AI e piattaforme digitali ridefiniscono il lavoro: come garantire transizioni e tutele efficaci?
Table of Contents
1. The Impact of the Ongoing Transformations on the Structures of Our Systems and the EU Regulations
2. The Principle of Human Control over Technologies: New and Traditional Regulatory Techniques
3. Information Obligations: Innovations and Implications
4. Risk Analysis and Impact Assessment: Connections and Overlaps between the GDPR and the AI Act
5. The Different Competences for Risk Assessment: Coordination among Regulations
6. Objects and Diversity of Risk Assessment Techniques
7. Limits of Risk Assessment. Mitigation or Elimination of Risk?
8. Interpretative Clarifications and a Plurality of Instruments for Human Control. Participatory Impact Assessment
9. The Evolution of EU Regulations: From the Due Diligence Directive to the Platform Work Directive
10. Innovations in the Digital Platform Work Directive
11. Human Control over Algorithmic Management
12. The Request for Review of Decisions
13. Obligation to Justify and Ascertain Algorithmic Discrimination
14. Continuity and Challenges of the European Social Agenda
1. The Impact of Ongoing Transformations on the Structures of Our Systems and European Union Regulations
The previous legislative term of the European Union, spurred by the extraordinary events of the period, was characterized by an intense wave of regulatory activism that also addressed topics which in the past received little attention—often at the very limits of EU competences.
New social issues, containing innovative content, were regulated and have stimulated a renewed interest among labour law scholars, myself included.
Measures were approved—not without arduous mediations—that, for the first time, directly addressed the transformations of the fundamental structures of our productive systems and, indeed, of our human relationships: the data-rich web networks managed by artificial intelligence systems, platform companies, and their supply chains organized within new, often transnational, organizational frameworks that even use these data as productive tools.
These regulatory interventions affect, with varying intensity, the functioning of such structures and have pervasive implications for many economic and social sectors, thus directly impacting the regulation of both individual and collective labour relations.
At least five instruments were approved during the legislative term, which—although developed and finalized within different time frames (from 2019 to 2024) and from different perspectives—share this broad regulatory field: the GDPR (General Data Protection Regulation), the Corporate Sustainability Reporting Directive (CSRD 2022–2024) and the Corporate Sustainability Due Diligence Directive 2024/1760; lastly, the Artificial Intelligence Regulation (AI Act 2024/1689) and the Directive on Improving Working Conditions in Platform work (2024/2831).
It is not yet possible, at least based on current evidence, to know whether these new rules governing productive and social structures already constitute a genuine “system” alternative to the historical Fordist model—in its various forms—in which labour law and employment policies were formed.
The speed of evolution and the transformative capacity of digital technologies create a situation of radical uncertainty, making even near-future forecasts difficult. They certainly do not facilitate the rapid formation of new, stable equilibria among the various forces and structures that now act disruptively on traditional frameworks.
Yet it is undeniable that the new material and immaterial realities referred to by the aforementioned regulations are not mere variants of the old productive world and labour system; rather, they transform their fundamental characteristics, with consequences that are still undefined regarding the implications for institutional structures and legal orders.
The complexity of recent European regulations—and the fact that they deeply affect traditional regulatory frameworks—makes it extremely difficult for jurists to clarify their implications both for the European legal order and for our national labour law. Moreover, the national transposition process itself suffers from the same difficulties and already exhibits some shortcomings. In fact, the growing number of commentaries on the aforementioned directives and regulations shows gaps and falls far short of shedding light on the main unresolved issues.
This interpretative work, however, is not the only one we should undertake. It must be integrated and could benefit from a broader reflection on the systemic implications of the interweaving of such regulations on existing legal orders.
If it is true that the ongoing technological transformations concern both the material and immaterial structures of traditional productive and social systems, they are bound to have equally structural implications for the main institutions of our legal orders, which must be analyzed in all their complexity, even amidst the present conditions of uncertainty.
The artificial intelligence systems and digital technologies through which they operate have transformative capacities throughout the entire economy and society, which manifest with specific modalities and effects in various sectors and fields. Analyses of the implications for the world of work—even though still in an early stage—already provide clear signals of these systems’ capacity to redesign the relationships between companies and workers.
As with other technological innovations, the possible effects are ambivalent.
Indeed, they receive different and often opposing evaluations: from those who highlight the opportunities that artificial intelligence systems open for productivity and work quality³, and from those who argue that algorithmic management would offer “potent new ways in which behaviours (of employees) can be redesigned enforcing a position of total subservience to the will (and whim) of superiors”⁴.
The ultimate consequences will depend on the decisions made over time and as a whole, according to the (changing) historical relationships between public institutions and social forces.
The declared objective of the European rules is to prevent the adverse effects of these technologies on human rights and the environment, without undermining their positive potential for productivity and employment.
2. The Principle of Human Control over Technologies: New and Traditional Regulatory Techniques
The pursuit of this principle is entrusted to a complex set of rules—both substantive and, above all, procedural—that are rather convoluted, with human control over these technologies as their guiding principle. In fact, making this principle effective is the main critical point of all the issues addressed by the aforementioned regulations.
It is critical because regulating and controlling digital technologies powered by artificial intelligence require tools different from those experienced under traditional labour law—and whose effectiveness has yet to be fully tested. This is also a structural problem: it stems from the fact that the languages and operating logics of these technological systems are very different from those prevailing in the analog world and adopted by human beings. In the most advanced versions, machine learning mechanisms add to the decisions under scrutiny “new layers of complexity that are more powerful and indecipherable than human managers”⁵.
For this reason, it is stressed that “insisting on the importance of uncovering the core of the algorithm constitutes a naïve reading of how these systems operate”⁶. In reality, the more autonomous neural networks become, the more we are forced to observe their functioning from the outside in order to seek the most effective means of control.
This is what the various regulatory bodies set up by the Union aim to achieve using a plurality of techniques—mostly procedural and remedial—that are variably directed at the stakeholders on the field, at companies, at their associations, and at their supply chains. As will be seen later in the analysis, this amounts to a series of sequentially imposed obligations.
First of all, there are obligations of transparency and information, detailed with regard to all phases and contents of corporate decisions—not for their own sake, but in order to establish processes for analysing and assessing the risks related to those decisions. These processes, in turn, prefigure the duty to activate direct interventions to prevent, minimize, or remove the adverse effects resulting from the occurrence of risks; they impose on decision-makers the obligation to set up systems of human oversight over the decisions made by automated systems, granting affected parties the possibility to request a review of such decisions, as well as the right to obtain an explanation and justification of those decisions—and, of course, the possibility to challenge them in court to verify their legitimacy under current rules.
These procedures for information, risk assessment, human oversight, and review are referred to the individual workers affected, but are accompanied by the collective participation—of varying intensity—of their representatives. Such obligations are further complemented by instruments designed to support the affected parties in researching and proving their claims, which are often obscured by the low transparency of these mechanisms. These instruments include well-known tools such as legal presumptions, the reversal of the burden of proof in favour of workers, and specific investigative powers of judges and administrative agencies.
The capacity of this multifaceted, still somewhat disorganized European regulatory framework to implement the principle of human control over digital technologies and artificial intelligence systems is subject to varying assessments by commentators. In any case, such assessments can only be partial and provisional, as they refer to manifestations of a rapidly evolving reality whose systemic implications remain undecipherable.
Nonetheless, this does not exempt us from engaging in analyses that go beyond mere commentaries on the various, fragmented European regulations to search for possible systemic implications. Such an attempt—even if preliminary—may be useful to facilitate the full understanding of the obligations for operators and the horizontal connections among the various regulatory fields, thereby promoting the search for the effectiveness of both procedural and substantive guarantees.⁷
Part of the investigation undertaken here aims to analyse the aforementioned European regulations to clarify the connections and possible coordination among them. I must immediately note that some of the coordination and clarifications needed to better define the impact of the various regulations are possible through interpretation; others, however, will require normative interventions or at least European guidelines, and, in other cases, adjustments by national legal systems through appropriate legislative instruments.
Firstly, I intend to reflect on the relationships among the regulations that require companies to identify risks and conduct impact assessments for the rights of individuals and for the environment regarding their activities (and those of their supply chains). Secondly, I propose to analyse the implications of employing artificial intelligence in the management of labour relations and the rules for addressing them provided by the recent European Directive on platform work.
3. Information Obligations: Innovations and Implications
Recent European regulations share several intervention inspired by the goal of ensuring effective human control over technologies and artificial intelligence systems. An initial set of rules, present in various directives, establishes widespread obligations of transparency and information regarding the processing of personal data and the corporate decisions that affect the interested people.
The importance of these provisions should not be underestimated, because the obligations are defined in rather stringent terms: they must be provided within the necessary time frames to be useful in relation to the decisions; they cover all the implementation phases, and they include all the informational content required to assess the meaning of the decisions.
Moreover, the information requested is not limited to providing descriptive elements or merely indicating the objectives to be pursued; it must also be accompanied by the identification of the “relevant, actual, and potential impacts of the company’s activities on its own employees and on those employed in the supply chain, in relation to all the social and environmental factors specified in the directive.”
It has been said that these are third-generation information rights that have a productive function and constitute an “enabling tool for empowering workers.”⁸
Furthermore, Italian Decree 104/2022, which implements Directive 2019/1152 on transparent and predictable working conditions, contains provisions that partially anticipate European regulation and further reinforce the importance of these information obligations (Article 4), particularly in the case of using automated decision-making and monitoring systems.
To specify the content of the obligations, the decree clarifies that the information must indicate the purposes and objectives of such systems, their underlying logic and functioning, the categories of data and the parameters used to program or train them, including the mechanisms for evaluating individuals, the control measures for automated decisions, any corrective processes, and the metrics used to measure those parameters, as well as the potentially discriminatory impacts of the metrics themselves.
The GDPR (Articles 22.1 and 22.4) also requires that, in the case of automated decision-making systems, the obligations must include “meaningful information about both the logic behind these systems and the significance and consequences of the data processing for the data subjects, in order to ensure fair and transparent processing”.
According to the prevailing interpretation, however, this regulation still does not recognize a genuine right for individuals to obtain an explanation of the decisions made by automated systems—a right that will be explicitly provided by the so-called Platform Directive (see below).⁹
Yet, even if the collection of these rules does not explicitly establish a right for workers to obtain an explanation of decisions made by automated systems, it certainly represents a “race towards explaining the logics of these systems”¹⁰.
More generally, the rules in question provide a high degree of transparency regarding decisions, offering tools to the affected parties to evaluate the correctness of decision-making processes and to obtain useful insights in case legal remedies are sought¹¹.
Moreover, the very same provisions regarding the duty of transparency and information on algorithmic decisions are, functional to further interventions that require companies to assume more stringent responsibilities for social and environmental accountability. In fact, another common regulation provides for the activation of procedures aimed at identifying and preventing the risks resulting from the activities in question, which are relatively new in the field of labour—even though they have long been used in the field of occupational safety—and are inspired by managerial practices widely adopted in companies.
The recourse to this type of procedural regulation is motivated by the great complexity, differentiation, and variability of the phenomenon to be regulated, which tends to elude the traditional, standard, and predetermined regulatory techniques¹².
Choosing such a technique has the effect of shifting the focus of intervention from regulating the negative effects of corporate decisions to a procedure aimed at preventing or mitigating them.
More generally, assigning the competence to organize these procedures to the companies themselves has implications for their role, as it recognizes their function in contributing to a public good such as the prevention and safeguarding of human rights and the environment¹³.
This is a rather significant choice regarding companies’ activities and responsibilities, but it is not without its criticisms—particularly when it entrusts the performance of this public function to the autonomous evaluation of the companies themselves, without providing for specific public rules and preventive controls. There is also the risk noted by commentators and public authorities that this obligation may be reduced to forms of self-assessment that are not very meaningful, especially since data processing regulation does not even provide for certification obligations by accredited bodies (which are only recommended) nor for public disclosure¹⁴.
In this regard, it will be important how this choice is followed up not only in national regulations, but also by the actors and authorities involved in developing guidelines, codes of conduct, and organizational models to ensure both a coherent application of the regulations and the correct implementation of impact assessments¹⁵.
This regulatory approach is expressed already in the European Commission’s White Paper on Artificial Intelligence of February 2020 and was reiterated by the European Parliament in the same year, as well as in the framework agreement of European social partners from June 2020.
I also recall that on September 5 the European Commission signed the Council of Europe’s framework convention on artificial intelligence, which is the first internationally binding agreement on AI in line with the EU AI Regulation 2024/1689.
The common approach of both is to ensure that AI systems are compatible with human rights, democracy, and the rule of law¹⁶.
4. Risk Analysis and Impact Assessment: Connections and Overlaps between the GDPR and the AI Act
Subsequent regulations have implemented this method in different ways, which present numerous overlaps and mutual interferences¹⁷. In this regard, the relationship between the GDPR and the Artificial Intelligence Regulation is significant.
The connection between these two regulatory areas arises from the fact that the analysis and processing of large quantities of data are fundamental to the functioning of artificial intelligence systems¹⁸. It is precisely this connection that motivated the European legislator—and subsequently the national legislators—to intervene primarily to ensure the protection of individuals’ privacy from the actions of automated systems. At the same time, they recognized that the protection of personal data constitutes “a precondition for the full enjoyment of other fundamental rights”¹⁹. Consequently, the European legal framework has acknowledged the need to establish safeguards for individuals in the face of the use of AI systems, with reference to a broad and open set of fundamental rights²⁰.
The approach between the two texts is described as one of non-belligerence²¹. In the latter text, the conditions and requirements set forth in the GDPR are recalled, and joint activities are envisaged—such as, for example, the impact assessment on fundamental rights for high-risk systems, which under certain conditions is conducted by the operator alongside the data protection impact assessment. However, there are also interferences between the two regulations, which present various critical aspects that have been the subject of commentators’ analyses.
Although the two regulations, despite their connections, have a fundamentally different framework—since, while in the case of the GDPR the applicability of the protective provisions is conditioned by the nature of the personal data, in the AI Regulation the determining element is the type of algorithm executed by the automated process (which must fall within the broad definition of an AI system)—this difference is significant for the risk assessment and management model. In the context of the GDPR, it is the right to data protection that qualifies the safeguard and the identity of the data subject, while also ensuring the broader guarantee of an individual’s fundamental rights; as will be seen, this influences the modalities of risk assessment²².
The risk assessment required by the AI Regulation does not concern the protection of a specific right or a single data subject, but rather the protection against adverse events affecting all the fundamental rights enshrined in international charters, including the health and safety of individuals. Hence, there is a diversity of consequences, clearly observable in the field of labour relations, whereby an employer’s decision originating from the use of AI will fall within the scope of the GDPR only if it involves the use of information that can be linked to an identifiable worker²³. But, it is very likely that similar employer decisions based on the use of AI systems will also include profiling of the personal characteristics of the data subject, thereby triggering the application of the GDPR.
Another point of contact between the two regulations is that AI typically processes a large volume of data, the quality and reliability of which are decisive for the proper functioning of the system in ensuring respect for rights (see below). The AI Regulation does not address this aspect of data collection and processing, which is intended to remain governed by the GDPR²⁴.
Both regulations adopt a risk-based approach in accordance with the method indicated by the Union’s guidelines, but with some significant differences. The underlying approach in applying the method is different: while the GDPR places responsibility on the data controller to manage risk in accordance with the specific characteristics of each case, the AI Regulation categorizes AI systems based on the severity of the risk, thereby establishing the conditions to be applied to prevent and mitigate risks. This is a solution which, although formally reassuring, may prove to be excessively rigid given the rapid evolution of these technologies—so much so that the Regulation itself provides that such categorization be subject to periodic review²⁵.
In reality, the definition of the various risk categories has been quite controversial and has been the subject of a compromise that has raised many doubts.
Further differences between these texts concern the parties responsible for risk assessment, the content of the information obligations to be provided to stakeholders, their involvement in the related procedures, as well as the sanctioning aspects and the remedies available to the affected parties. The differences and discrepancies identified between the two instruments—which are repeated in other interventions—signal the need for better coordination among the disciplines in an overall perspective of system simplification.
This need is particularly evident with regard to the sources applicable to the digital sector, which have diversified considerably in recent years, with AI being only the latest episode²⁷. The identification of the subject responsible for activating the procedures for risk identification and assessment, as provided by the regulations, is of great importance given the significance of these procedures in preventing adverse events affecting individuals’ rights and in safeguarding the environment. Here, the indications provided by the various regulations exhibit clear discrepancies among themselves.
5. The Different Competences for Risk Assessment: Coordination among Regulations
The Artificial Intelligence Regulation designates the provider of an AI system as the active party in this matter—as well as in other aspects of regulation—for example, in identifying the risk class of the system and specifying the technical methods for its use. This assignment of tasks follows from the design of the AI Regulation, which is built on the model of product safety regulation. It presents several problems in terms of adequacy and functionality, because it may undermine the procedure’s ability to capture the specific risks inherent to the different production contexts in which the system is intended to operate—contexts of which only the user (deployer) is aware.
This inadequacy is confirmed by the fact that only by taking into account the specific characteristics of the context known to the deployer can that party—and no one else—take the prevention and mitigation measures required by the regulation. A further confirmation of this argument is found in European health and safety rules, mirrored by our national legislation, according to which the sole party responsible for risk assessment—and for the measures needed to prevent and mitigate such risks—is the employer, an exclusive competence for which no derogation is provided under either European or national law.
Moreover, under the GDPR, in cases where the use of an AI system involves the processing of personal data, it is the data controller—that is, the user—who must manage and assess the emerging risks for the data owner. It is unclear why this party is deemed competent and obliged to carry out the impact assessment concerning data protection, rather than for assessing risks related to the protection of fundamental rights.
The need for coordination between these two texts (GDPR and AI Regulation) is acknowledged in the AI Regulation itself, in Articles 73 and 1.5, which provide that the deployer may use the information received from the provider to fulfill the obligation to conduct the impact assessment required by Article 35 of the GDPR. This coordination is significant for the safeguards established and is provided for in the AI Regulation since the information received from the employer-user is relevant for the protection of the rights of workers potentially exposed to harm from the use of AI algorithms.
Such a discrepancy does not arise in the Due Diligence Directive, which is not affected by the dependence on product safety regulations, but rather reflects the Union’s orientation to require companies to take the necessary measures to prevent their own activities and those of their supply chains from violating social rights and environmental integrity. In line with this perspective—one inspired by the sustainable development objectives adopted by the Union—this directive imposes on the employer, and not on other parties, the duty both to assess the risks in question and to implement the measures needed to prevent and mitigate infringements of the fundamental rights of workers and the environment, irrespective of whether the tools used by the employer include automated decision-making mechanisms based on artificial intelligence.
The problems resulting from the lack of coordination between the GDPR and the AI Regulation on this point can be resolved by recognizing the competence and the obligation of the employer-deployer to assess the risks to fundamental rights in accordance with the provisions of the Due Diligence Directive. This directive should, in any case, take precedence over the AI Regulation, because it is a specialized instrument specifically aimed at ensuring the protection of labour and environmental rights pursued by the European legislator. It remains to be verified whether the provider’s instructions to the employer-deployer, based on its own risk assessment, contain elements that can be usefully integrated with those emerging from the employer’s risk assessment carried out in fulfillment of the Due Diligence Directive’s obligations.
It is not easy to identify reliable criteria to resolve this issue and to define any potential integration between the different risk assessments. This matter should ideally be addressed not only by the national legislator during the transposition of the Due Diligence Directive but even more so by the new European legislature as part of a general effort to coordinate the disciplines in question.
Similar issues regarding the risk assessment procedure arise in the relationship between the Due Diligence regulation and the recent Directive on the platform work. This latter directive underwent a rather contentious negotiation process, which in its final version introduced several improvements compared to the Commission’s original proposal. Some of these specifically concern the regulation of algorithmic management of labour relations—a practice that has become increasingly widespread in companies—and to which the second part of the directive is devoted.
6. Scope and Diversity of Risk Assessment Techniques
In this regard, particular attention should be paid to issues that have so far received little in‐depth analysis concerning the techniques to be followed for identifying and assessing risk as provided by the relevant regulations.
I refer to the fact that applying these techniques to the issues covered by these regulations presents characteristics that are different and more complex than those employed in procedures used to assess risks in other contexts—from the criteria required to evaluate workplace accident risks to those related to the risk of committing predicate offenses under Italian Law 231/2001. The content of these procedures and the obligations imposed on companies presuppose specific expertise on the part of corporate management, especially since the information and data produced can also be relevant in subsequent disputes.
Concerns have been raised by companies, noting that these operations entail significant costs and complex analyses, with the risk that in the absence of European Commission guidelines excessively discretionary practices may be adopted, thereby offering insufficient protection to the companies themselves.
The greater complexity of the procedures provided by the directives under review compared to those previously experimented with can be explained by the fact that they reflect fundamental differences between these regulations.
The first, general difference derives from the characteristics of artificial intelligence systems, which operate according to statistical processes and associations that are alien to the logics of human reasoning. Moreover, they exhibit exceptional variability in their operating criteria as a consequence of the evolution of these technologies, which also influences the potential risks they produce. This intrinsic variability of technological risk factors renders impact assessments precarious, requiring necessary repeated adjustments over time primarily in relation to the evolution of these factors (31).
In addition to this underlying diversity, there are specific differences. In procedures provided by the regulations on occupational injuries and by Law 231/2001, risk assessment concerns events defined by the rules—for example, the occurrence of workplace accidents and predicate offenses resulting from company activities. Furthermore, the analysis can rely on past data and experiences to deduce both the probability of an event occurring and its likely severity, which are the two main elements upon which traditional risk assessments are based.
The risk assessment required by the European directives under review—both the CSRD and the Due Diligence Directive—covers a very broad and not always clearly defined range of activities, because it concerns not only those of the parent company but also its supply chain and all actions that may potentially harm the fundamental rights of workers and the environment. In both cases, the novelty of these assessments is such that they cannot normally rely on significant past experience as is possible for other risks mentioned. Therefore, the procedures in question and the subsequent controls must look not only backward—to information and data from the past—but also, and above all, forward, by making predictions about the future and then comparing those predictions with what actually happens in subsequent periods, either to verify or to refute them (32).
Furthermore, while the assessment of social risks concerns events identified by the rules—albeit in great number and detail, since these are the main international and national regulations on the fundamental rights of workers—the assessment of environmental risks, concerns events that are not entirely defined. In fact, the listing of the rights to be protected provided by the regulation, and further detailed by the European Sustainability Reporting Standards (ESRS) contained in the annexes to the Delegated Regulation of July 31, 2023 (2023/2772), is not closed but open-ended (33).
This is confirmed by the preambles to the CSRD, which repeatedly stress the need to adopt a non-restrictive and non-prescriptive concept of both social rights and environmental obligations. For example, preambles 41, 43, and 45 indicate the duty to take into account the international guiding principles of the OECD and the UN. Similarly, they mention the obligation to consider both the general EU environmental legislation and that concerning the climate—even if not directly related to sustainability (preambles 41 and 42).
The open and non-standardized nature of environmental risks reflects the fact that they are linked to variations in local and global contexts that are not easily predictable. Indeed, they are often characterized by radical uncertainty that cannot be overcome with the usual statistical predictive techniques, as unfortunately evidenced by recent experiences of exceptional and dramatic climatic events (34).
For this reason, the identification and assessment of environmental risks provided for by both the GDPR and the CSRD must be dynamic in a sense different from the usual—that is, they must not only recur over time but also be continually reviewed based on subsequent verifications.
Article 15 of the Due Diligence Directive specifies that a company “must periodically conduct an assessment of its own activities and measures, as well as those of its supply chain, in order to evaluate their implementation and to monitor the adequacy and effectiveness of the measures for identifying, preventing, mitigating, halting, and minimizing the magnitude of negative impacts.” It also establishes that the assessment must be carried out after every significant change, in any case at least every 12 months, and whenever there are well-founded reasons for new risks to emerge.
To ensure the effective use of these risk assessment procedures, the same directive provides that they be carried out by company officials endowed with various competencies and responsibilities, capable of forming judgments adequate to the complexity of the risks.
For the same purpose of ensuring the completeness and reliability of the procedure, Article 13 of the Due Diligence Directive stipulates that all risk management systems be developed following consultation with the company’s employees and their representatives. And it specifies that such consultations must be conducted at all stages of the Due Diligence process: from the collection of useful information for risk mapping, to the development of action plans for prevention, to any decision to suspend or terminate a relationship with business partners, to the adoption of the measures necessary to remedy negative effects, and to the development of indicators for monitoring (35).
This is a case in which the aforementioned directive goes beyond the routine indications present in other regulations because, in addition to providing for the involvement of stakeholders throughout the Due Diligence process, it specifies that the consultation must establish a meaningful dialogue with the stakeholders; and that the company must present its approach—even in the long term—by outlining the procedures for integrating the duty of due diligence into corporate policies, including the measures for verifying compliance with the code of conduct, as well as for extending its application to business partners.
7. Limits of Risk Assessment. Mitigation or Elimination of Risk?
The complexities of the risk assessment processes described above cast doubt on their ability to predict harmful events and indicate that they are not always adequate to control the impacts of new technologies. Much will depend on the measures adopted by companies—starting with the procedures implemented to conduct impact analysis and assessment.
The difficulty of this task requires that companies engage personnel with the appropriate expertise and authority, supplemented by independent experts to counterbalance risks of self-reference. In this regard, the provision in the so-called Platform Directive (Article 10, second subparagraph) is significant, as it requires the participation of workers’ representatives in these assessment procedures and obliges the platform company to provide sufficient human resources endowed with the necessary skills, training, and authority to perform this function.
The same uncertainty affecting risk assessment processes is reflected in the measures to be implemented to counteract the identified risks and to remedy their negative impacts. Concerns have been raised that these regulations—particularly those not originally designed for labour—may stop at merely mitigating such risks without establishing their elimination. In fact, the textual indications of these norms are once again not fully aligned or entirely clear regarding the priorities to be pursued.
A clear indication is provided by Recital 53 of the Due Diligence Directive, which governs the consequences of ineffective prevention by stipulating that if negative effects on human rights and the environment occur, the company must adopt adequate measures to halt them and, if this is not possible, must minimize their magnitude. Moreover, the company should periodically review the circumstances that prevented it from halting the negative impact and assess whether it is possible to stop it.
The text of the directive provides more detailed indications—albeit with some discrepancies among the various provisions. Article 11 reproduces the provision of Recital 53, establishing that Member States shall ensure that companies “adopt adequate measures in accordance with Article 9 and this Article to halt the actual negative impacts that have been identified or that should have been identified pursuant to Article 8 (i.e., with the impact procedures provided therein).” Article 9 specifies that if it is not possible to simultaneously and completely prevent, mitigate, halt, or minimize all identified negative impacts, Member States shall ensure that companies prioritize the identified negative impacts in accordance with Article 8 (i.e. taking into account the relevant risk factors based on specific risk mappings).
Article 10, after requiring companies to adopt adequate measures in line with the previous provisions to prevent the potential negative impacts identified, adds that if prevention is not possible or immediate, they must take measures “to sufficiently mitigate them.”
Detailed provisions on this matter can be found in the so-called Platform Directive, whose Article 10, entitled “Human Oversight of Automated Monitoring Systems,” requires that platforms regularly conduct an impact assessment of the decisions made or supported by such systems and, if a risk of discrimination emerges or if a worker’s right is violated, that the platform adopt the necessary measures—including, if necessary, modifying the system or ceasing its use.
Less uniform and more divergent are the provisions of the other regulations under consideration. Article 35 of the GDPR requires, regarding the measures to be taken following the impact assessment procedure, that an evaluation be made of the “necessity and proportionality of the processing in relation to the purposes.” This is a generic formulation with uncertain implications; it has been interpreted to mean that it would impose on the employer an evaluation of the organizational process aimed at eliminating risk, and that it would permit the adoption of mitigation measures only in the absence of less intrusive solutions for the rights involved.
Moreover, non-coinciding indications also emerge from the provisions of Article 9, paragraphs 3 and 5, of the AI Act. According to the former, the risks that the provider’s measures must address are only those that can be reasonably mitigated or eliminated. The latter stipulates that the risk management measures must be such that the residual risk is considered acceptable. Additionally, it requires that the provider “guarantee the elimination or reduction of the risks identified and assessed to the extent technically feasible” and, only where appropriate, “establish adequate measures to mitigate and control those risks that cannot be eliminated.” ⁸
8. Interpretative Clarifications and a Plurality of Instruments for Human Control. Participatory Impact Assessment
These textual variations, which undermine the effectiveness of the rules mentioned above, reflect the difficult compromises reached during the legislative process of the directives, which in the AI Regulation led to assigning the obligation for impact assessment solely to the provider, thereby excluding the deployer.
In this regard, proposals have been advanced that, not without sound arguments, have sought to interpret these formulations by aligning them with a preventive logic and with the sequence of objectives indicated by the Due Diligence and Platform Directives mentioned above; thus aligning them with the clear priority established by the right to workplace safety, according to the general principle set out in Article 2087 of the Civil Code.
It remains to be seen whether, on these controversial aspects, clarifications will be provided—as would be desirable—by national regulations or by European authorities. A clarifying intervention might come from the national Data Protection Agencies, which are assigned by GDPR-related legislation the task of interpreting its provisions in cases of dispute, as has indeed repeatedly occurred.
The analysis conducted so far of European regulations on risk assessment has confirmed the usefulness of the techniques they employ; however, it has also highlighted their limitations in terms of reliability.
In order to exercise the human control of artificial intelligence envisaged by the European texts, these procedures must therefore be accompanied by the other instruments mentioned above, ensuring that they intervene from the very early stages of the design of intelligent systems: from the analysis of the data that feed their operation to avoid the presence of biases and distorting elements, to the periodic verification of the systems’ functioning to ascertain any changes over time, and finally to the oversight of decisions concerning labour relations in order to identify any discriminatory impacts.
More generally, the entire set of such instruments should be aimed at implementing “proactive, anticipatory and integrated forms of precautionary governance that evolve in conjunction with the technology”².
It must not be forgotten that, as European trade unions remind us, human control over technologies presupposes that humans are involved in their governance—and not merely informed—so that they are enabled to understand the operating conditions of the technologies, to act as critical agents within the system, and to exercise “meaningful participation”³.
In conclusion, while the regulations under review open up new and potentially virtuous perspectives for the management of companies and labour relations, their full implications can only be appreciated after thorough analyses and long-term experimentation. In any case, their practical application in national legal systems is likely to present significant interpretative challenges for the operators tasked with implementing them.
For example, among many issues, it will be necessary to clarify what is meant by “adequate measures” for eliminating or minimizing negative effects, and to define what constitutes the remediation of the negative impact referred to in Article 12 of the Due Diligence Directive.
The notion of remediation is provided in Article 2(f) of the same directive, as well as in Recital 58, which defines it as “the restoration of an equivalent situation or one as close as possible to that which would have existed if the negative impact had not occurred, including financial or non-financial compensation to the affected person and, where appropriate, the reimbursement of the costs incurred by public authorities for any corrective measures that may be necessary.”⁴
The clarification and resolution of these issues is a task that falls to the Member States and will require difficult choices, opening both old and new controversies—particularly regarding the selection of remedial measures or of specific forms of compliance.
In cases of violations of workers’ fundamental rights, the indications provided by the directive, as derived from the recitals, should encourage the remediation of the violation by the company itself in the manner most appropriate to remedy its adverse consequences for workers. These indications are especially significant when violations of social rights occur serially and continuously—for example, the failure to ensure gender equality in its various aspects, starting with remuneration.
In this regard, the regulation under review should be coordinated with the provisions of Directive 970/2023 on equal pay between men and women (Article 10), which stipulates that when wage information reveals an average pay gap between men and women exceeding 5% and the employer does not justify it based on objective criteria, the employer is obliged to conduct a joint evaluation of pay levels with the workers’ representatives.
This is an innovative provision as it requires that the impact assessment technique be implemented not unilaterally by the employer but with the participation of workers’ representatives, and aimed at the same objective as the procedures examined above—that is, to prevent and correct unjustified wage disparities and to take appropriate measures to address them⁴⁵.
At first glance, such a provision has the potential to be applied beyond the specific case provided by the law. As mentioned, “it could provide a reference model for countering discrimination in the context of automated systems, to be implemented in the form of a periodic and participatory assessment of their impact on working conditions”⁴⁶.
Such an assessment should be aimed at verifying whether the differences in treatment—here, between genders—are due to factors irrelevant to job performance and that are potentially discriminatory.
9. The Evolution of European Regulations: From the Due Diligence Directive to the Platform Work Directive
The regulations analyzed so far share the common objective of establishing rules governing the activities of companies, which have significant implications for various aspects of labour relations—from the processing of data concerning workers to protect their right to privacy, to the employer’s obligation to provide information and transparency regarding business activities that affect working conditions, and to the duty to ensure that AI is used in a manner that respects human rights.
The so-called Due Diligence Directive occupies a central role in this regulatory framework due to the novelty and breadth of its contents, which integrate important provisions from previous instruments. These include, for example, procedures for assessing the risks of corporate activities aimed at preventing violations of human rights and environmental harm, as well as the establishment of specific due diligence obligations that require companies to reinforce the observance of these rights.
The importance of this directive has been recognized by all commentators, myself included, who have noted that it is the instrument most capable of changing the managerial practices of European companies, compelling them to contribute to the achievement of sustainability and a just transition.
While further in‐depth analysis of its provisions is needed, I would like to emphasize that the way in which these measures will be implemented will also determine if and how the “rules” on artificial intelligence and data protection will affect the operation of European companies and the working and living conditions of workers.
10. The Innovations of the Digital Platform Work Directive
The most recent Directive on the Improvement of Working Conditions through Digital Platforms (2024/2831) concludes the intense regulatory review of the Von der Leyen presidency and is the instrument that has the most direct impact on labour relations. The importance of this intervention is demonstrated, among other things, by the lengthy drafting process spanning the entire legislative term, culminating in its final approval on 23 October 2024. The approved version deserves particular attention because it contains significant innovations that, in my view, represent improvements compared to the original 2021 proposal.
The first part of the Directive, concerning the criteria for qualifying employment relationships activated through these platforms (the Directive applies only to platforms that organize both dependent and independent work), encountered the greatest resistance even from major countries such as France and Germany. The main concern was that the elements indicated in the first version of the Directive—deemed relevant to trigger a presumption for qualification—might lead to a widespread transformation of gig economy work into subordinate employment relationships, a transformation on which Member States’ jurisprudence has expressed divergent opinions, often recognizing such work as independent.
I recall that the use of the qualification presumption raised doubts from the outset regarding its effectiveness, as noted by many commentators and by European trade unions⁴⁷.
I devote some reflections—expanding on arguments developed elsewhere⁴⁸—to the second part of the Directive, which concerns the implications of using these digital platforms that are destined to assume a general significance for the future regulation of labour relations.
The so-called Platform Directive also introduces various innovations in regulatory techniques, partly similar and partly different from those adopted by the directives discussed so far, but all inspired by the goal of ensuring human control over the use of these platforms in managing labour relations to prevent it from undermining workers’ fundamental rights.
The time elapsed between the first version of the Directive and the current one has allowed for the integration of GDPR on points that experience and critical remarks have identified as deficient. This represents a virtuous case of coordination/integration among the regulations.
Thus, the regulation in question provides a comprehensive framework for regulating platforms and for safeguarding workers and their representatives. However, its limitation lies in its application being confined to the activities of platforms that offer labour management services through electronic means on behalf of clients (Article 2 of the Directive). The activity that gave rise to the European legislative intervention—and to various national laws—is that of platforms for the delivery of goods and services to the home; however, the services provided by such platforms have expanded and diversified into new organizational models.
This evolution is acknowledged in various recitals of the Directive, particularly Recital 19, which mentions cases in which the platform offers services whose recipient is the platform itself or a commercial entity within the same corporate group.
In light of this evolution, European trade unions have repeatedly stressed the need for a European legislative intervention that, going beyond the original scope of this regulation, extends its coverage to all workers subject to the algorithmic management of digital platforms in any organized form—even those operating within companies—recognizing that such management entails the same needs for safeguarding workers’ rights.⁴⁹
Article 8 establishes that the processing of personal data by the platform may pose a high risk to the rights and freedoms of individuals; consequently, it triggers the obligation for an ex-ante impact assessment pursuant to Article 35 of the GDPR, also requiring the platform to seek the opinion of both the individuals who work through it and their representatives, and to provide them with the outcome of the assessment.
A further integration with the GDPR is provided by Article 7 of the Directive, which adds a prohibition on processing a series of categories of personal data potentially dangerous for workers and their rights, overriding the exceptions permitted by the GDPR—such as data concerning individuals’ emotional and psychological states, private conversations (including exchanges with trade union representatives), personal data used to predict the exercise of fundamental rights, data used to infer racial or ethnic origin, migrant status, political opinions, biometric data, etc.
Article 7 further specifies that these provisions apply not only when automated systems make the decisions in question, but also when they are used to support human decisions.
Article 10 of the Directive also prohibits decisions aimed at limiting, suspending, or terminating the employment relationship—or any other decision with a similar effect—that are not made by a human being; here too, the prohibition excludes the derogations provided by the GDPR, which allow “contractual necessity clauses” (Article 22).
Furthermore, the rule prohibits the use of automated systems that exert undue pressure on workers (Article 12.3) or create risks to their physical or mental safety and health. This prohibition, moreover, applies only to subordinate workers, with a different approach from other parts of the Directive that extend safeguards to self-employed workers or those without a specific contract.
The tendency, present in other recent provisions—both European and national—to extend safeguards beyond the traditional scope of subordinate work has encountered a limitation that remains even in the final version of the text, despite the European Parliament’s amendment requests, and reflects the resistance—also evident in case law from the Court of Justice of the European Union—to apply collective rights to self-employed work.⁵¹
Recital 39 of the Directive also notes that in the case of platform work, it cannot be assumed that workers’ consent to the processing of their personal data is given freely, given the power imbalance with the employer. It concludes that therefore platforms should not process these individuals’ data on the basis of the presence of their consent, as is permitted by the GDPR.
A similar observation, already expressed by GDPR commentators,⁵² is not, however, reflected in the new regulation of the Directive.
11. Human Control over Algorithmic Management
The procedural safeguards provided even in the final version of the text focus on the obligations of information and transparency imposed on the platform company. These represent a common procedural toolkit in many recent regulations on the subject.
The novelty, however, is that the current version overcomes the ambiguities and uncertainties of these regulations by extending the obligation to “all types of decisions made or supported by automated systems” (Article 7, paragraph 3).
Similarly, significant meaning is attributed to the information obligations by specifying that they must include not only the fact that automated monitoring or decision-making systems are in use, but also the categories of data and actions upon which these mechanisms operate, the objective of the monitoring, the recipients of the processed data, the parameters used in the decision-making process and their relative importance, and the reasons for decisions affecting key aspects of the employment relationship—such as remuneration and termination.
It is further specified that such information must be provided in writing, prior to the use of the relevant mechanisms, and must include all of their characteristics. Moreover, the information must be made available not only to the individual worker concerned but also to trade union representatives (Article 9). Additional information must be provided to representatives and competent authorities regarding the quantity and nature of the work performed on the platforms (Article 17).
The Directive also requires a direct monitoring obligation to assess the impact of the decisions made or supported by the automated systems employed by the platforms on the working conditions of those involved, specifying that the platform must ensure sufficient human resources to effectively oversee and evaluate the impact, and that such evaluation must be communicated to workers’ representatives (Article 10).
This provision confirms a choice present in other European regulations intended to complete the sequence of activities —information, assessment, oversight, and monitoring—designed to ensure human control over decisions made by intelligent machines.
However, the scope of these collective rights remains limited—as in other European regulations—to information and consultation, even though the Commission’s proposal indicates that these rights are meant to promote social dialogue on algorithmic management (Article 25) ⁵³. In the same direction, the Directive, incorporating an amendment from the European Parliament, has introduced provisions aimed at encouraging the exercise of collective bargaining rights and rights related to algorithmic management (Article 25).
Nevertheless, these indications remain weak and insufficient to support trade union action in dealing with platforms and the novel dynamics of the digital economy. An attempt to establish a coordinated management of digitalization in labour relations was made with the European Framework Agreement on Digitalization of June 2020, which established a procedure for involving workers’ representatives in all stages of the design, introduction, and implementation of automated digital tools related to labour relations. However, this agreement met with distancing from some European employer associations and did not receive sufficient support at the national level, ultimately failing to produce effective follow-up⁵⁴.
Similarly, national collective bargaining has shown difficulties—except for a few exceptions—in addressing the issues raised by the new digital economy⁵⁵.
A provision of particular importance is found in the latest version of Article 11, which is titled “Human Review.”
12. The Request for Review of Decisions
In the initial version, it was established that affected workers could submit a request for review of the algorithmic decision, to which the platform company was required to respond. Such a regulation—providing only for the individual’s relationship with the platform—could not be considered adequate to effectively implement the principle of human control ⁵⁶.
The new version of the text integrates the regulation by establishing that the request for review may be submitted not only by the worker but also by trade union representatives, and that the platform company must respond to the review request both to the worker and to their representatives.
Furthermore, it stipulates that it is not sufficient for the platform company to merely provide any rationale for the decisions made by the automated systems regarding the employment relationship; rather, it must provide “a justification in a transparent and understandable manner, using simple and clear language” (Article 11.1).
Moreover, if the decision violates workers’ rights, the platform is required to rectify it without delay and, if that is not possible, to offer adequate compensation for the damage; in any case, it must take the necessary measures—including, if applicable, modifying the automated decision-making system or ceasing its use—in order to prevent similar decisions in the future (Article 11.3).
A similar regulation is contained in the AI Act (Articles 86 and 13), which provides that the affected person has the right to request clear and meaningful explanations from the operator regarding the role of AI in the process and the key elements of the decision.
The importance of the obligation to justify decisions made by automated systems can be appreciated by considering the difficulty of ascertaining the underlying logic of such decisions, the presence of discriminatory consequences and potential adverse effects on workers’ rights.
13. Obligation to Justify and Ascertain Algorithmic Discrimination
Issues concerning the risk of discrimination raised by AI systems have long been on the radar of commentators, and they are testing the traditional techniques of anti-discrimination law—even those that have been long established. It has been observed that this complex subject is not explicitly addressed by recent European regulations, despite the fact that the Commission’s positions include references to the increased risks of discrimination resulting from the use of algorithms.
Nevertheless, the regulations examined so far to control the decisions made by AI systems do offer several useful instruments to facilitate the identification and proof of any discrimination. One tool aimed at ensuring the anti-discriminatory use of AI is the verification of the essential elements of how these systems operate—namely, on one hand, the instructions given to them and their operating models, and on the other, the reliability and quality of the data they use—to detect any biases that might reproduce distortions and discriminatory practices in matters of labour and personnel relations.
It is significant that the GDPR guidelines recommend that companies verify the quality and accuracy of their data in order to prevent distorted outcomes and to implement measures aimed at preventing errors, inaccuracies, or discrimination.⁵⁷
However, this type of ex-ante control, though of great importance, may not be sufficient to prevent discriminatory outcomes when dealing with non-deterministic AI systems endowed with a degree of autonomous functioning. Therefore, it is necessary to complement these initial controls with periodic ongoing verifications of the AI processes and with ex-post investigations designed to ascertain the presence of factors that could produce discriminatory effects.
A similar approach is adopted by the aforementioned European Framework Agreement on Digitalization, whose procedural tools allow workers’ representatives to exercise oversight over the entire AI decision-making process. This represents the most pervasive intervention recognized for social partners on this issue, even though it still suffers from the limitations of weak participation and, in other respects, its application is conditioned by the opposition of European employer organizations, so that it has not been applied at the national level.
Furthermore, even such a detailed intervention by social partners may influence the impact of AI systems on labour relations from the outside, but it is not guaranteed that it can verify the correctness of the decisions and their potential discriminatory character.
The difficulty of verification using traditional techniques stems from the way algorithmic rationality operates—based on probabilistic correlations rather than on causal relationships. Moreover, in developing these systems, the elements of correlation intersect in “haphazard” artificial patterns and are recombined in processes that make it impossible to determine if and when a single element derived from a particular personal datum has influenced the system’s outcome.⁵⁸
Consequently, it is not ascertainable whether and under what conditions an algorithmic decision has been determined by a discriminatory factor, which is the criterion for establishing the existence of direct discrimination. It may be relatively easier to ascertain indirect discrimination, though this entails different consequences and limitations compared to direct discrimination.⁵⁹
Indirect support for countering algorithmic discrimination can be derived from the regulations related to impact assessment procedures and to the information rights regarding these automated decision-making mechanisms. As noted above, the data and informational elements derived from these procedures can be used to establish prima facie evidence of discrimination, thereby opening the door to strategic litigation by affected individuals. Similarly, the competent data protection authorities can facilitate data subjects’ access to documents that would not otherwise be available.
Moreover, case law has recorded a series of evolutions that have facilitated the identification of cases of not only indirect discrimination but also direct discrimination, thereby expanding the scope of the situations, the protected factors, and the individuals concerned.⁶⁰
In this evolving context, a useful contribution can come from the provision in the Platform Directive (as well as a similar provision in the AI Act) that requires a justification for the decision resulting from the use of automated systems, aimed at clarifying the system’s operating methods—that is, the procedures followed to arrive at the decision.
The right of affected individuals to obtain an explanation of decisions detrimental to them is in line with the aforementioned provisions, serving as a tool to facilitate the ascertainment of any discrimination. The eventual refusal to provide the requested explanation, or the presentation of an inadequate explanation, cannot go without consequences. If one does not conclude that the decision is invalid due to inadequate justification, the alternative is to consider that the absence of such justification reverses the burden of proof, thereby placing on the platform company the duty to provide non-discriminatory reasons for the decision.
Among the regulatory provisions that shift the burden of proof, one can refer to the provision in Directive 2023/970 (Article 18) on pay transparency, which provides for a complete reversal of the burden in cases of violations of the employer’s transparency obligations, as well as the possibility of a disclosure order in court.
Of course, it is always possible for the platform company to adopt the solution indicated by the Directive—that is, to rectify the unjustified decision and, if that is not possible, to offer adequate compensation.
14. Continuity and Challenges of the European Social Agenda
The body of legislation comprising recent directives and the Artificial Intelligence Regulation—more so than other initiatives of the Von der Leyen presidency—has given unprecedented momentum to the deepening of the European social acquis. A decisive push came from the extraordinary events of recent years and from the pervasive impact of artificial intelligence on all the structures of the economy and on social life itself.
The complexity of this regulatory framework poses a formidable challenge in terms of both understanding and implementation for all European and national institutions and stakeholders.
A novelty emphasized in the preceding pages is that this legislation has, for the first time, directly invoked the responsibility of companies in regulating and controlling these new technologies and in pursuing social and environmental sustainability objectives. The implications of these regulatory interventions—as well as those economic and financial measures that have accompanied them—are far from fully manifested; for this reason, they are likely to engage the analytical and innovative capacities of experts—not only legal scholars, but also practitioners and policymakers—for a long time to come.
The programs of the new Von der Leyen Commission, as well as the reports prepared by Enrico Letta and Mario Draghi, are based on the continuity of the policies initiated in the previous term, including in the social field.
A first condition for ensuring that such continuity is effective is to address the controversial issues left unresolved by the legislation adopted so far, with the aim of clarifying critical points and harmonizing the protection of workers against the negative effects of algorithmic management of labour relations.
Such an effort should be supported not only by national implementing legislation—which, on many issues, leaves wide margins for discretion—but primarily by European initiatives, both in the form of legislative texts and guidelines, and through the intervention of various authorities entrusted with tasks of assistance and advice to the Commission, as well as the interpretation of EU regulations.
1. For example, decree 125/2024 transposing the CSRD Directive and the accompanying comments by A. Candotti and T. Treu in Commentario, edited by A. Genovese and M. Libertini, currently in publication.
2. See the reference by M. Peruzzi, Intelligenza artificiale e lavoro, Giappichelli, 2024, p. 22 ff.
3. These opportunities are highlighted by the multidisciplinary research conducted by M. Molè, W. Been, and B. Haar, Multidisciplinary Perspectives on Work and Digitalisation, in Italian Labour Law – e-Journal, 2024, issue 2.
4. Cf. A. Aloisi, “Regulating Algorithmic Management at Work in the European Union: Data Protection, Non-Discrimination and Collective Rights,” in The International Journal Comparative Labor Law and Industrial Relations, vol. 40, March 2024, p. 45.
5. J. Adams-Prassl, “What if Your Boss Was an Algorithm? Economic, Incentives, Legal Challenges and the Rise of Artificial Intelligence at Work,” in Comparative Labour Law & Policy Journal, 2019, p. 6 ff.; A. Aloisi, “Regulating Algorithmic Management,” ibid., p. 46.
6. A. Aloisi, “Regulating Algorithmic Management,” loc. cit.
7. See in general M. Peruzzi, Intelligenza artificiale e lavoro, ibid., p. 22; and M. Biasi, “Problema e sistema nella regolazione lavoristica dell’intelligenza artificiale: note preliminari,” in Federalismi, December 18, 2024, p. 176.
8. L. Zappalà, “Appunti sul linguaggio, complessità e comprensibilità del diritto del lavoro 4.0: verso una nuova proceduralizzazione dei poteri datoriali,” in Janus, no. 24, p. 23 ff., p. 37.
9. J. Adams-Prassl, “What if Your Boss Was an Algorithm,” ibid., no. 6, p. 150 ff.; S. Wachter, B. Mittelstadt, and L. Floridi, “Why a Right to Legibility of Automated Decision Making Does Not Exist in the GDPR,” International Data Privacy Law, 2017, 1.1; with additional citations; see also the considerations by S. Wachter, B. Mittelstadt, “A Right to Reasonable Inferences. Re-thinking Data Protection Law in the Age of Big Data and AI,” in Columbia Business Law Review, 2019, no. 2, p. 51 ff., which point out the importance of including in GDPR transparency the right of individuals to know what inferences are drawn from their personal data; and G. Gaudio, “Algorithmic Management and the Burden of Proof: In Search of the Material Truth Hidden within the Algorithm,” Labour and Law Issues, 2020, p. 29 ff.
10. Equally relevant and innovative compared to other provisions on the subject is that this regulation explicitly requires that all such informational content be shared with workers’ representatives. L. Zappalà, Appunti sul linguaggio, ibid., p. 34 ff., according to which the set of rules on transparency and human review of automated decisions prefigures tools for constructing “a justifiable common world,” p. 39; see also A. Perulli and V. Speziale, Dieci tesi sul diritto del lavoro, Mulino, 2022, p. 69.
11. A. Aloisi, “Regulating Algorithmic Management,” ibid., p. 51; G. Gaudio, “Algorithmic Management,” ibid., p. 39 ff.; M. Peruzzi, Intelligenza artificiale, potere datoriale, e tutela del lavoro. Ragionando di tecniche di trasparenza e poteri regolativi, in Janus, no. 24, ibid., p. 84.
12. See in general P. Loi, La sicurezza, diritto e fondamento dei diritti nel rapporto di lavoro, Giappichelli, 2000, p. 47 ff.; and ID, Il rischio proporzionato nella proposta di regolamento sull’IA e i suoi effetti nel rapporto di lavoro, in Federalismi, no. 4, 2023, p. 252 ff.; also S. Villani, “Luci e ombre degli strumenti di tutela dei diritti nella architettura dell’IA Act basata su rischio,” in I nuovi confini del lavoro: la trasformazione digitale, edited by F. Lunardon and E. Menegatti, Italian Labour Law e-studies, 2024, p. 143 ff.
13. For this topic, see in general my paper L’impresa sostenibile: implicazioni per il diritto del lavoro, in T. Treu, Le trasformazioni del lavoro: discontinuità e interdipendenze, ibid., p. 225 ff.
14. See critical remarks in G. Gaudio, “Algoritmo e Management,” ibid., paragraph 4, and similar observations in the virtual roundtable cited in note 49.
15. For example, as provided by Article 6, paragraph 5, of the IA Act, which assigns the European Commission the task of establishing guidelines for the implementation of the rules with illustrative examples of both high-risk and non-high-risk AI cases.
16. The Council of Europe has long analysed the impact of such technologies on human rights, as foreseen by the European Commission, by establishing an ad hoc committee (CAHAI) to examine a possible legal framework on these aspects. See S. Villani, “Luci ed ombre degli strumenti di tutela dei diritti nell’architettura dell’IA Act basata sul rischio,” Italian Labour Law e-studies, in I nuovi confini del lavoro: la trasformazione digitale, edited by F. Lunardon and E. Menegatti, 2024, p. 153 ff.
17. The European Competitiveness Report of September 9, 2024, emphasizes that this regulatory fragmentation is particularly pronounced in the digital sector (with more than 100 specific regulations and 270 regulatory authorities) and hinders innovation and development.
18. See in particular T. Le Bonniec, Another Path for AI: Workers’ Unions and Data Protection Rights, in Italian Labour Law e Journal, 2024, pp. 116 ff., which underlines the reasons for this connection and the predominant role of data protection rules even in the management of artificial intelligence; the hegemonic role of the GDPR is also discussed by M. Peruzzi in Intelligenza artificiale e lavoro, ibid., pp. 41 ff.
19. G. Alpa, European Private Law, Giuffrè, Milan, 2016, p. 182; N. Lipari, Civil Law and Reason, Giuffrè, Milan, 2019, p. 183; F. Mollo, The Processing of Biometric Data in the IA Act: Intersections between Data Protection Regulation and the New European Framework for Artificial Intelligence, in Federalismi, issue 28/2024, pp. 114 ff., with citations from the case law of the Luxembourg and Strasbourg courts.
20. See the European Commission, Protection of Fundamental Rights in the Digital Age, 2021 Annual Report on the Implementation of the European Union Charter of Fundamental Rights, COM (2021) 819, dated 10 December 2021.
21. L. Tebano, Employer Powers and Biorhythmic Data in the Context of AI, in Federalismi, 25/2023, p. 211.
22. M. Peruzzi, Intelligenza artificiale e lavoro, ibid., p. 45.
23. M. Peruzzi, Intelligenza artificiale e lavoro, ibid., p. 65.
24. This is clarified in the accompanying report to the IA Regulation: cf. comments in P. Falletta and A. Marsano, Intelligenza artificiale e protezione dei dati personali. Il rapporto fra regolazione europea sull’intelligenza artificiale e GDPR, in Rivista Italiana di Informatica e Diritto, 6 June 2024, p. 9.
25. P. Falletta and A. Marsano, Intelligenza artificiale e protezione dei dati personali, ibid., p. 15.
26. See the remarks by S. Ciucciovino, Risorse umane e IA alla luce del regolamento UE 2024/1689. Fra norme legali, etica e codici di condotta, ibid., pp. 580 ff.
27. See, with regard to this, the indications from the study by the European Parliament Research Service dedicated to The Impact of the General Data Protection Regulation (GDPR) on Artificial Intelligence.
28. These issues are highlighted by commentators; A. Alaimo, Il regolamento sull’intelligenza artificiale. Un treno al traguardo con alcuni vagoni rimasti fermi, in Federalismi, 25/2024, p. 242; see also S. Ciucciovino, Risorse umane e intelligenza artificiale alla luce del regolamento europeo, 2024/1689. Fra norme legali, etica e codici di condotta, ibid., p. 586, which points out the prevalence of obligations on the provider.
29. A significant linkage between the texts is provided by Articles 73 and 9.5 of the IA Regulation, according to which the deployer may use the provider’s information to fulfil its obligation to conduct the impact assessment required by Article 35 of the GDPR.
30. See P. Falletta and A. Marsano, Intelligenza artificiale e protezione dei dati: il rapporto fra regolamento europeo sull’intelligenza artificiale e GDPR, ibid., pp. 7 ff.
31. See S. Villani, Luci e ombre, ibid., p. 143.
32. G. Gometz, Intelligenza artificiale, profiliazione e nuove forme di discriminazioni, in Teoria e storia del diritto privato, NS 2022, pp. 28 ff. The differences between traditional assessments of workplace safety risks and those required for the risks arising from the use of AI are also noted by C. Lazzari and P. Pascucci in Sistemi di IA, salute e sicurezza del lavoro: una sfida al modello di prevenzione aziendale – responsabilità e opportunità, forthcoming in Rivista Giurisprudenza del Lavoro, 2024, no. 4.
33. For this reason, one cannot claim—as some commentators argue—that choosing to adopt a selective approach to these rights, especially human and social rights, entails a limitation of them, given the United Nations principles that refer to all those contemplated by international sources.
34. J. Kay and M. King, Radical Uncertainty: Decision Making Beyond the Numbers, Norton & Co., London-New York, 2020, particularly parts III and V.
35. Cf. for analysis and commentary, F. Guarriello, Take Due Diligence Seriously: A Commentary on Directive 2024/1760, in Giornale Diritto del Lavoro e Relazioni Industriali, 2024, pp. 191 ff.
36. G. Gaudio, Valutazione di impatto e management algoritmico, forthcoming in RGL, 2024, no. 4, notes that where there is a risk of violating imperative labour standards, the employer should in any case be required to eliminate the violation; M. Peruzzi, Intelligenza artificiale e obblighi datoriali di tutela del lavoratore: necessità e declinazione dell’approccio risk based, in Federalismi, 18 December 2024, p. 231.
37. Thus, M. Peruzzi, Intelligenza artificiale e obblighi datoriali di tutela del lavoratore: necessità e declinazione dell’obbligo risk based, ibid.
38. See critical remarks on these limitations in A. Mantelero and M. Peruzzi, L’AI Act e la gestione del rischio nel sistema integrato delle fonti, forthcoming in Rivista di Diritto del Lavoro, no. 4, 2024.
39. M. Peruzzi, Intelligenza artificiale e obblighi datoriali di tutela, ibid.
40. Cf. analysis and examples in T. Le Bonniec, Another Path for AI Regulation, ibid., pp. 118 ff.
41. U.S. state law, as mentioned by M. Peruzzi in IA e obblighi datoriali (p. 241), is significant because it requires independent auditors to review automated candidate selection processes for potential biases.
42. Similarly, A. Ponce del Castillo, in Artificial Intelligence and Labor (p. 18 ff.), also emphasizes the long-established precautionary principle in European regulation.
43. See also A. Ponce del Castillo, Artificial Intelligence (p. 20).
44. F. Guarriello, Take Due Diligence, p. 190.
45. Cf. my remarks in La legge sulla parità di genere (Guida Lavoro, Sole 24 Ore, no. 49, December 10, 2021). See also D. Izzi, Gender Gap nel lavoro e regole di trasparenza in LD, 2024, p. 652 ff.
46. Thus, M. Peruzzi in IA e obblighi datoriali (pp. 240–242) notes that these same evaluation methods must confront the specific challenges of assessing decisions based on intelligence mechanisms that operate according to logics different from those of human rationality.
47. M. Magnani, La proposta di direttiva sul lavoro mediante piattaforma digitale, in CNEL, Rapporto sul mercato del lavoro e contrattazione collettiva, 2022, p. 265 ff.; T. Treu, Diritti e politiche sindacali nel contesto europeo in LD, 2023, p. 94 ff.; A. Perulli, Lavoro e diritti nella rivoluzione di internet in LDE, I, 2022, p. 5 ff.
48. See Guida al lavoro (November 24, 2024) and G. Smorto, A. Donini, L’approvazione della direttiva sul lavoro tramite piattaforme digitali: prima lettura in Labor Law Issues, I, p. 25 ff.; also in relation to Article 2, first paragraph, of Decree 81/2015.
49. See ETUC, Directives Needed to Make Work Safe for the Digital Age, October 23, 2024, www.etuc.org, and the comments by Stefano Gran in the virtual roundtable on the AI Act and the Platform Directive in the European regulatory space, forthcoming in RGL, April 2024.
50. See on these topics, A. Perulli and T. Treu, In tutte le sue forme e manifestazioni. Per un nuovo statuto del lavoro, Giappichelli, 2023.
51. Cf. T. Treu, Lavoro autonomo e diritti collettivi nella Unione Europea, LDE, 2023, p. 251; and the commentary on the directive by M.P. Aimo, Il management algoritmico nel lavoro tramite piattaforma in Federalismi, 25/2024, p. 227 ff., which traces this dual approach to the common legal basis of the directive, shared by all recent European measures.
52. Cf. in general, M. Peruzzi, Intelligenza artificiale e lavoro (p. 49 ff.).
53. Cf. regarding the content and limits of this “weak” participation, M.P. Aimo, Il management algoritmico nel lavoro tramite piattaforma: osservazione sulle prime regole di portata europea, Federalismi, no. 25/2024, p. 227; A. Alaimo, Il regolamento sull’intelligenza artificiale. Un treno al traguardo con alcuni vagoni rimasti fermi, Federalismi, no. 25/2024, p. 243 ff.; and M. Corti, Intelligenza artificiale e partecipazione dei lavoratori. Per un nuovo umanesimo del lavoro, DRI, special issue, p. 682.
54. See my comment in La legge sulla parità di genere (Guida Lavoro, Sole 24 Ore, no. 49, December 10, 2021).
55. Cf. M.P. Aimo, Il management algoritmico (p. 247).
56. See my critical remarks in La legge sulla parità di genere (Guida Lavoro, Sole 24 Ore, no. 49, December 10, 2021).
57. Guidelines on Automated Individual Decision Making and Profiling for the purposes of Regulation, 2026/679, at paragraph 28.
58. G. Gometz, Intelligenza artificiale, profilazione e nuove forme di discriminazioni, in Teoria e storia del diritto privato, no. 5/2022, p. 28.
59. For a discussion on the point and on the subtle distinction between cases of direct and indirect discrimination, cf. in general, M. Barbera, Discriminazioni algoritmiche e forme di discriminazione, in Labor Law Issues, 2021, no. 1, p. 12 ff.
60. Cf. the analyses and case studies in A. Aloisi, Regulating Algorithmic Management (p. 58 ff.), which points out that these instruments together reduce the need to peer inside the algorithm’s black box; see also M. Barbera, Discriminazioni algoritmiche e forme di discriminazione, ibid., particularly p. 8 ff.
61. See my considerations in Il futuro della Europa sociale: continuità e nuove sfide, Astrid Rassegna, 18/2024.
Compila il seguente form. Ti risponderemo il prima possibile!