The Ethics of Artificial Intelligence is a complex book that is composed, in Luciano Floridi’s typical style, of three layers of analysis and reflection: a first in which Floridi reconnects with his previous thematic strands to delineate a further aspect of the impact of digitization on our lives and reality; a second in which he brings into focus the current central cores and problematic issues of artificial intelligence from a broad normative perspective, bringing ethics into legal and political discourse; and, finally, a third in which Floridi sows the seeds of new research perspectives on the topic under consideration.

In this short paper, I would like to focus on the third layer. In particular, I will refer to some passages in Chapter 3,1 in which Floridi not only carries forward some reflections related to the technological and digital reconfiguration of space initiated in earlier works,2 but also lays down the groundwork for thinking about a certain redesigning of space. Such space is no longer understood empirically as a physical environment, concrete situation or transcendentally as an a-priori condition of sensory experience, but normatively as a sphere of play, the nature and extent of which need to be examined and evaluated in relation to the degree to which it is (relatively) structured.

This new approach is actually quite traditional and takes us back to the beginnings of modernity, where space is thought of as a horizon for the constitution of politics. Political thought in modernity investigates, ponders, and elaborates on the conditions of human coexistence as a political project mostly implemented through the law. Coexistence demands not only to think about but also to normatively and empirically design how to structure space as the horizon of such human coexistence across time. Since modernity, a national border, for example, is an empirical boundary that can never be fully separated from its normative function of organizing human coexistence in space (inside/outside) and time (inclusion/exclusion).

In political modernity, it is possible to think about space starting from the fact that an object placed in a specific place does not allow another object to be placed in the same place except through time. An object placed in a specific place raises a claim that sets off a conflict in space (two objects cannot persist in the same place at the same time), which can be moderated in time (two objects can coexist in the same space over time). We have realized that the digital space is largely different from the analogical one.3 This has already forced us to rethink space as a horizon for the constitution of politics in the contemporary world as a human project that redesigns normatively the forms and conditions of coexistence.4 Today, we must start thinking, along the lines outlined by Floridi, about the possibility, or the hazard, that AI contributes (or nudges us) to normatively redesign the space of such coexistence.

Floridi’s critical insight is that the space in which AI systems make decisions and act is a space that is redesigned (i.e., structured, not only as a function of how such systems are agents, i.e., operate in the world) but also, and above all, as a function of how such systems are epistemic subjects (i.e., perceive and represent the world in which they are called upon to perform tasks). In earlier works, Floridi used to refer to this consideration with the robotic notion of envelope by emphasizing that space is enveloped around a robot to enable it to function better.5 This consideration still has an empirical meaning, emphasizing the robot’s ability to adapt to the environment. This consideration is theoretically significant because it already shows the blurring of the culture/nature divide as a result of the technological redesign of the environment and especially the digital synchronization of the protocols and interfaces of the technologies communicating with the redesigned environment.6

In Ethics of Artificial Intelligence, Floridi goes a step further, conceiving the space in which an AI system operates as one that is not only empirically redesigned but also structured in epistemic and normative terms. Let me first clarify what I mean by Floridi’s step further and what this can entail on the level of coexistence of human or artificial agents. Floridi recalls the difference between constraining and constitutive rules and applies it to the distinction between structured and unstructured (or partially structured) contexts through the metaphor, many times used in the literature, of the game, for example, distinguishing the game of chess from the game of football.7

The game of chess, for example, is characterized by constitutive rules, that is, rules that constitute all and only lawful behavior on the board. The rules of chess tell us how to move the chess pieces on the chessboard: there are no possible moves other than those dictated and made possible by the constitutive rules. This is made possible by two concomitant conditions: 1) chess pieces are not so much empirical objects (tokens) as representations of sets of rules (types). In this sense, if we lost a horse, we could take whatever token (e.g., a cork) and use it as a chess piece, simply saying, “This is the horse”; 2) rules can constitute all and only allowed behaviors because moves take place on a chessboard that constitutes a predefined and structured space, both empirically and normatively.

The game of football, on the other hand, is characterized by constraining rules, which do not constitute behaviors in the sense above-mentioned, i.e., they do not so much tell us what we can do (they do not instruct us about how to do what we can do) but instead just tell us what we cannot do (what behaviors are unlawful). Players can play however they want as long as they respect the rules that forbid certain incorrect behaviors. This is made possible by two concomitant conditions: 1) players are embodied agents guided by intentions; 2) rules cannot constitute the behaviors because the moves of the game occur in a semi-structured space (the pitch is empirically predefined but is normatively unstructured or only partially structured).

Leaving metaphors behind, what does all this mean? A space characterized by a structured environment governed by constitutive rules is one in which behaviors are more predictable and computable (notably by means of artificial intelligence). It is not just that but something more radical and far-reaching. It is in a computable space that it becomes possible to simulate the interaction between agents, predict the possible outcomes of those interactions, and train artificial intelligence models (such as machine learning or adversarial artificial intelligence systems) in order to increase and improve their ability to predict behaviors. From this perspective, the gradual redesign of space into increasingly structured environments governed by constitutive rules, which in turn enable the transformation of difficult tasks into complex ones,8 is not only instrumental to the development and use of artificial intelligence but also to the making of space as a new horizon for the interaction9 and coexistence of agents’ external freedoms and, therefore, for the constitution of the political. Doesn’t the possibility of generating, simulating, analyzing, and predicting behaviors and interactions among agents extend and fulfill by other means the modern political dream of a computable space?

We are therefore dealing with three different but overlapping conceptions of political space, which nowadays coexist rather than exclude each other: 1) the modern one, for which political space is still empirically and normatively territory and borders,10 in which sovereign power is exercised as a relationship of inclusion and exclusion in time and space; 2) the contemporary one, for which the political space is a place of articulation and overlap between the analogue and digital world, in which sovereign power is exercised as regulation and control over the technological protocols of information and communication and their interfaces between the different spheres of experience of our life and reality; 3) the approaching one, for which the political space is an increasingly structured and computable space of prediction and calculation.

How are these three different forms of space destined to be articulated? How deeply can space be redesigned and adapted to the functioning and development of artificial intelligence models and systems? How will this computable space of data generation and prediction impact the way we conceive coexistence and collaboration between the human and non-human? What “game” will politics play to give itself new and powerful means of control and regulation over the world?

I am confident that Luciano Floridi’s forthcoming insights and thoughts on the politics of information will provide important clues and elements to answer this complex set of questions.

Luciano Floridi’s latest book, The Ethics of Artificial Intelligence, has manifold merits, one of which is taking the challenges of technological underuse seriously (Floridi 2023). This perspective sheds light on what most scholars overlook today and how ethics relates to further normative domains, such as politics and the law. As stressed in the preface of the book, The Ethics of AI should be understood as the first part of a forthcoming Politics of Information. By focusing on the challenges of technological underuse, the aim of this note is threefold: (i) to stress “the proper design of new forms of artificial and political agency” (op. cit., at xiii); (ii) to highlight what is still missing in this “study of the design opportunities available to us” (ibid.); and (iii) to ask how the author may intend to address this gap on what I call “the strength of soft law.”

Taking the Underuse of AI Seriously

In addition to the risks brought forth by misuses and overuses of technology—by far the most popular topic among scholars discussing the normative challenges of AI today—attention must be drawn to how “fear, ignorance, misplaced concerns, or excessive reaction may lead a society to underuse AI technologies below their full potential, and for what might be broadly described as the wrong reasons” (Floridi 2023, 169–70). Although the problem may appear invisible, it is highly relevant. The wrong reasons for technological underuse trigger what economists dub as opportunity costs. For example, work on national health services and their cost analysis estimated that the opportunity costs of ambulatory medical care in the U.S.A. are around 15% (Ray et al. 2015). In the U.K., the opportunity costs of the National Health Service may amount to around 10 million pounds each year, whereas such figures could even underestimate the phenomenon (UY 2021). In some countries, e.g., Italy, which invests around 9% of its gross domestic product (GDP) in the public health sector, the opportunity costs of technological underuse may amount to 1% up to 2% of the Italian GDP (Pagallo 2022a).

Floridi proposes 20 recommendations for a good AI society that should provide a clear ethical and legal framework to resolve “the tension between incorporating the benefits and mitigating the potential harms of AI—in short, simultaneously avoiding the misuse and underuse of these technologies” (Floridi 2023, 173). Those recommendations are divided into four general areas for action to assess, develop, incentivize, and support. The recommendations include the development of auditing mechanisms for AI systems to identify unwanted consequences; the development of agreed-upon metrics for the trustworthiness of AI products and services; the incentivization of technologies that are socially preferable and environmentally friendly; the creation of educational curricula and public awareness activities that should revolve around the societal, legal, and ethical impact of AI; etc. (op. cit., 174–79). Such recommendations, on the one hand, hinge on the tenets of soft rather than hard ethics and, on the other hand, on the means of good governance. Soft ethics refers to post-compliance ethics, which applies after legal compliance with legislation (op. cit., 77). Governance concerns “the practice of establishing and implementing policies, procedures, and standards for the proper development, use, and management of the infosphere” (op. cit., 79–80). Therefore, compliance with the law is merely necessary but insufficient; it is the least that is required, whereas the most must be done: “It is the difference between playing according to the rules and playing well to win the game, so to speak” (op. cit., 173).

However, this very difference between law and ethics should be further clarified. Although Floridi admits that his distinction between hard ethics and soft ethics “is loosely based on the one between hard and soft law” (op. cit., 82), the role of soft law is simply overlooked. Why would soft law be so relevant?

Putting the Secondary Rules of the Law First

Soft law refers to the opinions, recommendations, or guidelines of authorities, public agencies, and boards that do not aim to replace the hard tools of legislation—i.e., the top-down provisions of lawmakers that hinge on the threat of physical and/or pecuniary sanctions—but rather, intend to complement them. The soft tools of the law are indispensable under several circumstances to enforce, strengthen, clarify, or stimulate the adoption of the top-down provisions of the regulator through, e.g., the development of new standards. Moreover, it is worth mentioning that several legal systems have complemented the hard tools of the law with forms of soft law and new forms of coregulation, coordination, and cooperation to tackle cases of technological underuse. Drivers of technological underuse, such as public distrust and business diffidence, can hardly be addressed only based on the top-down instructions of the regulator. The response of most legal systems to the threat of AI underuse has thus intended to complement the binding rules of the law with guidelines and mechanisms of coordination, recommendations and methods of cooperation, etc. In health law, for example, this is the approach in both the USA and the EU (Pagallo 2021).

Rather than a symptom of weakness, the strength of soft law depends on both the requirements and functions of the law, namely, what the law is supposed to be (requirements) and what it is called to do (functions). In some fields, e.g., the legal regulation of autonomous vehicles, it is worth mentioning that the lack of any robust soft law appears in most jurisdictions—Germany’s 2021 Road Traffic Act being the proverbial exception—that which appears to be the by-product of an ongoing process to determine the rules of hard law (Pagallo 2022b).

Against this backdrop, we may wonder about the connection between soft ethics and soft law.

All in all, we may accept Floridi’s quasi-jusnaturalistic approach—according to which “hard ethics first precede and then further contribute to shaping legislation” (Floridi 2023, 77)—and still, his remarks on how soft ethics may apply to hard law cannot be simply expanded to the realm of soft law. Although soft law may rely on moral arguments, a crucial difference exists between the soft sides of ethics and the law. In legal theory, soft law and the adoption of mechanisms of coordination, cooperation, and even delegation of legal powers to stakeholders concern what is usually dubbed as the secondary rules of the law. Contrary to the primary rules of the law, i.e., the hard tools of legislation that aim to directly govern the behavior of groups and individuals through the top-down instructions of the regulator, the secondary rules of the law comprise norms of change and procedures. Advocates of legal positivism and, in particular, of exclusive legal positivism would claim that matters of validity, even in the case of the soft tools of the law, stand by themselves, that is, regardless of any moral constraint after compliance with legislation. Nevertheless, the problem with legal positivism has to do with the very idea of compliance with the valid rules of the game.

After all, attention has been drawn to the soft tools of the law and mechanisms of coregulation, cooperation, and coordination endorsed by lawmakers to tackle the challenges of technological underuse. However, the analysis of how effective these norms are requires subtler forms of evaluation for the status of adherence to the regulatory provisions of the legal system. Rather than the traditional stance, according to which either legal agents are compliant or not, more nuanced assessments should distinguish between different kinds of compliance, e.g., ideal, sub-ideal, and non-compliant statuses of the legal agents (Pagallo 2022c). Whereas legal positivism and the binary alternative of compliance or non-compliance do not provide any valuable information for the assessment and improvement of current institutional initiatives against the underuse of AI—through methods of cooperation, coregulation, or coordination that hinge on the soft tools of the law—I reckon that Floridi’s distinction between playing according to the rules of the game and playing well to win the game points toward the right direction. How does soft ethics interact with soft law?

The Politics of Information

The strength of soft law is twofold since it regards both the formation or clarification of the rules of the game and how well we play to win the game through different forms of legal compliance, such as average compliance, reasonably high compliance, very high compliance, and full compliance (Hashmi et al. 2018). Although such assessment is crucial to determine how much current policies and legislation have effectively tackled the opportunity costs of technology, this kind of work is still in its infancy. I think Floridi’s soft ethics plays a critical role in supporting the efforts of soft law. The interaction between ethics and the law does not only regard the difference between the rules of the game (hard law) and playing well to win the game (soft ethics) but also the different levels of legal compliance and how the application of ethical principles can help playing increasingly well in the legal domain. Remarkably, this stance fits like a hand into a glove with the “space of soft ethics” illustrated in Figure 6.2 of Floridi’s book (op. cit., 83). I am eager to know how he may further develop this stance in The Politics of Information.

References

Floridi, L. 2023. The Ethics of Artificial Intelligence. Oxford: Oxford University Press.

Hashmi, M., P. Casanovas, and L. de Koker. 2018. “Legal Compliance through Design: Preliminary Results of a Literature Survey.” TERECOM2018@JURIX, Technologies for Regulatory Compliance. http://ceur-ws.org/Vol-2309/06.pdf.

Pagallo, U. 2021. “The Governance of AI and Its Legal Context-Dependency.” In The 2020 Yearbook of the Digital Ethics Lab, edited by J. Cowls and J. Morley. Cham: Springer.

Pagallo, U. 2022a. Il dovere alla salute: Sul rischio di sottoutilizzo dell’intelligenza artificiale in ambito sanitario. Milan: Mimesis.

Pagallo, U. 2022b. “The Politics of Self-Driving Cars—Soft Ethics, Hard Law, Big Business, Social Norms.” In Autonomous Vehicle Ethics: The Trolley Problem and Beyond, edited by R. Jenkins et al. Oxford: Oxford University Press.

Pagallo, U. 2022c. “The Politics of Data in EU Law: Will It Succeed?” DISO 1, 20.

Ray, K. N., A. V. Chari, J. Engberg, M. Bertolet, and A. Mehrotra. 2015. “Opportunity Costs of Ambulatory Medical Care in the United States.” American Journal of Managed Care 21, no. 8: 567–74.

UY. 2021. Re-estimating Health Opportunity Costs in the NHS. University of York’s Centre for Health Economics. https://www.york.ac.uk/che/research/teehta/health-opportunity-costs/re-estimating-health-opportunity-costs/#tab-1.

Introduction

Legislators and international organizations are increasingly preoccupied with the governance and regulation of AI, since it touches upon significant domains of life, including but not limited to finance, healthcare, human resources, military operations, transportation, or public transportation (see Hunkenschroer and Lütge 2022; Schmidt et al. 2022). Consequently, the said family of technologies exerts an immense influence on individuals, business, and society at large.

The European Union in particular has positioned itself early on in the discourse on how to regulate AI. Particularly, the work of the High Level Expert Group on AI (HLEG, April 2019), but also the ethical discussions initiated by AI4People had a strong effect on the formulation of a joint European approach to AI regulation (Engstrom and Jebari 2023). This development was led by prominent philosophers and ethicists, but it was also driven by scholarship in constitutional law and organizations that expressed a keen interest in aligning AI with legal notions such as fundamental rights, rule of law, or democracy (Ulnicane 2022).

Thus, the regulatory response to AI involves two different conceptual pillars that represent different normative traditions. The discourse on technology ethics has early on scrutinized the impact of technologies, specifically in the context of life sciences, but also in respect to nuclear energy. These discourses have revolved around concepts such as individual responsibility, accountability, but also considerations related to values such as beneficence, autonomy, fairness, or justice that originate in traditional moral philosophy. Moreover, the EU AI Act but also earlier on the report of HLEG on Trustworthy AI has pointed at the relevance of fundamental rights and constitutional principles such as democracy or rule of law that are at stake when it comes to the massive use of AI in society (HLEG AI, April 2019).

The existence of the two mentioned traditions prompts the question on how the relationship between both conceptual pillars within current attempts of legislation, but also the place of ethics in emerging AI regulation. Is AI regulation replacing ethics by gravitating to fundamental rights, or are ethics and fundamental rights to be understood as mutually reinforcing concepts? In order to elaborate this, the remainder of our contribution is structured as follows. We begin with the examination of the concept of hard and soft ethics including the closely related idea of post-compliance as introduced by Floridi. In a further step, we concentrate on the question how the aforementioned normative traditions, namely AI ethics and fundamental rights have been integrated in the EU AI Act. Based on this, we examine the structural relationship between ethics and human rights along the lines of Floridi’s use of hard and soft ethics, in order to derive implications for the future role of ethics in AI regulation and governance, with a specific focus on the concept post-compliance.

Hard and Soft Ethics

In The Ethics of Artificial Intelligence, Floridi (2023) distinguishes between two concepts of ethics that impose different types of normative obligations on individuals and society. This distinction serves as a conceptual tool to derive a better understanding of the division of labor between upcoming AI legislation and AI ethics approaches developed by companies in the form of self-regulation. It also closely aligns with earlier approaches on the division between politics and ethics, but also formal and informal institutions.

According to Floridi (2023, 82), hard ethics pertains to what “we usually have in mind when discussing values, rights, duties, and responsibilities—or, more broadly, what is morally right or wrong.” Moreover, hard ethics has a purpose in defining the legal conditions, constraints, and circumstances which determine individual agency in society by “formulating new regulations and challenging existing ones (p. 82).” It concerns therefore quite fundamental questions in respect to social organization that require clarification on the level of legislation and are not just subject to individual deliberation. According to the author, the reflection on the underlying normative foundations of society has prompted the dismantling of legal structures that were unethical, such as Apartheid legislation. In that sense, it can be understood as a significant driver of moral progress. Furthermore, Floridi introduces the concept of soft ethics, which refers to the conduct of individuals and companies in areas that lie outside of the scope of regulation (82). In a game-theoretical sense, soft ethics determines the moves of the game played by individual players (Lütge et al. 2016). Floridi (2023, 82) states here that soft ethics “does so by considering what ought and ought not to be done over and above the existing regulation—not against it, or despite its scope, or to change it.” It is therefore the domain of individual moral choice and can be advantageous to act in line with soft ethics. Particularly, in the case of AI, soft ethics serves as a tool to inform corporate decision makers on the positive potential of technologies, depicting an opportunity strategy, but as serving as an approach to risk management.

The distinction made by Floridi between hard and soft ethics is deeply embedded in a wider philosophical tradition, tracing back to Aristotle, who sharply distinguished between politics, referring to the principles of how to govern a society, and ethics, pertaining to the right, but also smart way to act as an individual. This distinction is highly relevant from institutional economics, which emphasizes the role of formal (laws, constitutions, and binding guidelines) and informal institutions (such as traditions, moral perceptions, and customs) that coordinate human behavior (Lütge et al. 2016). It also involves the metaethical question of how to separate between the space to be governed by hard ethics as understood by Floridi and the space to be governed by individual moral commitments and strategic choices likely to be influenced by a particular cultural background. This debate manifests itself in policy choices, for instance the fiercely debated issue of voluntary self-regulation vs. mandatory legislation in the case foundational models. Based on this conceptual difference it is therefore relevant to identify the structural role of both approaches, namely hard and soft ethics, in the context of emerging AI regulation, and particularly the role “soft ethics” is likely to play in this.

The Regulation on Artificial Intelligence: Between Ethics and Fundamental Rights

In its beginning, the public debate on AI governance and regulation has been characterized by a dominance of ethical discourses and the search for a common normative foundation that should underpin upcoming legislation, in the best case anchored in international conventions. Doing so, the majority of frameworks has oriented itself to existing values such as fairness or accountability, resorted to consequentialist and deontological arguments, or sought to align themselves to particular frameworks including the United Nations Sustainable Development Goals or other concepts related to AI for Good (see Halsband and Heinrichs 2023).

Furthermore, ethical approaches to AI have attempted to formulate principles that would address certain features of AI as a technology that stand in conflict with ethical norms (Evans et al. 2023). This owes to the fact that AI tends to replace human agency, that it is prone to bias, that it depends on data, and that the complexity of machine learning and neural networks deprives us of a clear explanation as to why AI has made a particular decision (compare HLEG AI, April 2019). The EU AI Act has referred to these elements as “opacity, complexity, dependency on data, autonomous behavior” and put them at the center of its legislative approach (see “Explanatory Memorandum,” Regulation 2021/0106). Finally, the actual impact created by AI depends on the agenda of the involved parties, particularly users and developers, implying that AI can also be used and modified for unethical or illegal purposes. The presented characteristics have therefore been understood as the foundation of the design of mitigation and prevention measures. This includes for instance the concept of explainability as a means to address the algorithmic complexity, fairness addressing the issue of bias, and the focus on human autonomy when navigating the issue of algorithmic opacity or the non-deterministic nature of AI. However, the contextual meaning of what fairness or what prevention of harm would constitute in a given specific context has often remained vague (see HLEG, April 2019). While the mitigation measures paved their way into the AI Act high-risk requirements, the objective to be realized by developers and users of AI as well as the exact risk categories to be considered in assessments has not been fully addressed in the AI ethics discourse.

On the contrary, regulatory frameworks can rely on an arsenal of normative concepts, which informs on harm, fairness, or privacy considerations in a given context. In the European context, the EU Charter of Fundamental Rights (2000) codifies not only basic rights of individuals such as the right to privacy, freedom of opinion, or the right to non-discrimination, but also rights of citizens, for instance the right to good administration. Fundamental rights are understood as principles that (1) create boundaries for state, in this case EU, interventions, that (2) imply measures taken by state legislation to protect individuals from adverse impacts on these rights by third parties and that impose duties on states to fulfill entitlements, for instance the duty of states providing individuals with access to education. According to the preamble of the EU CFR, technologies are relevant for this constitutional mission, as they can unfold major effects on the norms protected by human rights. This coincides with tendencies in international law, which calls on state organizations to prevent adverse effects on human rights—a concept closely related to fundamental rights—created by business entities (Kriebitz and Lütge 2020).

A key aim of European legislation on technologies has therefore been to ensure that business activities are not leading to a decline in fundamental rights. This explains why fundamental rights are repeatedly mentioned throughout the act as determinants of risks, as they give a contextual understanding of harm, for instance when it comes to violations of the physical integrity of a person, or when an AI system undermines freedom of speech. Thus, fundamental rights considerations are found in different contexts of the EU AI Act, particularly when it comes to the identification of “prohibited practices” and “high-risk” AI (see “Explanatory Memorandum,” Regulation 2021/0106). The former case refers to situations in which the very use of AI conflicts with the content of fundamental rights, for instance in the case of social scoring (see ibid. Art. 5). Moreover, high-risk use cases describe settings in which the stakes are higher due to the proximity of AI use to norms protected by fundamental rights. This would be the case if AI is deployed in the context of determining access to education.

To conclude with, the EU AI Act draws from two different conceptual pillars. The AI ethics discourse has largely described the nature of the problem, namely that AI has certain characteristics that can undermine norms and that AI can also be used for multiple purposes. On the contrary, fundamental rights depict normative expectations and give indications about the risks characterizing a particular setting, for which AI is used.

Fundamental Rights and Ethics: Hard or Soft?

Owing to the conceptual difference between hard and soft ethics on the one side, which correlates with the normative strength of claims on the spectrum between “must” and “may,” and on the other side the existence of two normative pillars of AI regulation, namely “ethics” and “fundamental” rights, the question arises how the two conceptual tools stand with each other. In the following, we elaborate that the existence of fundamental rights and ethics as two different normative traditions in AI regulation adds a further layer to the concept of hard and soft ethics as introduced by Floridi.

This might be counterintuitive on the first glance. One might be tempted to make the case that fundamental rights given their legal background and their prominent role in the EU AI Act represent an example of hard ethics. They would directly imply a “must,” in the sense of immediate action for changing the legal framework, but also imposing strong normative obligations on individual or corporate agencies. Likewise, one could argue that ethics would be reduced to post-compliance as a principle informing on the right conduct, following the introduction of fundamental rights as the major normative approach behind AI regulation.

On a second glance, the exact interplay between hard and soft ethics, but also between ethics and fundamental rights is a bit more complex than it seems. While fundamental rights portray clear normative guardrails for states, the implications of fundamental rights for companies are slightly different. Other than states, fundamental rights in their own right do not legally bind business enterprises (compare: Kriebitz and Lütge 2020). The notion of fundamental rights impact hints therefore at certain norms (in German Rechtsgüter), protected by fundamental rights such as freedom of speech or freedom of movement that should be considered by companies in their operations. Companies need therefore to demonstrate that they did engage in measures to prevent and to mitigate against particular interventions in areas protected by fundamental rights. For instance, a social network evaluates the impact of certain design choices and privacy settings on freedom of speech (see HLEG, April 2019). However, not all entities in the AI ecosystem are likely to have such an influence on design choices, particularly small business enterprises that use AI solutions created by others. It is therefore plausible to assume that exact normative urgency on the spectrum of “must” vs. “may” would depend on the leverage companies have in understanding but also preventing a negative impact created by an AI solution. Such a rationale would be at least plausible considering the focus in the business and human rights discourse on the mitigation of adverse impacts created by corporate activities. Detailed codes of conduct specifying the implementation of fundamental rights in AI development and use would be difficult to define anyhow, given the rule of constitutional courts in resolving tensions between different implications of human, and in the EU context, fundamental rights. Moreover, fundamental rights are instrumental for understanding the exact interplay between soft and hard ethics, as legislation—hard ethics—bears always implications for the exercise of fundamental rights for example the freedom to conduct business. Tighter regulation usually implies higher market entry barriers, particularly for companies that seek to challenge already existing monopolies. AI regulation can impose high costs on individuals and companies that seek to participate in AI development, and needs to strike a proportionate balance between the risk posed by the technology and the consequences of legislation on fundamental rights understood as freedom. Doing so, a fundamental rights-centered approach would create room for individual actions that can be guided by ethical considerations and that are restricted by an overarching legal framework (compare: Lütge et al. 2016; Häußermann and Lütge 2022).

The role of ethics remains therefore in identifying traits of AI that are pertinent for a given use case and how these traits create negative effects on AI solution, but also to identify the right mitigation and prevention measures when balancing positive and negative impacts of a particular AI solution in terms of human rights. An example would be the use of care robots that require immense data in order to provide highly individualized patient care (Boch et al. 2023). Here again, ethics will have to assess the existence of hard boundaries, namely clear violations of ethical principles for instance, that must not be crossed for instance the lack of consent in patient interactions and the existence of principles, which require a contextual interpretation. Moreover, the resolution of dilemmatic situations and other moral trade-offs is likely to be a main field of examination, particularly as regulation will hardly regulate all potential trade-offs and dilemma situations. Nevertheless, this debate often involves facets regulated by law, for instance in the case of risk assessments along the lines of fundamental rights, and by broader and not-codified ethical conventions, where developers and users of AI will have to rely on normative instruments to diffuse tensions between different norms for example the notion of proportionality or the understanding of leverage.

Considering the initial question on the role of ethics in world after AI regulation, we would propose seeing ethics not merely as an area of post-compliance. Instead, we would recommend to see ethics as a means to identify the normative relevance of a given AI solution (“pre-compliance”), an indicator for the right choice of mitigation and prevention measures and finally as a tool making it easier for companies to comply with upcoming legislation (“compliance”). While the conceptual distinction between hard and soft ethics makes sense as it informs on the division of labor between society, organizations, and individuals, AI regulation constitutes an in-between necessitating not only hard and soft ethics, but a closer understanding of the conceptual differences between ethics and fundamental rights.

References

Boch, A., S. Ryan, A. Kriebitz, L. M. Amugongo, and C. Lütge. 2023. “Beyond the Metal Flesh: Understanding the Intersection between Bio- and AI Ethics for Robotics in Healthcare.” Robotics 12, no. 4: 110.

Charter of the Fundamental Rights of the European Union. 2000. https://www.europarl.europa.eu/charter/pdf/text_en.pdf.

Engstrom, E., and K. Jebari. 2023. “AI4People or People4AI? On Human Adaptation to AI at Work.” AI & Society 38, no. 2: 967–68.

Evans, K., N. de Moura, S. Chauvier, and R. Chatila. 2023. “Automated Driving Without Ethics: Meaning, Design and Real-World Implementation.” In Connected and Automated Vehicles: Integrating Engineering and Ethics, 123–43. Cham: Springer Nature Switzerland.

Floridi, L. 2023. The Ethics of Artificial Intelligence. Oxford: Oxford University Press.

Halsband, A., and B. Heinrichs. 2022. “AI, Suicide Prevention and the Limits of Beneficence.” Philosophy & Technology 35, no. 4: 103.

Häußermann, J. J., and C. Lütge. 2022. “Community-in-the-Loop: Towards Pluralistic Value Creation in AI, or—Why AI Needs Business Ethics.” AI and Ethics, 1–22.

High-Level Expert Group on Artificial Intelligence (HLEG AI). 2019, April. “Ethics Guidelines for Trustworthy AI.” European Commission. https://ec.europa.eu/futurium/en/ai-allianceconsultation.1.html.

Hunkenschroer, A. L., and C. Lütge. 2022. “Ethics of AI-Enabled Recruiting and Selection: A Review and Research Agenda.” Journal of Business Ethics 178, no. 4: 977–1007.

Kriebitz, A., and C. Lütge. 2020. “Artificial Intelligence and Human Rights: A Business Ethical Assessment.” Business and Human Rights Journal 5, no. 1: 84–104.

Lütge, C., T. Armbrüster, and J. Müller. 2016. “Order Ethics: Bridging the Gap between Contractarianism and Business Ethics.” Journal of Business Ethics 136: 687–97.

Regulation 2021/0106. “Regulation of the European Parliament and of the Council Laying Down Harmonized Rules on Artificial Intelligence.” European Parliament, Council of the European Union. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206.

Schmidt, L. 2022. “Mapping Global AI Governance: A Nascent Regime in a Fragmented Landscape.” AI and Ethics 2, no. 2: 303–14.

Ulnicane, I. 2022. “Artificial Intelligence in the European Union: Policy, Ethics and Regulation.” In The Routledge Handbook of European Integrations. Taylor & Francis.

It is hard to be aware of transformations as they are unfolding. Even harder it is—for most of us—to accept transformations once they have unfolded. Few things elicit more resistance than change, especially when it is deployed before our eyes, in plain sight. Humans have this mysteriously fascinating tendency to deny that what they thought was the case, in fact, is no more. And yet we have embraced the digital transformation of our previous (analogue) world into a global ecosystem in which our presence is the presence of our data, our identity is the identity of our data.

Or, haven’t we? Have we really accepted this new digital world, or are we experimenting with (futile? pointless?) forms of resistance to an already accomplished process? The coming of age of AI—not AI per se, that has been around for decades, but its recent large-scale implementation in all domains of individual and associated life—is sparking global debate about the merits and the limits of digitalization.

We share presence and identity with tools and objects that literally do not exist for us (meaning, cannot be used, rented, bought and sold) unless we can appear as data that such tools and objects can compute and transact alongside the data of myriad other users, perfectly identical to ourselves—at least in the “eyes” of such digital objects. To many of us this is just what it is: the present way humans inhabit the worlds. But what shall we do of the deeply uncomfortable feelings that our existence is somehow reduced to that of our data, that it is inexorably mediated by technological objects, and that there is no coming back from this point of (digital) alienation?

Perhaps a simpler question is whether the alleged digital transformation of our world has actually already occurred, or whether it is a process that is still unfolding, perhaps an irreversible one. But yet again such question triggers more complex thoughts, as we inevitably need to ask how shall we live in such transformation, whether it is an accomplished or an ongoing process.

These and similar questions hover overhead as the reader engages with Luciano Floridi’s new book The Ethics of Artificial Intelligence. Right from the first pages, Floridi gives center stage to the phenomenon of digital transformation as the bedrock of his argument(s). The book dovetails anthropological, sociological, and ethical reflections as it presents a number of normative observations about AI ethics and, while the title does not really convey that, AI governance.

The philosophical starting point of such articulated normative discussion is an epistemological assumption: we can no longer know the world we are in unless we see it for what it is, meaning, a digital world. An ontological assumption is coupled with the epistemological one—or rather nested within it, as it were: the very nature of the world is different (digital), not just the phenomenic surface of how we do stuff in it. Floridi has defended such two philosophical theses elsewhere. In this new book, he takes them as the basis for two historically laden (sociological) claims. First, the digital transformation of the world has happened and cannot be unmade. Second, the digital transformation of the world will not happen a second time. Floridi, however, maybe unlike the way he has been making this argument in the past, is now insisting more on the ongoingness of digital transformation. And he does so, precisely in relation to AI. The digital world, he maintains, is still evolving—or being remade. In particular, AI is the force that is shaping such digital retransformation. AI is providing unprecedented computational opportunities. In this sense it is part and parcel with the very digital world it is nonetheless transforming. However, for AI to be able to deploy its agency (I’ll get back to this shortly) in the digital world, the digital world needs to be properly arranged in the first place. The current status of the digital world is not—yet—conducive to the deployment of AI’s full potential in terms of predictive, classificatory, and decision-making capabilities. What needs to happen is what Floridi calls “enveloping.” Very simply stated, enveloping means making the world AI-friendly, or better, AI-ready. More specifically, it means arranging the digital world in such a way as to provide AI systems with the kind of data they need (in terms of both kind, quality, and quantity); restructuring industries and institutions (e.g., the healthcare sector) so that key services can be delivered in an AI-automated way; reforming ethics norms and regulatory ecosystems so that AI-driven decisions can be trusted and be perceived as legitimate.

It seems that many of the challenges Floridi identifies around AI relate in one way or the other to enveloping. The biggest of such problems from an ethics point of view, Floridi clearly spells out—with overtly Kantian overtones: are humans (their wellbeing, their happiness, their flourishing) the ultimate end of this enveloping process, or is it rather the case that humans too are the means through which enveloping is taking place (28)?

What is particularly interesting in this way of interrogating the ethics of AI is that it shifts the focus from the consequences of using and misusing AI, to exploring which “forms of life” AI is giving rise to. Philosophy of technology and Science and Technology Studies scholars have long insisted that technological artifacts do derive from and, at the same time, reinforce imagined forms of social life (ethics) and social order (governance)—a process that is not necessarily visible, let alone straightforward to reconstruct while it is still taking shape.

And this takes us back to the point we have started from, namely, Floridi’s (now) sociologically and historically laden appreciation of digital transformation as an ongoing process. It is precisely because the process is ongoing that we may have a hard time recognizing it or figuring out its ethically fundamental characteristics. Nonetheless, it looks like we cannot avoid questioning (in Kantian, philosophical, or STS ways) the process whereby the world and its inhabitants get enveloped for the sake of, well, AI itself.

As briefly alluded to before, Floridi defines AI as an agent, or to be more precise, as a reservoir of agency to solve problems and perform tasks (whole chapter 2). This is a promising viewpoint to start exploring whether and how humans are adapting to the presence and behavior of new agents (AI models) and what specific principles, guidelines, and regulations need to be adopted in order to deal with those agents. Are humans able to resist certain unethical or unpolitical forms of adaptation? Should we be worried that resistance is—as we said at the beginning—already a sign of pointless protest against change that has already inexorably occurred? If the digital transformation of the world will not happen a second time—as Floridi maintains—then what AI is realizing is not a second digital revolution. It is either the latest stage (the end stage?) of digitalization or something different altogether.

Floridi seems to think the former to be the case. The recent explosion of AI and its associated ethical and governance challenges are part of a broader and longer phenomenon. To some extent, the reader has the impression that questioning this phenomenon cannot stretch to the point of arresting it, or, for that matter, attempting to substantially interfere with it. The process has momentum of its own and it is pretty much irreversible. Hence the book goes back and forth between foundational issues and practical strategies of piecemeal adaptation and mitigation, focusing also extensively on principles, guidelines, and governance practices as means to channel innovation in AI toward what Floridi calls “AI for social good” (AI4SG).

The ultimate aim of any philosopher of technology—a label I wish Floridi will be comfortable with—is not to understand what technology is, but what technology is for. In this respect, philosophy of technology converges toward, or maybe even overlaps with ethics—a disciplinary label with which Floridi has become increasingly comfortable with in recent years. Floridi’s book is a book about the ethics of AI. But it is more generally a book about the ethics of technology as a historical phenomenon that invests, at the same time, human self-understanding and our capacity to conduct ourselves with practical reason—yes, in an overtly Kantian sense. However, this book pursues an aim that Kant might have found too mundane—that of situating practical reason historically. The backbone of Floridi’s strategy is spelling out the ethical salience of AI as the incarnation of a historical condition. Hence, his interest—at the same time ambitious and humble—for AI as a determinant of social good. I think recognizing this philosophical move is key to understand the book as a book about social ethics, but also, more in general, to appreciate where ethics as a discipline is going and how it can help with some of humanity’s most pressing questions. This book invites scholars in ethics to look at the big picture, and to analyze how our questions relate to AI as a piece of our (present) world.

Whether we should think—with Hegel—of historically determined technological conditions, or—with Marx—of technologically determined historical conditions is not a problem that bothers Floridi more than it should. Or maybe it will, in his next book.