. This is Article 70 septies of the Italian Copyright Law, which allows the TDM exception to be applied to reproductions and extractions (from works or other materials contained on the internet or in databases to which one has legitimate access) carried out “for the purpose of extracting text and data through artificial intelligence models and systems, including generative ones”.
First of all, it can be noted that the new Italian provision expressly admits that extraction (of text and data) can also be carried out through Artificial Intelligence models and systems, including generative ones. In this meaning (and so interpreted) the new Article 70 septies appears to be fully in line with European legislation, given that the Digital Copyright Directive (Directive 2019/790), in introducing the TDM exception, adopted the principle of technological neutrality, i.e., not limiting the reproduction and extraction of text and data to specific tools. In other words, it did not say that reproduction can only be carried out under the exception with a certain scraping tool, cloud platform, or specific software.
However, it is legitimate to ask whether the express reference to generative AI systems aims, at least in its intentions, to include any Large Language Models (LLMs) training operations within the scope of the TDM exception.
The wording of the provision itself would seem to argue against this, as it limits the purpose of reproduction and extraction to ‘text and data extraction’ without mentioning model training.
However, it could be argued that training is in fact reducible to a mere activity of information extraction and analysis.
The Munich Regional Tribunal addressed this point in the GEMA/OpenAI case, analyzing the characteristics of the various acts of reproduction that take place in the different stages of preparation (pre-training - stage 1) and training (training - stage 2) of an AI system, prior to its use (in this case, by a chatbot).
The German Tribunal began by observing that certain preparatory acts (phase 1) are necessary to implement an AI system and may affect intellectual property rights, such as the reproduction of a work by transposing it into a different (digital) format or saving it in working memory. These reproductions are, in principle, lawful (provided, of course, that they come from sources to which there is a legitimate access) insofar as the copies “are created solely for subsequent analysis purposes and, as such, do not prejudice the author’s interests in exploiting the work.”
Articles 70 ter and 70 quater of our copyright law, issued in implementation of the Digital Copyright Directive (Directive 2019/790), expressly refer to this type of reproduction, allowing it to the extent that it is carried out “for the purposes of text and data mining” (i.e., precisely to benefit from the TDM exception).
However, it may happen that in the subsequent LLM training phase (phase 2), the system does not merely extract information from the training data, but reproduces works, or portions thereof, that are creative. This operation, the judges specify, “does not constitute text and data mining.”
The distinction is based on a finding of fact: it was inferred from observation of the output results of the OpenAI system that, during the training process, the model did not merely extract statistical regularities from the dataset, but retained portions of structured content attributable to the original works, making it possible to reproduce them in the output.
Where this occurs, the German Tribunal emphasized, the act of reproduction is relevant under the InfoSoc Directive (Directive 2001/29) and does not fall within the scope of the TDM exception introduced by the Digital Copyright Directive (Directive 2019/790).
In light of the InfoSoc and Digital Copyright Directives, therefore, the new Article 70 septies of the Italian law seems to confirm that deep learning activities through deep neural networks may fall within the scope of the TDM exceptions to the extent that:
- they analyze text, images, video, or audio,
- they automatically extract patterns, characteristics, relationships, and structures,
- and they transform them into numerical representations useful for learning.
However, acts of reproduction aimed at incorporating creative portions of the work into the model remain excluded from the scope of the TDM exception.
Such acts are subject to authorization by the rights holder, without the TDM exception being applicable - the German Tribunal specifies.
In its recent decision, the Tribunal focused on the act of making works used in the pre-training phase publicly accessible—whether in whole, in part, or in adapted form—through the outputs of a generative language assistant (chatbot). We will reserve a future contribution for the prospects of the German decision, which are particularly interesting even beyond the scope of AIGen.
With regard to the relationship between AI training activities and the TDM exception in Europe, this first precedent in Europe highlights the distinction between acts of reproduction of ‘creative portions of works’ and analysis of ‘information’ that can be extracted from a work, bringing the classic issue of the requirements of creativity and originality of the work back to the forefront.
It therefore seems reasonable to conclude that, also under Italian AI Law, the development of AI systems can be fully ensured in compliance with the principles — including those of balance — that underpin copyright protection.
(Beatrice Cunegatti)
Articoli Correlati