The extremely rapid technological evolution, which has characterized the last decade, has placed intellectual property in front of new challenges, posing different interpretative and applicative issues, which respond to the needs of protection and definition of the new scenarios, among which deserves, certainly, mention artificial intelligence.

The spread of this technological tool has given rise to several novelties that have, from the outset, directly impacted the legal system, highlighting new (or pre-existing) regulatory gaps, which have required the intervention of European jurists.

In the field of intellectual property it should be noted that, among several recent interventions, the EUIPO, has addressed this technological evolution through the drafting of the so-called IP Infringement and Enforcement Tech Watch Discussion Paper 2023, in which the potential implications and repercussions of new technologies (especially in the field of AI) vis-à-vis intellectual property rights were addressed.

The doctrinal and institutional fervor created made it possible to identify in the so-called "generative AI," in terms of input (in the implementation phase of the AI model) and in terms of output (due to the work created), some of the issues that needed ad hoc regulation, which led to the approval of the European AI ACT Regulation.

The Regulation, which consists of 113 articles and 13 annexes (unlike the GDPR based on a principle of "accountability"), provides for a different classification according to the different types of AI due to the degree of risk related to their enforceability.

Specifically, the structure of risk assessment is divided into four categories:

1. unacceptable risk: unacceptable risks for systems are expressly prohibited;
2. high risk: systems will be assessed based on compliance with minimum requirements;
3. limited risk: for systems held to transparency requirements;
4. low and minimal risk: for systems exempt from obligations.

This mapping procedure becomes of paramount importance for companies, as it contains the guidelines to be followed to plan any corrective actions in order to comply with the provisions contained in the AI Act itself.

It should be noted, insofar as it is of interest here, that specific provisions in the area of intellectual property emerge among the indications stipulated in the Regulations under review.

Consider, in fact, that AI-based technologies collect and process numerous data, becoming of paramount importance to prepare, on the one hand, the appropriate proactive measures suitable to protect trade secrets used by AI (i.e., through the preparation of specific NDAs) and, on the other hand, to prepare preventive control measures, in order to avoid that the data (and AI models themselves) may infringe the intellectual property rights of third parties (trademarks, designs, patents and copyrights).

Specifically, in the area of copyright, there still remains the long-standing issue, in the field of generative AI, concerning the copyrights associated with "training data", which can be extrapolated from material designed specifically for training procedures. However, it recurs more and more frequently, that such inputs are extracted from public domain databases, but copyrighted. Since generative AI training involves the reprocessing of pre-existing content, the latter is often copyrighted (despite the fact that it is on freely accessible and searchable databases) and/or available online without the knowledge or permission of the respective owners.

Regarding this type of data, the AI Act, in Article 53, provides specific provisions and obligations for generative AI training procedures:

• "put in place a policy of compliance with Union copyright, in particular to identify and respect, including through state-of-the-art technologies, reservations of rights expressed in accordance with Article 4(3) of Directive (EU) 2019/790."
• “make public a sufficiently detailed summary of the content used in the formation of the general AI model, according to a template provided by the AI Office".

As can be seen from the above provisions, Article 4 of EU Directive 2019/790, which grants holders the power to impose limits (with the exception of specific cases listed in the Directive) on the reproduction and/or extraction of texts and data, also applies to the new Regulation.

It therefore becomes a must for AI technology providers to request and obtain permission from rights holders for any use of protected data/content, during AI model training procedures.

Finally, the AI Office is in charge of reviewing the summary reports submitted by AI-based system providers, which must contain a detailed summary concerning all data and banks used for the extrapolation of training data.

From a jurisprudential point of view, at the international level, two recent conflicting pronouncements in the field of generative AI are worth mentioning, regarding the ownership of generated works and copyright infringement.

Regarding the former, The Beijing District Court held that the prompts used by the author of the work (through software called Stable Diffusion) are protectable by the rules in the field of intellectual property, recognizing the originality of the image created, due to "the repeated adjustment of such parameters, reflecting his aesthetic choice and personalized judgment".

Of a different view is the order issued by Federal Judge William H. Orrick of the U.S. District Court for the District of Northern California, which opened up the possibility that content used as training data for AI systems may not be subject to protection.

In that proceeding, the defendant was accused of "downloading or otherwise acquiring copies of billions of copyrighted images without permission to create Stable Diffusion and used those images to train the model, causing the works to be stored and incorporated into Stable Diffusion as compressed copies".

The court rejected the plaintiff's claims for failure to identify which works were actually used and because the alleged infringement was not based on irrefutable research results.

Currently, the scenario envisions the existence of two strands of interpretation regarding the training data of generative AI systems, on the one hand, recognizing the protectability, in terms of copyright, of the data used and, on the other hand, recognizing the possibility that the training data may fall under the so-called U.S. Fair Use, thus protecting the transformative uses made by the author, which bring originality to the work processed.

To conclude, we recall the interesting study Futures of innovation and intellectual property regulation in 2040 carried out by the European Commission, aimed at defining some possible and possible scenarios in the field of intellectual property, briefly summarized as follows:

• Scenario 1: The end of IP as we know it: the digitization of the economy drastically reduces the use of traditional IP protection tools (patents), in favor the know-how and trade secrets inherent in AI software;
• Scenario 2: "Creative destruction of the IP regime": New AI systems redefine patent preparation and examination procedures;
• Scenario 3: IP as a battlefield of geopolitics: geopolitical tensions allow some states to implement protection and secrecy strategies that collide with a uniform European patent system;
• Scenario 4: Global and balanced IP for open innovation: technological development allows for rapid commercialization and dissemination of information, enabling global and digitized IP systems;
• Open-source collaboration globalized innovation: predefined licensing arrangements in a context of open-source sharing of digital content.

Key factors of change on innovation and their relevance for the five scenarios (source: European Commission)

1. These are the AI models that apply different algorithms to the data inputs relevant to the implementation and/or execution of the predetermined task, or output, for which they have been programmed. Although the terms “Algorithms” and “Models” are often used alternately, they constitute different steps in the AI model implementation procedure. Algorithms, in fact, are procedures (defined in mathematical language) that must be applied to a specific set of data in order to achieve a given function. Models, on the other hand, constitute the output of an algorithm, previously applied to a specific dataset.
2. EU Data Protection Regulation No. 679/2016 (General Data Protection Regulation - GDPR), which governs the processing and protection of personal data of natural persons within the EU, as well as their subsequent movement, which defines how such data should be processed, collected, used, protected, and shared, conferring certain obligations and guidelines on individuals, companies, and organizations charged with this task.
3. Principle of accountability of those involved in the processing of personal data, used as a parameter for assessing the level of risk, through a decentralized system.
4. Risk-based approach - centralized risk assessment. Internet Law - Digital Copyright and Data Protection 2024 No. 2, Pacini Giuridica, p. 201.
5. The AI Act provided for certain compliance obligations on the companies concerned, among which emerges the obligation to ensure adequate protection of fundamental rights in their potential applications, within AI systems. Indeed, the Regulatory Text, taking as an example some of the provisions contained in the GDPR, has included specific provisions aimed at avoiding violations of the right to human dignity, respect for private life, and protection of personal data (i.e., avoiding gender discrimination).
6. Synthetic data (creation of digital images from a catalog of real images) created by an AI as a training medium, Mischitelli, Synthetic data new frontier of artificial intelligence: opportunities and limitations, agendadigitale.eu, July 14, 2021.
7. The use of digital content obtained through access to open access platforms does not prove the lawfulness of the relative use of the data itself, since the latter may be subject to certain terms of use or, as seen earlier, published without the author's authorization: Internet Law - Digital Copyright and Data Protection 2023 No. 4, Pacini Giuridica, p. 619.
8. The European AI Office is the strategic center for the implementation of the AI Act, established within the European Commission. This office provides assistance and cooperation to member states to ensure safe and reliable implementation of the AI.
9. Directive on copyright and related rights in the digital single market.
10. Unless exempted or limited by law.
11. For an in-depth study of pending AI cases: AI Lawsuits Worth Watching: A Curated Guide, Tech Policy Press.