Keywords

Introduction

Recent years have witnessed an increasing amount of research and interest in the complex matter of digital platforms, which has been examined from diverse and multiple academic perspectives. The phenomenon of platformisation, defined as the penetration of infrastructures, economic processes and governmental frameworks of digital platforms in different economic sectors and spheres of existence, as well as the reorganisation of cultural practices and imaginations around these platforms (Poell et al., 2019), still represents a complex and deeply investigated subject, with visible repercussions on many aspects of public and private life. A concrete example is the widespread diffusion of globally operating platform businesses—from Facebook to Google, to Amazon, etc.—that are becoming increasingly central to public and private life, transforming key economic sectors and spheres of life.

Following the evolution and the increasing significance of platforms and digital intermediaries, institutions have been focusing their attention and their policymaking more and more on the issue of data and digital transition, which requires digital governance to adapt to each country’s regulatory culture and capacity, as well as understanding that these structures will continue to change over time (OECD, 2019:147). In this context, at least 73 countries worldwide have adopted a digital strategy or plan (ITU, 2020:3), while another recent trend is for countries to adopt strategies tailored to specific technologies or issues, such as automation, robotics, 5G, artificial intelligence (AI) and the Internet of Things (IoT). At the EU level, the latest Proposal of Regulation (EC)2020/825 on a Single Market for Digital Services (Digital Services Act, DSA) represents the culmination of a long-standing regulatory process that, over the years, has touched on numerous issues related to digital regulation. Particular attention has been paid to the implications for democratic stability caused by the extensive power of digital platforms and intermediaries: in the document published in 2020 by the European Commission’s Joint Research Centre, “Technology and Democracy: understanding the influence of online technologies on political behaviour and decision-making”, causal connections are established between how platforms work and how fundamental rights online are impacted. The research highlights how automated newsfeeds and recommendation systems are designed to maximise the attention and engagement of users by satisfying their alleged preferences, which may mean giving relevance to polarising and misleading content.

All these subjects—data-driven performances, data-agile economies and services, content’s ranking and recommendation systems—are linked by a common component: algorithms. Understanding algorithms, their role and functions in our digital world, the possible repercussions of their being more and more at the centre of the co-production of meaning in our society, is a key challenge of our times. The algorithmic power is transversal to many areas of life, and a single lens cannot magnify all of its complexity: a multidisciplinary approach can fruitfully help understand this all-encompassing phenomenon.

On the one hand, we need to understand the meaning of the “power of algorithms”: how does it concretely work, how it is related to platform performance and what are the possible implications, especially in terms of real interference with human knowledge and perception of reality. On the other hand, we will be examining the “counteractions” set out in the European regulation: how it has been developing over time and how algorithms have become increasingly central to the debate, subsequently highlighting the importance of algorithmic transparency and accountability. We will be looking at how lawmaking—in particular the European legal framework for AI—is trying to intercept that same algorithmic power and establish new benchmarks for transparency. In conclusion, we will assess how the most recent regulation proposal addresses the matter and what are the criticalities that might occur.

Performative Intermediaries and Their Political Power

As individuals, we are deeply immersed in a “datafied reality”, and we constantly interact with and through data-driven intermediaries, the vital functions of which depend on the capacity of algorithmic components to aggregate and elaborate users’ digital traces and to transform them into means of production, both of meaning and profit. Simple actions like looking for services through a search engine, shopping online, chatting with friends and so on, all implicate a high degree of interaction with algorithmic media (Deuze, 2012). Through the processes of selecting, ranking and suggesting contents—information, services, products, friends—algorithms show the capacity to build and shape our digital reality based on our previous activities and choices within the digital space.

On these premises is built the conception of digital platforms as performative intermediaries (Bucher, 2018:1) that is, types of media relying on predictive analytics, whose nature is inherently algorithmic and that don’t just represent our many worlds but actively participate in shaping them. When we define platforms as “inherently algorithmic”, we are pointing at some patterns, relations and processes that the structure of platforms intrinsically activate. For instance, let’s consider users’ registration when first accessing a platform: a definition of one’s identity according to fixed, measurable, comparable standards. This definitory process needs to provide information that is compatible with the algorithmic logic driving the system, that is, to provide comparable and compatible data for the algorithm to process. To be “algorithm-ready”, data need to be encoded according to schemes and a fixed set of standards, to be categorised. Categorisation is a powerful semantic and political process: how categories are defined, what belongs in each one of them and who decides how to implement these categories in practice are all powerful assertions about how things are and are supposed to be (Bowker & Star, 2000; Gillespie, 2014).

Information needs to fall within a database, which is much more than a simple “collection of data”: it is an ordered space, a structured repository that streamlines management and updating of data according to fixed criteria, without which complex searches would be impossible. As Tarleton Gillespie puts it, “algorithms are inert, meaningless machines until paired with databases upon which to function” and “before results can be algorithmically provided, information must be collected, and sometimes excluded or demoted” (Gillespie, 2014:169). There’s a precise political valence—intending “politics” as ways of world-making—in what the author defines as patterns of inclusion: the subsequent phases of data collection, readying and exclusion of “less relevant” information all imply upstream decisions about which pieces of information will be indexed in the first place and which will be excepted. Since the main characteristic of algorithms is to be functionally automatic, namely, to be activated and prompted without any regular human intervention or oversight (Winner, 1978), this means that the information included in the database must be rendered into data, formalised so that algorithms can act on it automatically. To the purpose of database design and management, categorisation represents a critical process. Categories draw demarcations, help establish and confirm standards and imply by nature the exclusion of certain items, while “algorithms can be touted as automatic, the patterns of inclusion are the ones that predetermine what will or will not appear among their results” (Gillespie, 2014:173).

As information technology has converged with the nature and production of scientific knowledge, we assist the social and political process of creating an explicit, indexical memory of what is known, the making of “memory infrastructures”. In Geoffrey Bowker’s words, today we “database the world” in a way that excludes certain spaces, entities and times: “the archive, by remembering all and only a certain set of facts/discoveries/observations, consistently and actively engages in the forgetting of other sets (…) The archive’s jussive force, then, operates through being invisibly exclusionary” (Bowker, 2006:12–14).

I’ll allow myself just a brief digression on the theme of archives since the matter is too vast to be addressed here and that would fall beyond the scope of this contribution. At the same time, this reflection is nonetheless essential to understand the in-built partiality of the indexing process and the power of database infrastructural invisibility deployed on, or against, certain sets of items and subjects. I will borrow the concept of the silence of the archive (Hartman, 2008:3–4) from the historiographical reconstruction outlined by Saidiya Hartman in her essay “Venus in Two Acts”. The theme of her study is very far from what we are dealing with—her potent dissertation addresses the scarcity of African narratives of captivity and enslavement opposed to the abundance of direct testimonies of colonial ferocity. However, her concept powerfully represents the exclusionary forces mentioned also by Bowker: sometimes histories are destroyed, aren’t collected or aren’t even told—in all the cases, in our collective history something extremely significant gets lost.

Jacques Derrida reminds us of the etymology of the word “archive”, the ancient Greek arkheion: literally “the house of whom command”. In ancient Greece, the citizens who held and signified political power (the archons) “were considered to possess the right to make or represent the law. (…) it is at their home (…) that official documents are filed” (Derrida & Prenowitz, 1995:9–10). From the beginning of history, adding or excluding something from the archive has been an essentially political action, determined by the archivist and by the political context in which he lives. At times, information and proof have not been entrusted to the archive because they are not considered important enough to be preserved; at others, acts of intentional destruction have prevented the collection of those pieces of history in the archive and, as a consequence, have inhibited human awareness: “the complete archive is a myth, it is only theoretically possible. Maybe it is in some recess of Jorge Luis Borges’ Library of Babel, buried under the detailed story of the future” (Machado, 2020:12). The “digital version” of the concept of the silence of the archive depicts a situation where “some information initially collected is subsequently removed before an algorithm ever gets to it” (Gillespie, 2014:172).

On the one hand, we have large-scale information services, providers and indexes that are not comprehensive by nature, since they act as censors as well. For instance, YouTube algorithms bring down the rankings of certain videos, so they do not become “most watched” nor do they get suggested on the customised home page of new users (Gillespie, 2010). Twitter, at the same time, does not censor profanity from public tweets, but its algorithms remove them from the evaluation of “trending terms/hashtag” (Gillespie, 2012). On the other, we must acknowledge that the digitalisation of all information and the acceptance of computational tools as our primary means of expression lead to specific implications, especially when we use algorithms—as above, automated tools that don’t need human oversight—to select what is “most relevant” from a corpus of data composed of traces of our activities, likings, languages.

Algorithms, Relevance and Influence on Perception

By definition, algorithms are encoded procedures for transforming input data into the desired output, based on specified calculations. These calculations have never been simple, but today we are facing an unprecedented expansion and growth in their complexity, as billions of people daily use the services provided by algorithmic media. The algorithmic components lying behind automated newsfeeds, sharing platforms or search engines, scrape and select contents and information—according to predetermined criteria of relevance—and build up a custom configuration of each individual’s digital reality, which can influence our perception of the world not only within but also outside the digital environment. Just think of all the pieces of information we searched for but were eventually lost because they were ranked low on our Google results: at the end of the day, a complex, automatic, profit-driven mechanism drew a line between what we could easily find and access and what was to remain silent. Again, this represents an exercise of power, both in terms of control over the availability and concrete accessibility of knowledge and in terms of influence on the meaning we subsequently construct over that same knowledge.

The criteria by which algorithms determine what is relevant, how those criteria are obscured from us and how they endorse political choices about the portion of knowledge that results in being appropriate and legitimate all fall within the process that Gillespie defines evaluation of relevance (2014:168). When users perform activities that depend on automated systems, they are running a query. In computer science, this means to search a database to extract or update data that meet a certain criteria, specified in the query and based on variables and values in the source data. The search algorithm then examines a vast amount of signals to retrieve cases that match the specified criteria. Through these signals, the algorithm approximates “relevance”, and that is the most concerning aspect: “relevant” in fact is a fluid term, loaded with connotations of value and attributes that depend on the subjectivity of who’s performing that specific query. “Relevant” is not quantifiable nor measurable, and it’s a concept at the other extreme of objectivity. In Gillespie’s words, since “there is no independent metric for what actually are the most relevant search results for any given query, engineers must decide what results look “right” (…) treating quick clicks and no follow-up searches as an approximation, not of relevance exactly, but of satisfaction” (Gillespie, 2014:178, italics added). We are not getting what could be important for us but what is going to satisfy us based on our collected, layered digital traces: this is a relevant way in which our perception of the world is influenced by heteronomous criteria, implemented by algorithmic components.

As we pointed out above, the algorithm is a set of coded instructions to be followed to perform a given task; in the digital context, they are set to select and make visible “in a meaningful way” portions of the endless amount of data produced and available on the web. Therefore, classifying, ranking and recommending (or, on the contrary, reducing visibility) are the elemental activities that sustain algorithmic functioning and performance: the question that arises is then who or what has the power to establish the conditions for what can be seen and known (Bucher, 2018:3). It is not possible to interrogate the underlying criteria of algorithms: in nearly all cases, the evaluative assumptions that algorithms are potentially requested to make remain hidden, in accordance with the precise wishes of platform owners. Being unambiguous or accurate about the working of algorithms would likely allow competitors to replicate and improve that service—a potential commercial suicide. The opacity of the underlying criteria in algorithmic performance means that how those criteria are measured, weighed against one another, incorporated with other criteria—it all remains unstated: this leaves algorithms open to the suspicion of their inclination to the provider’s commercial or political benefit (Gillespie, 2014:177).

We know, more or less, that Facebook uses a machine learning-based algorithm to give users a more personalised newsfeed, which relies on constantly increasing categories and subcategories of affinity (how close the relationship is between the user and the content/source) and multiple weight levels (the type of action that was taken by the user on the content). When Facebook attempts to measure the strength of a relationship between users, that measurement is not only based on personal interactions, but global interactions on the platform can outweigh them if the signal is strong enough. This means that if Facebook shows an update to a given amount of users and only a few of them interact with it, “we may not show it in your newsfeed. But if a lot of people are interacting with it, we might decide to show it to you, too” (McGee, 2013). This practice is what Bucher defines guilt by association (2018:12), an advertising technique used in the case of users who are not particularly active or engaged in the platform. Since they don’t provide enough personal information and data on which the algorithm can build their newsfeed—that is, the portion of reality they perceive on that platform—these pieces of information are inferred by the users’ friends: advertisement and contents are then targeted and customised on the base of the friends’ online behaviour. That is, our activity online plays a crucial role not only in determining the content we will ourselves see on our preferred platforms but also in determining what other people will see on theirs.

Another interesting and concrete sample of the reality-shaping power of algorithms brought by the author is computable friendship, a model example of how algorithms actively intercept human behaviours (Garzonio, 2021). On social media platforms, friendship is put at the centre of the business model, “because better-connected users tend to increase their use of the social networking system (…) [bringing a] corresponding increase in advertising opportunities” (Schultz et al., 2014). Online friendship becomes subjected to mechanisms that privilege quantification and automation: the algorithm “measures social impact, reputation and influence through the creation of composite numbers that function as score” (Gerlitz & Lury, 2014:175), which is typically used to feed rankings and enhance predictions on users’ behaviour. In this way, this computational and computable friendship is nothing more than an equation geared toward maximising engagement with the platform (Bucher, 2018:11), ultimately serving revenue purposes. The “drive toward more”—more friends, connections, likes, interactions—is materialised by the pervasive enumeration of everything on the user interface, pushed by algorithmic rankings and scores and compelling people to reimagine friendship as a quantitative space (Grosser, 2014).

We have so far observed several declinations of algorithmic influence on “world-making”, intended here as the capacity to impact on people awareness through relevance evaluation and on people’s online behaviour through different patterns of predictive analysis. To conclude, I’d like to bring another example of algorithmic active influence on perception, with its worrying implications in terms of polarisation: echo chambers. The relation between algorithms and echo chambers is clarified, among the others, in an interesting comparative study on four different social media platforms (Quattrociocchi et al., 2021) which displays that feed algorithms—the ones that select what we visualise compared to the entire content offer of our social network, based on similarities and interactions—play a fairly important role in polarisation dynamics. This is because, despite the unprecedented potential of social media platforms in terms of free expression and exchange of public information, these same platforms are articulated on self-regulatory mechanisms pushing customised content, tailored to users’ interests, tastes and beliefs. The portion of reality that we are driven to see confirms our previous ideas and nourishes ideological polarisation (Van Alstyne & Brynjolfsson, 2005). Echo chambers are attributable to the existence of structures of exclusion, which actively obstruct the flow and consumption of information and prevent large groups of social media users from acquiring awareness of certain types of data and contents (Nguyen, 2018).

The human tendency to aggregate with people with the same attitudes and interests, when we shift to the online world, is widely encouraged and directed by automated components filtering for us contents we are most likely to engage with. This implies that the inputs provided by users—I like it, I buy it, I interact with it—and the patterns that emerge from them are transformed into means of data-driven content production. What we see on platforms, the world that is built online before our eyes, is the result of what we have previously done and chosen on these same platforms. Our inclinations and preferences will be accommodated and we will be exposed to content we enjoy and believe in, which don’t question our opinions and don’t require to enlarge our views: in a world of minimised effort, this sounds more than reassuring.

AI, Algorithms and Transparency: The European Agenda

In the European context, the current definition of AI can be retrieved in the Communication of the European Commission “Artificial Intelligence for Europe”, COM(2018)237 of April 25th, 2018:

Artificial intelligence (AI) refers to systems that display intelligent behaviour by analysing their environment and taking actions—with some degree of autonomy—to achieve specific goals. AI-based systems can be purely software-based, acting in the virtual world (e.g. voice assistants, image analysis software, search engines, speech and face recognition systems) or AI can be embedded in hardware devices (e.g. advanced robots, autonomous cars, drones or Internet of Things applications). (…) Many AI technologies require data to improve their performance.

Predictive algorithms fall within this domain: they are software-based automated systems that learn from the surrounding environment, improving their performance as they process more and more data. Targeted ads, suggestions and content pushed by algorithms on e-commerce or social media platforms, for instance, are the result of machine learning procedures, made possible by the automated processing of data and traces provided—consciously or not—by the users.

The communication was preceded by a study on the protection of human rights in the field of automated data processing techniques, published in March 2018 by the Council of Europe, in which clear requirements for algorithmic transparency were already enucleated. The document denounced the increasing opacity of algorithms, which stems not from technological needs, but precise entrepreneurial choices, and highlights the negative impact of this lack of transparency on the exercise of various important rights and freedoms enshrined in the European Chart of Human Rights. The potentiality of introducing by legislative means certain minimum standards of technical accountability and transparency for algorithms was already envisaged in the CoE document, which also proposed mediation between the entrepreneurial need to protect intellectual property in the creation of algorithms and the public need for transparency (Allegri, 2020:57). In fact, “the provision of entire algorithms or the underlying software code to the public is an unlikely solution in this context, as private companies regard their algorithm as key proprietary software that is protected. However, there may be a possibility of demanding that key subsets of information about the algorithms be provided to the public (…)” (CoE, 2018:38).

Following this study, “The European Ethical Charter on the use of Artificial Intelligence in judicial systems and their environment” was adopted in December 2018 by the European Commission for the Efficiency of Justice (CEPEJ). The fourth principle proposed in the Charter replicates the need to balance between intellectual property and the need for transparency (access to the design process), impartiality (absence of bias), fairness and intellectual integrity (prioritising the interests of justice) when tools are used that may have legal consequences or may significantly affect people’s lives. This last aspect specifically refers to the ingrained habit by companies and governments of relying upon algorithms to make decisions on very delicate aspects of human livelihood—from loan approvals to recruiting, legal sentencing and college admissions.

In the context of weighing of interests, a more viable alternative to complete technical transparency is to explain the data processing system and to describe how results are produced “in clear and familiar language, by communicating, for example, the nature of the services offered [and] the tools that have been developed” (CEPEJ, 2018:11). Moreover, independent authorities or experts could be tasked with certifying and auditing processing methods or providing advice beforehand, while public authorities could grant certifications to be regularly reviewed. We have here two core elements which will also be retrieved in the Digital Services Act: the objective of public accessibility to data processing methods through comprehensible explanations and the possibility of external audits by groups of experts or independent authorities, to certify the completeness and reliability of such explanations.

The “Ethics Guidelines for Trustworthy AI” published by a High-Level Expert Group on AI set up by the European Commission in April 2019 further specifies the concept of technological transparency. It provides AI systems to be auditable, comprehensible and intelligible by human beings at varying levels of understanding and expertise: “business model transparency means that human beings are knowingly informed of the intention of developers and technology implementers of AI systems” (AI HLEG, 2019:18). Another important transparency requirement that AI systems should meet is related to traceability: AI systems should indeed document both the decisions they make and the whole process that yielded the decisions. While traceability is not (always) able to tell us why a certain decision was reached, it can tell us how it came about, enabling reasoning as to why an AI decision was erroneous. Traceability is thus a facilitator for auditability: whenever an AI system has a significant impact on people’s lives, laypersons should be able to understand the causality of the algorithmic decision-making process and how it is implemented by organisations that deploy the AI system.

Dealing with the interpretations and decisions made by learning algorithms, a known issue is the difficulty to provide clear reasons for the results delivered, because the training process implies setting the network parameters to numerical values that are difficult to correlate with the results. For a system to be trustworthy, it is necessary to be able to understand why it had a given behaviour and why it has provided a given interpretation: a whole field of research, explainable AI (xAI), as we will see shortly, is trying to address this issue.

Concluding this overview on the challenges of algorithmic accountability, it seems appropriate to address some critical issues that the approaches so far presented might raise, especially when dealing with technical transparency. As we have commented thus far, the line proposed in the various guidelines is to compromise between private interests and public disclosure. Enlarging the focus to binding norms, Regulation (EU)2016/679—General Data Protection Regulation (GDPR)—somehow reflects this approach, and it also offers a tool, indicating in its Recital 63 that a data subject should have the right of access to his or her collected personal data and “to know and obtain communication in particular with regard to (…) the logic involved in any automatic personal data processing”. The recital specifies that this right should not adversely affect the already mentioned trade secrets, intellectual property and the copyright protecting the software, but it also states that “the result of those considerations should not be a refusal to provide all information to the data subject”. In other words, while trade secrets cannot be extended as to refuse to disclose any information on automated data processing, the same right to disclosure still cannot reach the source code but only the features and the specific logic of the employed algorithms. In this difficult balancing, wide discretion is assigned to the Independent National Authorities and the extension of the right to access the algorithmic logic remains variable, “it shortens or lengthens according to the recipient of the explanation. If the information is addressed to the data subject, the communication will extend to the logic of the algorithm functioning, but without reaching the source code (…) if the conflict of rights arises in the court, the judge will have the authority to open the source code and conduct the judicial review over it” (De Minico, 2021:30).

Despite offering a key for a correct interpretation—that the trade secret cannot be an alibi to refuse any information to the data subject or the judge—what’s written in the Regulation still clashes with the fact that technical transparency is not attainable in most cases of algorithms managed by for-profit companies. Some scholars have also suggested that revealing in case of conflict the source code to regulators or auditors merely shifts the “burden of belief” from the algorithm itself to the regulators (Hosanagar & Jair, 2018). In addition, technical transparency makes algorithms vulnerable to gaming, but the biggest problem is that source code in modern AI, after all, is less relevant compared with other factors in algorithmic functioning—which is to say, some of today’s best-performing algorithms are often the most opaque. Specifically, machine learning algorithms logic is mostly built on training data and is rarely reflected in its source code: high transparency might involve having to untangle countless amounts of data and then still only being able to guess at what lessons the algorithm has learned from it (idem). Focusing on the sole disclosure of source code appears to be an important criticality.

An achievable declination of transparency could be the one carried on by xAI: to provide basic insights on the factors driving algorithmic decisions. This approach to transparency carries with it the theme of the right to explanation, not explicitly mentioned in GDPR articles but outlined in Recital 71. The fact that this requirement is defined in a nonbinding part of the regulation has created another scholarly debate on the very existence of this right (Wachte et al., 2017; Kaminski, 2019), nevertheless dismissing the right to explanation because of the nature of recitals “would be too formalistic, and less attentive to the Court of Justice case law which regularly uses recitals as an interpretative aid” (De Minico, 2021:31). The GDPR thus establishes that users be able to demand the data behind the algorithmic decisions made for them, including in recommendation systems, credit and insurance risk systems, targeted advertising and social media platforms. XAI systems help combine this right with the technical challenges associated with transparency because they analyse the various inputs used by the algorithm, measure the impact of each input individually and in groups and finally report the set of inputs that had the biggest impact on the final decision. This kind of analysis could help programmers get around the black box problem—that is, they don’t always know what is motivating the decisions of their machine learning algorithms—while identifying relationships between inputs and outcomes and spotting possible biases (Hosanagar & Jair, 2018).

The European Legal Framework for AI and the New DSA

The digital transition is one of the key priorities established by the European Commission under the presidency of Ursula von der Leyen. In her political guidelines for the period 2019–2024 period, the President announced that the Commission would put forward legislation for a coordinated European approach on the human and ethical implications of AI. Following that announcement, on 19 February 2020, the Commission published the “White Paper on AI: a European approach to excellence and trust”. The Conclusions of the Council of the European Union, on 21 October 2020 (“The Charter of Fundamental Rights in the context of Artificial Intelligence and Digital Change”) further called for addressing the opacity, complexity and bias a certain degree of unpredictability and partially autonomous behaviour of certain AI systems, to ensure their compatibility with fundamental rights and to facilitate the enforcement of legal rules.

The European Parliament has also undertaken a considerable amount of work in the area of AI, adopting various resolutions related to AI in October 2020, including on ethics, liability and copyright. In 2021, these were followed by resolutions on AI in criminal matters and in education, culture and the audiovisual sector. The Resolution EP A9-0186/2020 “on a Framework of Ethical Aspects of Artificial Intelligence, Robotics and Related Technologies” specifically recommends the Commission to propose legislative measures to harness the opportunities and benefits of AI while ensuring the protection of ethical principles; it also includes a text of the legislative proposal for a regulation on ethical principles. Moreover, it establishes the need for any AI (including software, algorithms and data used or produced by such technologies) to comply with Union law and respect human dignity and other fundamental rights set out in the Charter (Article 5), and it outlines the ethical characteristics and of “human-centric and human-made artificial intelligence” (Article 7). These principles have been taken into account by the subsequent “Proposal for a regulation laying down harmonised rules on AI” of April 2021 (Artificial Intelligence Act, COM(2021)206), which composes the current European legal framework concerning AI in conjunction with the aforementioned Framework Resolution, the GDPR and the Proposal of regulation (EC)2020/825 or Digital Services Act (DSA), all representing the foundations of the “regulatory work in progress”.

The GDPR, in particular, has been scrupulous in placing narrow limits to the use of algorithms: its Article 22 defines the minimum standard which cannot be downgraded by Member States and specifies that the data subject shall have the right not to be subject to a decision based solely on automated processing; joined with Articles 13 and 14, it recognises the core right to be immediately informed about “the existence of automated decision-making”. Nevertheless, this provision sets out just a mere declaration of the right without specifying its content, and, more in general, the GDPR seems not so prescriptive as it should be; in some of its parts, the text does not excel in clarity, again leaving too much room to Member States’ discretionary power (De Minico, 2021:28).

In principle, the regulation provides that private codes of conduct are subordinate to the European regulation in progress, in line with the co-regulation model, “characterized by a hierarchical distribution of the normative power between public and private sources” (De Minico, 2021:20). However, both the EP Framework and the DSA refrain from laying down the basic rules to which the private soft law has to conform: EU norms just deal with the distribution of competencies among subjects, but they abstain from establishing the constituent elements of illegal conduct. In other terms, they are stating who is entitled to issue the rules but not how these rules should manage conduct, and this translates into delegating the task to private self-regulation. This criticality can be traced back to “an old vice affecting the European legislation, namely not taking a well-defined and courageous position (…) [concerning the] co-regulation regime” (ibidem).

The DSA represents so far the last stage of this “regulatory journey”, proposed in December 2020 by the European Commission to upgrade the rules governing digital services. The reform of the European digital space is contained in two legislative initiatives, the DSA and the Digital Markets Act (DMA). Both are based on a comprehensive set of new rules for all digital services that connect consumers to goods, services or content operating in the EU: intermediary services offering network infrastructure (e.g. Internet access providers, domain name registrars), hosting services (cloud, web hosting) and online platforms of different nature.

The DSA establishes binding EU-wide obligations to tackle specific concerns raised by the accelerating digitalisation and consolidation of online platforms as systemic intermediaries, and it identifies different categories of systemic risk (Recital 58 and Article 26) connected to very large platforms (reaching more than 10% of the EU’s population, 45 million users). Among the obligations, a new oversight structure to which big platforms are subjected is established: platform representatives are obliged to provide qualified researchers access to data, to allow transparency monitoring. These data will include information on the accuracy and functioning of algorithmic content moderation (Article 13), recommendation (Article 29) and advertising systems (Articles 24 and 30). Moreover, European Commission representatives, auditors and experts can carry out on-site inspections and require access to the platform’s algorithms and database (Articles 28 and 57).

Regarding the processes interested by algorithms, in particular the issue of recommender systems that prevent users from finding and interacting with online information, the general principle of transparency is outlined in Article 12. It’s the users’ right to be adequately informed on any restrictions that providers of intermediary services impose in relation to the use of their service: that information “shall be set out in clear and unambiguous language and shall be publicly available in an easily accessible format”. Users should also understand how the content presented to them is filtered, and they should be offered alternative options that are not based on profiling.

On paper, the DSA would constitute a major qualitative step forward on the issue of large platforms accountability and algorithmic transparency, setting a new benchmark for the regulation of digital services. This higher purpose, nevertheless, clashes with some of the criticalities we have identified. In particular, the tendency to leave space for self-regulation by private subjects can be spotted as a potential weakness. This “kind of blank endorsement from heteronomy to autonomy” (De Minico, 2021:21) can be detected for instance with the issue of unfair and misleading information. The DSA does not specify when “news has to be removed because they stop being a lawful exercise of a fundamental right and become an illegal act (…) Hence, the norm in blank about misleading information opens the way for self-regulation codes or, more precisely, for private platforms to mark the borderline between right and wrong” (ibidem). The subjects that should be regulated are the same that will assess whether the content they’re sharing is misleading or not, within a “vicious cycle” that witnesses private platforms establishing their codes of conduct and being at the same time the only judge to decide upon them. Consequently, the co-regulation model fails to be factually implemented, while the way norms are framed “appears not far from giving rise to an anarchic soft law”, in contrast with the expected and desirable combination between heteronomous sources and private acts.

Conclusions

To identify the meeting point between the power of performative intermediaries and digital regulation, we started our analysis with the definition of “algorithmic power”, and the investigation of the influence algorithms can display over human knowledge, perception and behaviour. We have reviewed approaches, ideas and contributions from very different fields of research, keeping in mind the multidisciplinary attitude proven to be the most useful for the understanding of complex and transversal phenomena. Enucleating and analysing algorithmic characteristics and influence—the “political power” of algorithms—helped us defining what are the most complex problems to be solved and the most obscure points of their underlying workings.

It’s exactly at these problems and opacities that digital regulation should aim. We have reconstructed, then, both the European agenda built along the years to create ethical guidelines for AI and the legal framework for this same issue, trying to define its strengths and potential weaknesses. The most evident criticalities encountered are related to the lack of prescriptive intention of the regulation, which still leaves much space both to Member States’ discretionary powers, as in the GDPR, and to private self-regulation, in the case of DSA. In addition, a major definitory clearness could avoid the “blank endorsement” that ultimately leads to private subjects’ autonomy in establishing their own rules to follow.

The subject of algorithms thus represents a concrete case study to test two regulatory alternatives, self-regulation alone or in conjunction with binding regulation. Having analysed the contributions on the theme of exclusionary powers and patterns of inclusion, we can fairly assess that algorithms and performative intermediaries are not neutral elements, but they replicate biases and discriminations already existing in the offline world, and on top of that, they have the power to silence voices, to inhibit awareness and to influence our perception. The discriminations one could quite easily detect in reality don’t appear as such if they’re kept hidden. This is why algorithms should be kept under policymakers’ control and why a binding regulation, although held to a minimum standard, could be able to design algorithms consistent with European Constitutional values, respectful of fundamental rights and acquiescent to the democratic institutional framework.