Introduction
In today's digital age, artificial intelligence (AI) is revolutionizing numerous sectors, while simultaneously raising new and complex legal challenges, particularly concerning the protection of AI software and the ascertainment of infringements. This article aims to explore strategies for protecting artificial intelligence software, with a specific focus on the European system and, crucially, the pivotal role that Italian jurisdiction and the "Descrizione" (description) tool can play in gathering evidence in cases of counterfeiting.
Various Forms of Software Protection
Software can be protected in three primary ways: through copyright, patents, and trade secrets.
Copyright protection safeguards the creative expression of software, specifically the expressive form in which the software is written, including both source code and object code. It is important to note that copyright only protects the expressive form of software, not the underlying idea. For instance, reproducing the same idea with a different expressive form of the source and object code does not infringe copyright. Furthermore, mathematical algorithms are not protected by copyright as they lack creativity.
Patent protection safeguards the underlying inventive idea of the software, focusing on its innovative functionality. Under the European patent system, an invention must possess "technical character" to be eligible for a patent. This means the invention, in the case of software, must solve a specific technical problem using technical means. For a computer program to be patentable, it must produce a "further technical effect" that extends beyond the normal physical interaction between the program and the hardware on which it operates.
Trade secret protection safeguards confidential business information, such as algorithms and programming techniques, which provide a competitive advantage.
There are three fundamental requirements for trade secret protection:
This form of protection is particularly useful for software that does not meet patentability requirements (e.g., due to a lack of technical character) and for protecting algorithms and programming techniques.
Specifics of AI Software Protection and Challenges in Proving Infringement
Artificial intelligence (AI) software presents specific characteristics that pose unique challenges in terms of legal protection and, crucially, in proving counterfeiting.
In particular, AI software operates through neural networks, computational models inspired by the structure and functions of the human brain. Neural networks are composed of interconnected "neurons" organized into layers, with each connection between neurons possessing a "weight" representing its strength. These weights determine the importance each neuron assigns to its input to achieve the desired result. The "training" process involves adjusting the weights of these connections, thereby enabling the artificial intelligence system to "reason".
The way a neural network reasons raises issues related to the complexity and lack of transparency of such systems. Due to the enormous amount of data processed, it is particularly difficult to fully understand how neural networks reason. From a legal standpoint, this difficulty in understanding the functioning of neural networks makes it challenging to compare the features of protected AI software with those of potentially infringing software. Therefore, it is difficult to obtain evidence of AI software infringement.
Technical and Legal Measures for Gathering Evidence of Infringement
To mitigate the "black box" problem of neural networks, AI distillation algorithms can serve as a valuable technical tool. These algorithms are useful for examining data classes and their relationships, thereby gaining a partial understanding of the logic behind the software. In the context of artificial intelligence, the aim of a distillation algorithm is to transfer knowledge from a complex and large-scale model (the "teacher model") to a smaller and lighter model (the "student model"). The student model aims to mimic the behaviour of the teacher model, effectively distilling its knowledge. These algorithms offer a powerful way to analyse the internal logic of the teacher model, which in our case is the suspected infringing software. Training a student model to emulate the teacher model allows us to gain partial insight into the teacher model's reasoning processes by observing the student model's reasoning processes. This process enables us to partially reverse-engineer the teacher model’s reasoning process, in order to identify functional similarities or differences between it and the protected software.
Furthermore, AI software is often protected through process patents. Article 34 of the TRIPs Agreement establishes a legal presumption that facilitates the proof of infringement of a process patent when it is difficult to gather evidence on the process used by the infringer. The presumption of infringement arises when an identical product, produced without the consent of the patent owner, is deemed to have been obtained by the patented process if:
This legal presumption reverses the burden of proof; instead of the patent holder having to prove how the allegedly infringing software works, it is up to the defendant to demonstrate that the patented process is not used by their artificial intelligence system. This helps to overcome the problem of having to ascertain the reasoning process of a neural network.
The Key Role of Italian Jurisdiction and the "Descrizione" Tool
Finally, it is crucial to emphasize the important role played by the Italian system for preserving evidence.
Italy possesses one of the most efficient systems for preserving evidence in Europe, known as "Descrizione" (description). This system is considered so effective at gathering evidence that it has also been adopted by the new European Unified Patent Court for cases involving infringement of a European Unified Patent (Rule 192 of the Unified Patent Court's Rules of Procedure).
The Italian "Descrizione" is an urgent measure aimed at gathering evidence of suspected infringing products (for example, suspected infringing AI software) when this evidence is not freely available. Italian Courts can order a court-appointed patent or IT expert to inspect the suspected infringing software and all relevant data and documents, in order to gather evidence of its features and reasoning process.
During the inspection, a copy of the software is acquired for further examination. Technical, commercial, and accounting documents are also typically acquired. No other inspection or evidence-gathering measure in Europe enables the acquisition of such a large quantity of documents and information relating to infringing products.
These orders are typically issued as "Ex Parte Orders". This means the inspection is carried out before the defendant is summoned, to prevent them from concealing or destroying evidence of the infringement.
It is important to note that an Italian Court's order to preserve evidence can be enforced in all European Union countries under EU Regulation no. 1215/2012, relating to the recognition and enforcement of foreign judgments in civil and commercial matters.
For example, on June 18, 2025, the Court of Venice issued an order, confirming that the Italian “Descrizione” can be enforced abroad for the acquisition of evidence that is in the possession of a foreign third party. In particular, in that case, the order to preserve evidence was enforced against a French client of an Italian infringer and the measure was very useful for describing and acquiring evidence of how the infringing machine and its software operate at a plant located in France.
In the event of copyright infringement committed through software available on the internet, it can be argued that Italian Courts may have jurisdiction in accordance with the principles established in the Hejduk decision of the European Court of Justice (issued on January 22, 2015, in case C-441/13).
In the Hejduk case, the European Court of Justice ruled on jurisdiction in cases of copyright infringement committed through the internet, establishing the "mosaic principle". This principle allows claimants to sue for damages in the courts of each EU Member State in which the infringing work was made accessible through publication on the internet, even if the server is located elsewhere. Therefore, if software published on the internet is accessible to an Italian user connected from Italy, Italian jurisdiction can apply, even if the server is located, for example, in Germany.
Consequently, if it is necessary to gather evidence on an infringing AI software, the principles of the Hejduk decision can justify a request to an Italian Court for an order of "Descrizione" to acquire a copy of the suspected infringing software, along with all relevant documents and information. This evidence can also be used in legal proceedings in different countries.
Conclusions
In conclusion, protecting AI software in the European system requires a multifaceted approach, considering the different forms of protection available. Furthermore, the difficulties in understanding the reasoning process of neural networks present unique challenges in proving infringement.
One way to overcome this is to employ innovative strategies and technical tools such as AI distillation algorithms to partially reverse-engineer the reasoning process of suspected infringing software.
Additionally, legal instruments such as the presumption of infringement for process patents can shift the burden of proof, while Italy's efficient "Descrizione" evidence preservation system provides a powerful mechanism for acquiring crucial data and documents, even across EU borders.
Combining these technical and legal instruments can significantly strengthen the position of right holders in safeguarding their valuable AI software.
Article published in today's Lexology Newsletter, written by Lorenzo Gyulai.