by Raffaella Aghemo
With the advent of generative Artificial Intelligence, there is an ongoing debate about how to protect creativity.
Alongside the issue of recognizing authorship in a future increasingly at the mercy of “pirate” training, which appropriates creative experiences without any respect or restraint, there is a further question, namely the importance, on the part of companies that use “creative” AI, to protect their intellectual property rights through patentability, but also and above all through policies that cover trade secrets through contractual confidentiality clauses. Although the problem is a general one, the approach to the situation varies depending on the country in which we find ourselves, and above all it relates both to the protection of AI products and to the TDM regime.
Trade secrets and policies to protect AI
Trade secrets protect “secret” information. They are a useful tool for protecting artificial intelligence technologies. The definition of “trade secret” under UK law is broad and can encompass a wide range of information, including source code, training data, model weights, and proprietary algorithms. Trade secret protection is not limited in time and can, in theory, last indefinitely, provided that the information remains secret and continues to have commercial value.
Protecting AI through trade secrets offers a major advantage in that it does not require the information to be made public, unlike patents. This method is particularly useful in a rapidly developing industry, where patented technologies can quickly become obsolete. Trade secrets effectively protect the know-how essential for design, data collection, algorithms, and source code, and contracts can be quickly updated to include new technologies. However, relying on trade secrets comes with challenges: companies must demonstrate that they have taken concrete measures to protect information and ensure the enforcement of NDAs and contractual clauses, for example through staff training. To obtain compensation in the event of infringement, it is necessary to prove that the information is not public and that it has real commercial value.
The treatment is different in the United States, where the difference between trade secrets and protection through copyright or patent is that ownership of a trade secret does not derive from the way the information was created, but rather from the legitimate ownership of the information. Therefore, the use of trade secrets circumvents issues related to their ownership, a problem that could arise if protection through patent or copyright were sought.
In Italy, a company’s intangible assets, mainly know-how, are extremely important for efficiency and competitiveness, as well as being a significant part of the company’s assets, which can be protected as trade secrets: the greatest economic investment in protecting these assets is in maintaining adequate security measures to protect the confidentiality of information. Directive (EU) 2016/943, known as the “Trade Secrets Directive,” on the protection of know-how and trade secrets against their unlawful acquisition, use, and disclosure provides, in Article 2(1), that information eligible for protection must be secret and have economic value as such, in that it is subject to reasonable measures to keep it unknown to the public (requirements transposed into the Industrial Property Code in Article 98).
Pursuant to Article 3 of the Trade Secrets Directive and Article 99 of the Italian Industrial Property Code, the prohibition on the misappropriation of trade secrets is subject to certain exceptions, such as independent discovery or, in certain cases, reverse engineering. However, the extreme complexity of artificial intelligence systems, in terms of programming and data selection for training and validation, makes this possibility quite remote. All this while providing, in accordance with the AI Act, specific rules on the need for human supervision in the operation of artificial intelligence systems, compatible with the necessary requirements of transparency (Trustworthy AI pursuant to Articles 1, 13, and 50 of the AI Act) and intelligibility of the output, training data, and decision-making processes that lead to the result. The regulatory framework therefore calls for a fair and proper balance between the essential values of secrecy and confidentiality of proprietary company data and the necessary transparency and explainability of the output, with an obligation to disclose, even if only to the competent authorities, if the ‘decision’ of the AI system can be said to be discriminatory and/or prejudicial to the rights of the recipient.
Different regimes for text data mining
In the United Kingdom, the Copyright, Designs and Patents Act 1988 allows copyrighted works to be used for text and data mining for non-commercial purposes, provided that the user has legitimate access to them, for example through a subscription or license.
Data mining refers to the use of automated tools to analyze large amounts of data, and may include the extraction of information from copyrighted works.
However, many rights holders argue that artificial intelligence systems are trained using protected works without their consent, raising issues of copyright infringement. That is why, in a consultation in December 2024, the UK government reported that, in its view, changes to copyright law are needed in the context of artificial intelligence, a need that could lead to a decision to amend the current TDM exception to allow its use for commercial purposes, but only in cases where the rights holder has not waived their consent to such use. These changes would be made through legislation rather than a code of conduct, to provide welcome clarity and greater certainty.
English courts are considering the issue of artificial intelligence and copyright, in which a company has filed claims for copyright, trademark, and database rights infringement against a UK-based company. The High Court has yet to rule. Two original issues were brought to the court’s attention: the first concerned the primary copyright infringement by the artificial intelligence company, which had copied its works during the training and development phase of its artificial intelligence tool. Let’s take a small step back: primary copyright infringement concerns unauthorized copying or communication to the public, while secondary copyright infringement concerns secondary acts such as sale, distribution, importation, and marketing. However, the part of the claim relating to primary infringement has been withdrawn and therefore the High Court’s decision will not rule on this issue, although the secondary infringement remains, the outcome of which could set an important precedent if other rights holders wish to successfully claim copyright against AI developers in the UK. Nevertheless, the decision provides important clarification for the future, both for copyright holders and for developers of artificial intelligence technologies.
The appellant abandoned these claims because the evidence presented related to alleged acts of infringement that took place outside the jurisdiction of the United Kingdom. Since UK copyright law does not have extraterritorial application, in order to invoke its protection, at least part of the acts of infringement must take place within UK territory. It should also be noted that the claimant’s decision to abandon its two main claims highlights the difficulties faced by claimants seeking to prove copyright infringement in relation to artificial intelligence resources.
Extraterritoriality creates a significant obstacle for rights holders seeking to enforce their claims under UK law, especially considering that AI systems are often trained outside the country.
As we know, the US approach is quite different, where the doctrine of ‘fair use’ covers many more cases than in any other jurisdiction in the world.
In the United States, courts assess on a case-by-case basis whether the use of copyrighted works for training artificial intelligence tools can be considered “fair use.” Compared to the United Kingdom, American case law has been more active in addressing this issue. So far, three US district courts have ruled on this issue, analyzing the possibility of invoking the fair use defense in relation to the training of artificial intelligence models.
In the first of these cases, the decision concerned a non-generative AI model and ruled out the applicability of the defense, mainly due to the commercial nature of the use and the lack of transformation of the result produced by the model. The other two rulings, on the other hand, examined cases relating to generative AI models, recognizing that in such circumstances the fair use defense can be invoked to justify training.
In the United States, the Supreme Court has ruled on several issues concerning the relationship between copyright and artificial intelligence, establishing, on one occasion, that a work created entirely by an artificial intelligence system cannot enjoy copyright protection, as protection can only be recognized for works that include a significant human contribution.
Many regulatory rulings focus on these issues, such as the AI Accountability and Personal Data Protection Act and the TRAIN Act: the former bill aims to prohibit the use of personal data or copyrighted works for training AI models without authorization, also providing for the possibility of bringing federal action in the event of misuse of such materials; the TRAIN Act, on the other hand, would introduce an administrative procedure that would allow copyright holders to request district court officials to issue subpoenas to AI developers in order to obtain the disclosure of copies or recordings useful for identifying with certainty the protected works used for training.
The Italian approach is different, where the newly enacted Artificial Intelligence Law, 132/2025, establishes a further approach: the law inserts a new article into copyright law (Article 70-septies) which explicitly states that ‘reproductions and extractions’ from works or materials to which one has legitimate access are permitted if carried out using artificial intelligence systems (including generative AI), in accordance with the provisions of Articles 70-ter and 70-quater, without prejudice to the provisions of the Berne Convention for the Protection of Literary and Artistic Works, ratified and implemented pursuant to Law №399 of June 20, 1978.
However, the text leaves uncertainty as to the scope (full training vs. analytical extraction), without completely resolving the question of whether the act of training (massive ingestion and memorization in the model parameters) automatically falls within the notion of extraction/reproduction covered by the exception. Implementing decrees will probably address and explore this issue in greater depth.
But Italian law goes further, introducing amendments to the criminal code and copyright code providing for penalties for illegal reproductions or extractions in violation of TDM rules. In the first case, the new Article 612-quater on the unlawful dissemination of content generated or altered by artificial intelligence systems provides that: “Anyone who causes unjust damage to a person by transferring, publishing, or otherwise disseminating, without their consent, images, videos, or voices that have been falsified or altered through the use of artificial intelligence systems and are likely to mislead as to their authenticity, shall be punished with imprisonment for a term of between one and five years.” In the second case, referring to Article 2637 of the Civil Code, to which the following sentence has been added: “The penalty shall be imprisonment for a term of between two and seven years if the offense is committed through the use of artificial intelligence systems.”
The practical effect is a tightening of deterrence against cyber abuse/unauthorized use of content for training purposes, with two significant effects, the first on developers and the second on copyright holders: in the first case, the law states that TDM operations carried out with AI are formally covered, but before considering training on any material to be ‘free’, it is advisable to check, first and foremost, whether the material is legally accessible, secondly whether the owner has exercised an opt-out, and finally whether the documentation and transparency requirements have been met. For content owners, the law strengthens protective measures, with penalties for unlawful use, but at the same time formalizes the TDM exception, for which the right may be exercised through opt-out.
This results in an overview that, although consistent, is fundamentally different depending on the country in which the case is dealt with, making it very difficult to effectively pursue any interests, whether on the part of the recipient user or the owner company, as the pervasiveness of AI systems makes it very difficult to place them within a well-defined geographical perimeter.