Regulating Generative AI towards the future1 Giovanna De Minico* - Michela Tuozzo* University of Naples Federico II, via Marina 33, Naples, 80125, Italy Abstract This intervention is focused on two issues. The first one aims to consider the current regulatory framework of Generative Artificial Intelligence systems, with specific attention to the obligations of providers and deployers and system governance as dictated in the AI Act. The second issue is dedicated to exploring points of intersection with other regulations applicable to AI systems within the European digital ecosystem. Keywords Generative AI systems, AI Act, Governance 1. Introduction potential implications on a crucial issue: hate speech and online misinformation. Generative intelligence, a cutting-edge development in the realm of artificial intelligence, has significantly 2. GPAIs’ classification influenced the legislative trajectory of the European Regulation on Artificial Intelligence - 2021/0106(COD). The critical characteristics of general-purpose AI AI systems utilizing Large Language Models have (GPAI) models include their large size, opacity, and the capacity to generate a wide range of outputs, potential to develop unexpected capabilities beyond including texts, translations, images, sounds, videos, those intended by their creators. According to article and more. The prospect of these systems 3 (63), a general-purpose AI model means an AI model harmoniously integrating with other AI systems trained with a large amount of data using self- amplifies their usefulness for users, both supervision at a scale that displays significant professionals and non-professionals, as well as for generality and is capable of competently performing a public and/or judicial authorities. The latter can wide range of distinct tasks regardless of the way the leverage them for forecasting, adopting model is placed on the market and that can be recommendations, or making informed decisions. integrated into a variety of downstream systems or The unique characteristics of generative AI have applications. raised questions about the applicability of the On December 6, 2022, the General Secretariat of traditional risk management approach, which forms the Council classified GPAIs as high-risk systems. This the foundation of European technological Regulation, classification enforces specific compliance obligations and how to effectively categorize this new form of AI. (Articles 10-15), requires a risk impact assessment on The proposed intervention aims to highlight the fundamental rights (Article 27), mandates auditing status of generative AI under the AI Act and its prior to market entry (Article 43), and necessitates governance. Consequently, critical reflections will be registration in the EU database, along with post- elaborated upon regarding these elements and their market surveillance obligations. 1 The article reflects collective thoughts; however, the © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). paragraphs can be attributed as follows: paragraphs 1, 2, 2.1, and 3 to Dr. Michela Tuozzo; paragraphs 2.2 and 4 to Prof. Giovanna De Minico. Ital-IA 2024: 4th National Conference on Artificial Intelligence, organized by CINI, May 29-30, 2024, Naples, Italy ∗ Corresponding author. giovanna.deminico@unina.it (G. De Minico); michela.tuozzo@unina.it (M. Tuozzo); CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings The high-risk classification of GPAIs has sparked In addition to those already stipulated for "standard" two types of criticism: the failure to adhere to a GPAIs, providers shall perform model evaluation to precautionary approach and the imposition of identify and mitigate systemic risk; assess and obligations that are perceived as difficult to achieve. mitigate possible systemic risks at the Union level; Through its amendments on June 14, 2023, these report without undue delay to the AI Office and, as critical aspects led the European Parliament to appropriate, to national competent authorities, develop an autonomous classification of GPAIs as relevant information about severe incidents and foundation models. Article 28 ter outlined three possible corrective measures to address them; ensure categories of obligations for the provider: risk an adequate level of cybersecurity protection. identification and mitigation, testing and evaluation, and documentation. 2.2. Governance2 However, it's important to note that the regulatory model was significantly altered in the final version of The governance of the artificial intelligence the adopted AI Act, a result of the compromise market is complex due to AI's diverse applications in reached among the European institutions during the both public and private sectors across national and EU trilogue phase. This change was influenced by levels. It is important to note that responsibility for lobbying efforts (Bareis), underscoring the political actions like fundamental rights impact assessment dynamics at play in the regulatory process. and conformity evaluation falls on the entrepreneur's initiative, reflecting a confusing blend by the Commission of individual centrality and a trend 2.1. Tiered approach toward system privatization. The governance system could have taken three forms: complete decentralisation, relying on national The Regulation of generative intelligence became its oversight systems like telecommunications, complete category in the final version of the AI Act approved on centralised supervision disregarding national March 13, 2024. Within this category, generative AI is variations, or a mixed system with tasks entrusted to classified into three types: general-purpose AI model, the Commission in a designated Directorate-General general-purpose AI model with systemic risk, and and to European agencies. The governance of the AI open-source general-purpose AI model (articles 51 – Act takes different forms depending on the level 55). As a result, we have different rules for different considered and even the type of intelligence, as shown GPAIs. by the case of GPAIs. Examples of entirely European In essence, Spanish Presidency of the EU Council Independent Authorities, independent from National has aimed to strike a balance between the Council's Governments and the Commission, were established hands-off approach and the Parliament's earlier in 2010 with the European Banking Authority, the stance of establishing uniform rules for all generative European Insurance and Occupational Pensions AI systems. Authority, and the European Securities and Markets Specifically, providers of "standard" GPAIs, when Authority. generating synthetic audio, image, video, or text While traditional AI systems have an intermediate content, must ensure that the outputs of the AI system governance form balancing decentralisation and are marked in a machine-readable format and centralisation, generative AI systems follow a fully detectable as artificially generated or manipulated centralised approach. This led to a complex (Article 50, paragraph 2). In addition providers shall governance structure: the original proposal involved (Article 53): draw up and keep up-to-date the three authorities (Commission, National Authorities, technical documentation; comply with Union and AI Board), but the final act expanded to at least 5 copyright law; make a sufficiently detailed summary (including the Advisory Forum and Scientific Panel). of the content used for training publicly available. The entire AI system obediently follows a Conversely, providers of open-source GPAIs must government-centric approach. It is justified solely by only comply with copyright rules and those regarding the fact that AI will become the engine of public the synthesis of content used for training. policies that must firmly remain in the hands of the The obligations for providers of GPAI with current political majorities. systemic risk are more extensive than for "standard". 2 This part is due to Professor Giovanna De Minico. As for the Commission, it has the power to have superficial rule because it is unclear how to ensure the final and definitive say on corrective measures functional independence without first guaranteeing proposed by national supervisory authorities, organisational independence, known as genetic confirming an approach centred on the community independence. Additionally, Article 70 does not executive. specify which entities should respect this As for the AI Board, it is ensured only functional independence. Previously, Article 59, par. 4, stated independence; indeed, genetic independence from the that “members of each national supervisory authority, Commission and Member States is not required. (…), shall neither seek nor take instructions from Evidence of this is the freedom of each State to send anybody and shall refrain from any action whomever they want and the presence of the incompatible with their duties”. Removing this part of Commission on the Board. Consequently, even the rule suggests that independence is only aimed at functional independence is at risk, as already those being regulated, not political representatives. evidenced by the comparison between European This suggests that the AI Act has accepted a partial Parliament amendments (Article 56) and the final text risk of capture because it only focuses on regulating of Article 65. This rule states: "The Board shall be solid entities. organised and operated to safeguard the objectivity The National Authority possesses regulatory and impartiality of its activities". The original powers, allowing it to mandate actions such as formulation of Article 56 stated: "The ‘European suspensions, corrective measures, or removing AI Artificial Intelligence Office’ (…) shall be an systems from the market. Could personal ablation be independent body of the Union. It shall have legal constructed as a complex administrative action with personality" and the Office "act independently when unequal powers? To this question, we respond carrying out its tasks or exercising its powers” (Article affirmatively because the National Authority has the 56 quater). Currently, in terms of its structure, the authority to propose the action, with the Commission Office is integrated within the administrative having the final say. Therefore, two authorities – the framework of the Directorate-General for National and the Community – intervene in the same Communications Networks, Content and Technology decision but at different times and with different (DG-CNECT) of the Commission. It does not have contributions: one proposes, and the other finalises operational autonomy from DG-CNECT. Furthermore, the procedure. This procedural collaboration achieves unlike national authorities, no dedicated coordination between authorities operating infrastructures or technical, financial, or human simultaneously within the National Union network. resources are provided. In a summary overview, the governance of the AI Regarding National Supervisory Authorities, it is Act reserves central political authority for the permissible for the State to designate them within an Commission, whose verbum is communicated affiliated entity such as the Government (as downstream to the National Authorities. Then, it emphasised in the Privacy Commissioner's letter returns to the Commission itself, with an AI Board dated March 25, 2024). intervening occasionally to address gaps in the In Italy, the agencification approach is confirmed discourse. by the legislative initiative on a delegated law In contrast to the system described above, the regarding artificial intelligence approved by Council generative system governance is centralised between of Ministers on April 23, 2024. In the draft, Article 18 the AI Office and the Commission. designates the two Authorities: the Agency for Digital The Commission shall have exclusive powers to Italy (AgID) and the National Cybersecurity Agency supervise general-purpose AI models and request (ACN). In both cases, these government agencies measures and shall entrust the implementation of achieve functional independence only with respect to these tasks to the AI Office, a European Agency of the the regulated entities but not the representative Commission. political body. The Government's choice seems clear: The AI Office plays a central role in developing a to maintain control over "intelligent policies" in its Code of Practice and monitoring its application. hands, rejecting the model of independent authorities. The Scientific Panel serves as a qualified advisory In addition to functional independence, body to the AI Office. organisational independence was expressly Unlike traditional AI models, governance for requested in the Parliament's amendments (Article generative models is fully Eurocentric to the extent 59, par. 1, EP). that the Commission fulfils its administrative role and Article 70 (AI Act adopted) only recognises possesses law enforcement powers. The functional independence, which seems like a centralisation of governance is at its maximum. While centralisation under the Commission is intervene. Civil society organisations, industry, justified by the sector's sensitivity and the need for a academia, and other relevant stakeholders, such as unified implementation approach, it raises concerns downstream providers and independent experts, may about deviating from the independent European also support the process (Article 56). authorities established in 2010. Co-regulation in the technology sector has shown various reasons for fallibility, with the primary concern being the risk of capture by regulated solid 3. Boundaries entities. Lastly, there is an emphasis on the responsible 3.1. Disadvantages of the AI Act behaviour of providers but a need for proper division of responsibility with 'downstream' users. It should The qualification of so-called 'systemic risk' will be considered that the distribution of responsibilities initially depend on the capability, either based on a along the value chain should involve multiple parties, quantitative threshold of the cumulative amount of each with different responsibilities, particularly compute used for its training measured in floating considering the user's role depending on whether point operations (FLOPs set as 10^25) or based on a they use the output for professional purposes decision of the Commission, ex officio or following a (Hacker, Engel, Mauer). qualified alert from the scientific panel. It is presumed that a model trained with large amounts of data and advanced complexity has foreseeable adverse effects 3.2. Addressing Challenges Within and on public health, safety, public security, fundamental Beyond the AI Act rights, or the society as a whole that can be propagated at scale across the value chain. Beyond the regulatory aspects addressed in the AI Act, Upon closer examination, this definition of the proliferation of this new type of AI also presents systemic risk consider the combination of the interpreters with the issue of the rapid pace of probability of a harmful event occurring and the technological transformations. From a constitutional severity of that harm, as well as the values and assets law perspective, it is essential to clarify the categories of constitutional relevance. This appears consistent involved to assess whether the discipline outlined in with the framework of the AI Act, which prohibits all the article respects fundamental liberties. AI systems whose use is deemed unacceptable Consider that the prerogative of the most because it contradicts Union values. However, there widespread GPAIs – such as chat GPT – is needs to be coherence between the means, communication. represented by numerical indicators such as FLOPs, We need to raise the following questions: Does and the end, which is the protection of common generative AI produce ideas? Is it a new form of constitutional values. These parameters «describe the media? Or a digital private communication? foundational model but not its impact on society, Article 15 of the Constitution protects the safety or fundamental rights» (Helberger et al.). limitation of the freedom of communication by the Another criticism concerns the quantitative and guarantees of legal reservation and jurisdiction, but qualitative reduction of provider obligations, even for with a significant difference compared to Article 21. GPAIs with systemic risk. It is surprising to note that The limits expressed (good conduct) and unexpressed for such models, the activity of demonstrating (protection of personality rights such as reputation compliance before placing the AI system into the and privacy) would not apply to communications as market or service is not accompanied by guarantees they do to Article 21. Furthermore, protecting of a prior conformity assessment for the provider and freedom and correspondence from undue a fundamental rights impact assessment for the interference would extend to the recipient and the deployer. Compliance can be demonstrated by relying sender (in our case, OpenAI, Google, and others). on codes of practice within the meaning of Article 56 The constitutional coverage of Article 21 would until a harmonised standard is published. imply extending the guarantees of the press medium The AI Office, in collaboration with the Board, to chat GPT as well: the prohibition of censorship, the encourages the drafting of the Code. They aim to possibility of adopting inhibitory acts with the ensure that the Codes of Practice comprehensively guarantees of legal reservation and jurisdiction, and address the obligations in Articles 53 and 55. finally, the possibility of limiting its contents. However, all providers of general-purpose AI models and relevant national competent authorities will Finally, similarly to the issue of advertising The criterion for identifying GPAIs at systemic risk information, it could be argued that informational relates to computational capacity: “when the content has an economic purpose. Therefore, it would cumulative amount of computation used for its be more appropriate to adopt the limits of Article 41. training measured in floating point operations is Thus, it cannot be carried out in contrast to social greater than 10(^25)” (Article 51 AI Act). Unlike for utility or in a way that harms health, the environment, the DSA in this case, the number of clients is security, freedom, or human dignity. Upon closer disregarded because the taxable entity is examination, the protection here is also twofold: identified because of the capacity to input data and towards the end-user citizen and other commercial occupy the data world. operators. The second difference is an objective one that Providing a definitive answer to the question of concerns the service. Platforms render an constitutional coverage requires further examination. intermediary service: they connect those who It is essential to consider the communicative context generate information with those who receive and the communicating subject: We are within the information, and this encounter between supply and protective sphere of Article 15 when the intention is demand happens on the platform. So, the platform to maintain the secrecy of the content of virtual does not put its hand on the information; it hosts it, correspondence, the recipients are specific and rationalizes it, organizes it, and categorizes it, but immutable, and the means are suitable for achieving someone else is generating the idea. Generative AI, on secrecy. It is precisely this last requirement that leans the other hand, does precisely what the platform does towards the category of Article 21. not do; that is, it occupies that space vacated by the When considering the fine line with Article 41, we platform because the AI is not a host; it is the author must examine the purpose of freedom of expression. of an idea of its own can argue about whether Chat- If it serves an economic aim, such as profiling, then the GPT creates the idea out of thin air, or whether the broader protection of Article 21 may not apply. idea is generated by fishing around the network, how (Ruffolo). it articulates it, and so on but it is still something that involves creative energy. Even though Chat-GPT, unlike the human mind, does not create from anything 4. Next steps: Implementing the but from a background, it still realizes a vital, active, innovative contribution that is not there in the Regulatory Framework in the platform. Digital Ecosystem3 So I have come to say that there is a subjective and Upon closer examination, we found that adopting objective difference, and because of this difference, to the AI Act only addresses some of the issues raised by be strict, the discipline of DSA cannot be applied. generative AI systems. 2) If we combine all the regulations, they would It could be a reason for adding the general still fall under the guidelines of the AI Act, which has discipline of the AI Act with some particular deliberately exempted GPAI. GPAI was only included disciplines, such as Digital Services Act – Regulation in the final negotiations as a last resort, with less strict 2022/2065 – and Digital Market Act – Regulation rules compared to other forms of intelligence. 2022/1925, which follow a community goal: to create Essentially, GPAI has been shielded from excessive a single digital ecosystem. regulation. Those advocating for additional rules are This cumulation operation cannot happen going against the original purpose. Since EU laws are automatically, but it should be conducted with some meant to be interpreted beyond just the words, this questions showing us how to integrate the disciplines. combination of regulations would contradict the 1) Is there a difference in the passive legitimates intentio legis of the AI Act, burdening GPAI with of the two disciplines? By passive legitimates, we regulations it was meant to avoid. While we might mean the recipients of the rules of the Digital Service regret not imposing more regulations, it is worth Act and the AI Act. The former are the platforms that acknowledging that GPAI has been protected as are identified in the discipline because of a quantity: intended. “several average monthly active recipients of the Finally, let us ask what the purpose of DSA is. The service in the Union equal to or higher than 45 DSA is for keeping the net clean of blatant million” (Article 33, DSA). malfeasance, misleading, hate speech, fake news, and 3 This part is due to Professor Giovanna De Minico. so on. Provided that the DSA is not intended to reasoning that hateful conduct is prohibited because derogate from the general principle that the platform it is prohibited. Rather, tautology can conceal is not the editor-in-chief of a newspaper, it is exempt dangerous liberticidal theses and easy slides toward from a prior and general control obligation, but it does only permissible speech: state speech. Thus, the put in place a punctual control obligation, thus not control of permissibility conceals insidious forms of generalized, and not ex-ante but ex-post. This merit-based scrutiny of the manifestation of thought, obligation of control signifies that the ultimate goal is opening the door to digital censorship on the Web. to hold together a control that does not impose a This undergoes a radical change from a place generalized vigilance to which no platform will ever unscathed by heteronomous interventions and free want to submit, surrendering fundamental freedoms, from information intermediaries to a space thus the right to say and have expressed those who supervised by private individuals who have the keys then put news on the platforms. to open and close the information agora in their Suppose we have a positive answer on positive hands. This risk cannot be avoided due the absence of DSA’s purpose. In that case, it could be shifted to the the normative definition of falsehood, with the Chat-GPT as well? Moreover, why not if this policy of paradoxical consequence of a caesura between the cleaning up the network is so positive, even if some offline environment, where the dissemination of false (De Minico d) see it as a form of censorship entrusted ideas does not constitute a crime unless it attacks to the private entity? When the Commission other goods-interests other than the truth, and the designates platforms as “providers of large online virtual one, where instead the idea if false ceases to be platforms” (Articles 15 and 33 DSA), they become the the exercise of a right to become an illicit fact. subjects of a timely and subsequent obligation to Therefore, I do believe that the DSA has control the information stored and transmitted by aggravated the limit of lawfulness, making certain their platform to ensure a transparent and secure conduct that is lawful offline unlawful when it sees the digital environment. playing field changed. The fulfilment of this selective cleanup duty-which The most severe thing about this blank reconfirms the absence of a generalized duty of endorsement to platforms of the power to control the control by object and, over time, according to the merit of others' ideas is the emergence of an unseen philosophy of the e-Commerce Directive, now re- function assigned to private individuals, which in the proposed in Art. 8 DSA, should equalize the real world does not even exist in terms of power asymmetrical relationship between the platform referred to as a public authority. I prefer to trust in the owner and the author of the hosted content, just as it beneficent virtues of the marketplace of ideas, which, should put the author of the content and its recipient, in allowing the coexistence of the false with the true, the end user of the information flow, on the same lets citizens distinguish between the two entities level. because it believes in their ability to mature a Without prejudice to my doubts about the responsible idea without being pre-addressed by suitability of this asymmetrical measure to equiordize those who claim to know for them what objective misaligned social partners, I would instead call truth is. attention to a possible effect that could affect This failure of the D.S.A. to predetermine the platforms. These, in order not to incur a liability concept of forgery degrades abstractness into judgment due to the maintenance of illicit content concreteness and generality into particularity. The online, will be inclined to delete rather than preserve law is resolved in the ordinal provision of the private the ideas of others, based, moreover, on summary strongman, while the equality of citizens before the evaluations pronounced inaudita altera parte. Adding law is attenuated in the concrete provision ad certam to this is a further consideration: the lack of an personam. abstract and general definition of the concept of false Adopting instead the point of view of those who leaves platforms free to confuse false with politically consider it positive and desirable the extension of its inappropriate news or news that does not conform to contents to the A.I. Act, then this operation cannot be dominant thinking. No more complete is the carried out by making an automatic addition of prohibition of hate speech, which lacks a prior disciplines but can be applied in interpretation. typification of hate speech, even regarding the 3) Who does the interpretation? The Commission necessary causal link between the saying and the has governance over the GPAIs and, therefore, could activating effect toward the recipients of the impose on them obligations arising from the digital prohibited conduct sought to be solicited. This service, not by automatic acquisition (Hacker, Engel, normative gap in the DSA points back to tautological Mauer; Botero Arcila) but according to reasoning by analogy legis, assuming that between the two cases, Available at SSRN: the regulated, the digital and the unregulated Chat- https://ssrn.com/abstract=4539452 GPT, via is the identity of facts. At the end of this [3] A. Cefaliello, M. Kullmann. “Offering False discussion, we have partially applied the rules of the Security. How the Draft Artificial Intelligence Act DSA to chat, but only by interpretation and assuming Undermines Fundamental Workers Rights”. that scrupulous reasoning is conducted. European Labour Law Journal 4(2022), pp. 542- While maintaining a negative judgment on the 562. content of the DSA, adopting instead the perspective https://doi.org/10.1177/20319525221114474 of those who consider it positive and desirable to [4] G. De Minico a. "Too many rules or zero rules for extend its contents to the AI Act, then such an the ChatGPT?" BioLaw Journal 2(2023), pp. 491- operation cannot be carried out by automatically 501. https://doi.org/10.15168/2284-4503- adding disciplines (Hacker, Engel, Mauer; Botero 2723 Arcila), but can be applied interpretively. [5] G. De Minico b. “Una norma unica che tenga Who does the interpretation? The Commission, insieme ChatGpt e privacy”. Il Sole 24 ore, June which has governance over GPAIs and could therefore 14, 2023. impose obligations on them arising from the DSA, not [6] G. De Minico c. “The Challenge of the Virtual through automatic acquisition but through analogical World for Independent Authorities”. European legal reasoning, admitting that between the two cases, Public Law 29:1(2023), pp. 1-26. the regulated digital pathway and the unregulated [7] G. De Minico d. “Internet e le sue regole”. Atti del GPAI, there is a factual identity. Convegno Processi politici e nuove tecnologie At the end of this discussion, we have partially tenutosi a Università degli Studi di Bari - Aldo applied the rules of the DSA to chat, but only Moro nel 22 giugno 2023, edited by M. Calamo interpretatively and provided that a careful reasoning Specchia, Giappichelli, Torino, forthcoming. is conducted. [8] G. De Minico e. “Nuova tecnica per nuove Similarly, one should reason about the extension diseguaglianze. Case law: Disciplina of the contents of the DMA. Telecomunicazioni, Digital Services Act e In conclusion, Regulations in the digital Neurodiritti”. Federalismi.it 6(2024). environment provide solid foundations for [9] L. Floridi, M. Chiriatti, “GPT-3: Its Nature, Scope, addressing the risks posed by the rapid evolution of Limits, and Consequences”. Minds & Machines 3 generative AI. Additionally, it is crucial for the (2020), pp. 681–694. European Commission to fully utilise its https://doi.org/10.1007/s11023-020-09548-1 implementation power during the execution phase [10] P. Hacker. “AI Regulation in Europe: From the AI (Articles 290 and 291 TFEU), adapting the pace of Act to Future Regulatory Challenges”. Oxford technology to existing rules to protect fundamental Handbook of Algorithmic Governance and the rights. Law edited by I. Ajunwa, J. Adams-Prassl, Oxford University Press, 2024. [11] P. Hacker, A. Engel, M. Mauer. “Regulating Acknowledgements ChatGPT and other Large Generative AI Models”. FAccT 7(2023) pp. 1112-1123 The initiative falls within the scope of the FAIR Project https://doi.org/10.48550/arXiv.2302.02337. FAIR (Future Artificial Intelligence Research), WP 3.8 [12] N. Helberger, N. Diakopoulos. “ChatGPT and the – Resilient AI, Ethical - Legal and Societal Issues in AI Act”. Internet Policy Review. Journal On Resilient AI Systems coordinated by Professor Internet Regulation 12:1 (2023), Giovanna De Minico. https://doi.org/10.14763/2023.1.1682. [13] C. Novelli, P. Hacker, J. Morley, J. Trondal, L. References Floridi. “A Robust Governance for the AI Act: AI [1] J. Bareis. “BigTech's Efforts to Derail the AI Act”. Office, AI Board, Scientific Panel, and National VerfBlog, 2023/12/05. Authorities,” May 5, 2024, available at SSRN: DOI: 10.59704/265f1afff8b3d2df https://ssrn.com/abstract=4817755 or [2] B. Botero Arcila. “Is it a Platform? Is it a Search http://dx.doi.org/10.2139/ssrn.4817755 Engine? It's Chat GPT! The European Liability [14] U. Ruffolo. “Piattaforme, A.I. generativa e libertà Regime for Large Language Models”. Journal of di (formazione e) manifestazione del pensiero. Il Free Speech Law, 3 (2023), pp. 455-488. caso ChatGPT”. Giurisprudenza italiana (2024), pp. 472-480.