Artificial virtuous agents: from theory to machine implementation

Abstract

Virtue ethics has many times been suggested as a promising recipe for the construction of artificial moral agents due to its emphasis on moral character and learning. However, given the complex nature of the theory, hardly any work has de facto attempted to implement the core tenets of virtue ethics in moral machines. The main goal of this paper is to demonstrate how virtue ethics can be taken all the way from theory to machine implementation. To achieve this goal, we critically explore the possibilities and challenges for virtue ethics from a computational perspective. Drawing on previous conceptual and technical work, we outline a version of artificial virtue based on moral functionalism, connectionist bottom–up learning, and eudaimonic reward. We then describe how core features of the outlined theory can be interpreted in terms of functionality, which in turn informs the design of components necessary for virtuous cognition. Finally, we present a comprehensive framework for the technical development of artificial virtuous agents and discuss how they can be implemented in moral environments.

1 Introduction

As artificial systems enter more domains of human life, the last decades have seen an explosion of research dealing with the ethical development and application of AI (AI ethics), and how to build ethical machines (machine ethics).¹ While efforts of the former kind have seemingly converged on a set of principles and guidelines (Floridi and Cowls 2019), their capacity to have any substantial impact on the ethical development of AI has been called into question (Hagendorff 2020; Mittelstadt 2019). Lacking mechanisms to enforce their own normative claims, AI guidelines might instead serve as “ethics-washing” strategies for institutions. To bring ethics into the very core of the research and development of AI systems, it has instead been suggested that ethicists should adopt the role of designers (Van Wynsberghe and Robbins 2014). Accordingly, several attempts have been made to implement ethical theory in machines, with the majority of them taking one of three approaches: consequentialism (Abel et al. 2016), deontology (Anderson and Anderson 2008), or hybrids (Dehghani et al. 2008).

Virtue ethics has several times been proposed as a promising recipe for artificial moral agents (Berberich and Diepold 2018; Coleman 2001; Gamez et al. 2020; Howard and Muntean 2017; Wallach and Allen 2008). Due to its emphasis on moral character and moral development, it offers a path to equip artificial moral agents (AMAs)² with the ability to learn from experience, be context-sensitive, adapt, and conform to complex human norms. However, hardly any technical work has attempted to implement virtue ethics in moral machines.³ The main reason is that virtue ethics has been proven difficult to approach from a computational perspective, especially in comparison to its more popular alternatives. Simply put, it is easier to implement particular deontological rules and consequentialist utility-functions as opposed to generic virtues and moral character.

The main goal of this paper is to tackle the challenge head-on and demonstrate how virtue ethics can be taken all the way from theory to machine implementation. The rest of the paper is structured as follows. In Sect. 2, we explore four major benefits virtue ethics could offer to the prospect of moral machines. In Sect. 3, we face up to four critical challenges for artificial virtue, including (3.1) the uncodifiability of virtuous language, (3.2) its reliance on human-like moral capacities, (3.3) the role of virtues, moral exemplars, and eudaimonia, and (3.4) issues regarding technical implementation. Simultaneously, we outline a path to artificial virtuous agents (AVAs) based on moral functionalism, bottom-up learning, and eudaimonic reward. Section 4 describes how features of the outlined virtue-ethical framework can be interpreted in terms of functionality, which in turn can guide the technical development of AVAs. We then present a generic architecture that can act as a blueprint for algorithmic implementation. In Sect. 5, we discuss remaining challenges and identify promising directions for future work.

2 Machine ethics and virtue ethics

2.1 Machine ethics

With a growing number of self-driving vehicles on public roads, and a variety of robots being in education, medicine, and elderly care, it is hard to question the urgent need for AI systems to have some form of ethical considerations factored into their decision making. Were AI to continue on its course of replacing traditional human roles such as drivers, medical doctors, soldiers, and teachers, we should also expect them to adequately meet the moral standards entailed by such roles. For these reasons, machine ethics have attracted a growing amount of interest among academics in the intersection of moral philosophy and computer science, and the resulting body of work ranges from more or less detailed prototypes of ethical machines to theoretical essays on what moral agents ought or ought not to do (Anderson and Anderson 2011; Sparrow 2007; Winfield et al. 2014).

There are many pathways towards ethical machines, and different paths provide their own distinct benefits and disadvantages. The rule-based nature of deontological ethics elegantly corresponds to the type of conditional statements often associated with computer code. Similarly, the utility-maximizing aspects of consequentialism seem to resonate well with objective functions found in mathematical optimization, or the reward functions used in reinforcement learning. Deontology and consequentialism are thus fruitful frameworks for the pursuit of moral machines in their own ways, each corresponding to important aspects of moral behavior found in humans. However, by building AMAs based on these approaches, there is a risk of cherry-picking particular features of the theoretical counterpart without offering any account of how these are situated in the general cognition of the agent and its relation to the complicated dynamics of our every-day ethical lives. Behind rules and utility-functions, there is no moral character to speak of; no learning or adaptation; no thorough account of what it is to be moral besides following a simplified version of the normative theory it is based upon.

2.2 The appeal of virtue ethics

“Virtue ethics” refers to a broad family of related ethical views, with variations found in Buddhist, Hindu, and Confucian traditions (Flanagan 2015; Perrett and Pettigrove 2015; Yu 2013). In the Western tradition, the most influential version stems from Aristotle, and in the modern age it has emerged on to the central stage owing a lot to Anscombe (1958), Nussbaum (1988), Hursthouse (1999), and Annas (2011). While a comprehensive summary of the tradition and its variations could fill a large library,⁴ we will introduce some of its central aspects by highlighting four major benefits the theory offers to moral machines.

2.2.1 Moral character

First and foremost, a person following virtue ethics puts her main focus on her character by fostering the dispositions that enable her to act in a morally good way. In this sense, virtues are the morally praiseworthy character traits one has or strives to possess. The courageous agent puts itself at risk to save another agent, not because it follows a rule, nor because it would result in the best outcome, but simply because it is what a courageous agent does (Hursthouse 1999). Consequently, the virtuous agent blurs as well as bridges the gap between, on the one hand, actions as a result of conscious deliberation and reasoning, and on the other hand, the psychological and biological dispositions that enable her to act in certain ways. While normative theories can be useful heuristics that guide us towards ethical conduct in terms of principles and reasons, in everyday life, there is often a gap between how we ideally ought to act and how we actually act.⁵ In fact, our general behavior is influenced by a range of conscious and unconscious processes; from emotions and motivations at the psychological level to mood-altering hormones and gut bacteria of our biological systems (Tangney et al. 2007; Teper et al. 2011). Taking character as the principal subject of moral evaluation, virtue ethics therefore enables us to account for these mechanisms, which in turn allows us to conceptualize a more comprehensive picture of what it means to be moral.

2.2.2 Learning from experience

Another key feature of virtue ethics is the emphasis on learning from experience.⁶ A child and an adult might share the same good intentions, but due to a lack of experience, the child is often unaware of what she needs to do to effectively reach the intended result. Only through experience can we acquire the practical wisdom (phronesis) that helps us exercise good judgment and promote the excellence of our character and habits. Fueled by the same intuition, experimental studies in moral psychology have grown into full-blown research paradigms that seeks to illuminate the ways cognitive development and experience are necessary for certain forms of moral conduct.⁷ In a similar vein, Annas (2011) has developed an influential version of virtue ethics based on the idea that the way we learn to be virtuous is similar to how we acquire a practical skill (Dreyfus 2004). Following Annas, moral competence is acquired both in terms of judgement (e.g., to follow reasons) and action (e.g., a moral know-how) through an active intelligent practice, akin to how we acquire and exercise skills such as farming or playing the piano.

One advantage of taking a learning approach to machine morality is that it enables the AMA to be context-sensitive to particulars and adapt to changes in ways that are difficult to encode in static rules. After all, real-life moral dilemmas rarely present themselves in the abstract and distilled manner as they are often portrayed in thought-experiments such as the trolley problem. It is perhaps even rarer that we find a moral dilemma to be in every sense similar to some we have encountered before, which in turn curtails the applicability of general principles. Following Aristotle, it is rather through our repeated encounter with particulars that we practice our practical wisdom (NE 1141b 15). Consequentially, as the virtuous agent develops through a continuous interaction with its environment, it would ideally be able to conform, not only to certain values and rules, but to the subtler details of social norms and cultural customs, with the additional ability to adapt accordingly as these change over time.

2.2.3 Relationship to connectionism

A third selling-point for the prospect of artificial virtue, and a natural extension of the second, is that virtue ethics resonates well with connectionism⁸ and the correlated methods that are frequently employed in modern AI. Although ideas about artificial neural networks and learning algorithms have circulated since the 1940s (McCulloch and Pitts 1943), connectionism truly rose to prominence through the speed of twenty-first century computer chips combined with internet-age amounts of training data (Miikkulainen et al. 2019). Faster processing and more data have allowed increasingly larger networks, which in turn has made machine learning and neural networks the most dominant AI tools of today, reaching human and expert-level performance in areas such as pattern recognition, game playing, translation, and medical diagnosis (Deng and Yu 2014; Senior et al. 2020). Due to the learning emphasis, and the ability to capture context-sensitive information without the use of general rules, several authors have pointed out the appeal of uniting virtue ethics with connectionism (Berberich and Diepold 2018; Gips 1995; Howard and Muntean 2017; Wallach and Allen 2008). Some would even go so far as to claim that connectionism holds the essential keys to fully account for the development of moral cognition (Casebeer 2003; Churchland 1996; DeMoss 1998), while others have noticed the historical link between virtue ethics and connectionism through Aristotle.⁹ Essentially, the relationship between the two could therefore provide AVAs with a compelling cognitive framework in combination with the technological backbone of modern learning methods.

2.2.4 Relationship to general cognition

The last appeal is that virtue ethics, compared to other theories, cuts deeper into the relationship between moral cognition and cognition in general. That is, virtue ethics situates morality, not separate from, but rather alongside general capacities and functionality. To use an analogy: we often measure the performance of artificial systems in functional terms, i.e., to the extent they are able to perform a certain task.¹⁰ If they were equipped with more salient forms of moral behavior, we would also judge their behavior in relation to their other capacities. For instance, the moral competence of a self-driving car is intimately linked to its general ability to drive safely without supervision, including capacities such as speed control and collision-detection. A self-driving car with faulty brakes would simply lack the ability to avoid certain collisions even if the control-system of the vehicle was determined to do so. Similarly, you can only be courageous if you have the means to act courageously; to save a person from drowning in the ocean, you need to know how to swim. The point is that morality cannot be viewed isolated from non-moral capabilities. This reflects Plato and Aristotle’s shared view on how virtue is intimately related to function; a virtue is a quality that enables you to be good at performing your function (ergon).¹¹ A good knife has the virtues—of being durable, sharp, etc.—that allows it to carry out its function (cutting). A good life, according to Aristotle, is thus to fulfill one’s ergon through virtuous living (arete). In a modern context, this can be seen to emphasize the intimate relationship between moral cognition and general cognition. For instance, some authors have argued that there is no sharp distinction between moral and non-moral cognition on the basis that they have coevolved throughout the evolution of mankind (Flanagan 2009; Kitcher 2011).¹² Or as Johnson (2012) claims: there is no special moral faculty besides the general faculties. In the growing imaging literature of moral cognition, the emerging picture is that morality relies on a highly diverse and decentralized neural network that selectively uses specific regions depending on the associated context (FeldmanHall and Mobbs 2015).

Grounding morality in a general cognitive framework is constructive for the pursuit of moral machines in several regards. It allows us to more clearly determine what the appropriate virtues would be for an artificial agent in relation to its role, and help us to focus on the relevant traits that enable it to excellently carry out its function. A social companion robot used in elderly care should not share the same virtues as a self-driving car; they are equipped with different functionalities, serve different purposes, and face their own distinct problems. By contrast, the prejudice of universalist moral philosophy, i.e., the idea that there are general answers to particular moral problems, might lead one to implement one and the same “generic moral module” in machines across all domains, which would obstruct the nuances and domain-specific challenges that machines face with regards to their particular purpose.

Placing artificial morality within general cognition would also enable the development of AMAs to continuously draw from insights from the growing body of brain science, which in turn could shed light on aspects of morality that are only possible through the use of other complex and highly distributed cognitive abilities.

In summary, virtue ethics provides a smorgasbord of attractive features for the pursuit of moral machines. However, we have so far only explored the prospect of AVAs in rather idealistic terms. To construct an AVA that would fully realize the discussed benefits, one would need to create something more or less similar to a virtuous human being, which is unrealistic given today’s technology.

3 Challenges for artificial virtue

Approaching virtue ethics from a computational perspective presents several novel challenges. In this section, we will focus on issues stemming from (i) the equivocal nature of the theory and its concepts, (ii) its reliance on human-like moral capacities, (iii) the difficulty of deciding the role of virtues, and (iv) technical implementation. Using the moral functionalism of Howard and Muntean (2017) and Hursthouse’s virtue-ethical framework (1999), we will argue that there is a feasible path towards artificial virtue, but only if we give up the idea of trying to capture the full depth of the theory’s anthropocentric roots.

3.1 The uncodifiability of virtuous language

The first challenge is to translate the concepts of virtue ethics into implementable computer models. This immediately becomes a difficult task since virtue ethics originated in ethical traditions through rich vocabularies of often interrelated, ambiguous, and higher-order mental concepts. In other words, the language of virtue ethics relies on thick descriptions, thick concepts,¹³ and folk psychology.¹⁴ To promote the traits that enable her to be courageous and fair, a virtuous person needs to have a thick understanding of courage and fairness to relate them to her own experience and motivations.¹⁵ “Courage” and “fairness” are also paradigmatic examples of thick concepts, i.e., terms can be characterized descriptively while simultaneously having an evaluative quality.¹⁶ Furthermore, since folk psychological notions such as belief, desire, and intention play a crucial role in our everyday ethical lives, to foster one’s character arguably implies an ability to grasp the way such concepts are grounded in mental states of others and oneself (Dennett 1989).

However, no AI systems can apprehend rich contexts, nor can they make use of a catalogue of subjective experience; nor do they possess interpretative mechanisms to disentangle value-laden terms or follow the logics of commonsense psychology. To that end, virtue ethics can be hard to compile even for a human being. One common criticism of the theory is that virtue ethics is “uncodifiable” and does not offer an applicable decision-procedure (Hursthouse 1999, pp 39–42).¹⁷ Faced with a particular moral dilemma, virtue ethics does not provide any straight-forward solutions; we simply have to trust that we do what a virtuous person would do.

The aforementioned difficulties have led many machine ethicists to avoid virtue ethics completely, others to argue that it is inferior to other approaches, or that it is simply incomputable due to its uncodifiability (Arkin 2007; Bauer 2020; Tolmeijer et al. 2020). But we will argue that there is a path for artificial virtue, provided that we give up on the project of trying to fully accommodate the theory’s anthropocentric foundations.

A simple version of the incomputability-argument can be constructed on the assumption that machines are essentially systems of automated rule-following. Since machines are governed by rules, and a virtuous person is not, it follows that virtue ethics is incomputable. This line of reasoning, however, ignores the fact that AI systems can be constructed of rudimentary rule-adhering units while the behavior of the larger system is not rule-following in the same sense. The relevant analogy is found in the similarity between the neurons of biological minds and the nodes of an artificial neural network. Biological neurons receive and transmit impulses according to the all-or-none law, meaning that they either produce a maximum response or none at all.¹⁸ Still, the human brain—consisting of roughly 86 billion neurons—is able to support complex processes that are not rule-following in the same way as its smallest components. After all, it is the very same network that gives rise to the thought and comprehension that enables a person to act virtuously. By extension, a large artificial neural network can produce a variety of behaviors that are not rule-adhering in the same narrow sense its nodes are.¹⁹

Howard and Muntean (2017) have extended a series of similar analogies between human and artificial cognition based on connectionism that they believe can pave the way for AMAs based on a form of virtue ethics. Drawing on Jackson and Pettit (1995) and Annas (2011), at the core of their framework is a “moral dispositional functionalism” which emphasizes “the role of the functional and behavioral nature of the moral agent” (Howard and Muntean 2017, p. 134). In their view, virtues are seen as dispositional traits that are nourished and refined through active learning of moral patterns in data, similar to how a cognitive system adapts to its environment (Howard and Muntean 2017). It is possible that Howard and Muntean’s vision might in the long-term solve some of the challenges posed by virtuous language; through active exposure to particulars, the AVA eventually learns to approximate the functional role of generic virtues, and how they are related to each-other in a complex whole. The first step towards AVAs is therefore not to create an artificial human being, fully equipped with the abilities required to grasp virtue ethics in a “top-down” fashion. It is rather to construct “bottom-up” learners who continuously interact and adapt to a dynamic environment, and through experience develop the appropriate dispositions depending on their functional role. Albeit lacking a reference to virtue ethics, previous technical work has already explored learning-methods to tackle the ambiguity of moral language. Using neural networks, Guarini (2006, 2013a, b) have taken a “classification” approach to moral data with a focus on the gap between particularism and generalism. After learning, Guarini’s models are able to classify cases as morally permissible or impermissible without the explicit use of principles. Similarly, McLaren have developed two systems—TruthTeller and Sirocco—that can learn and reason from moral data with the purpose of supporting humans in ethical reasoning (McLaren 2005, 2006).²⁰

Due to its inability to provide a decision-procedure, some still view virtue ethics as merely a supplement to the action-guidance provided by deontology and consequentialism. But others have found consolation in Hursthouse’s virtue ethics since it both provides (a) a decision-procedure in terms of rules, and (b) accounts for the developmental aspects of morality (Hursthouse 1999). According to Hursthouse, virtue ethics can offer action guidance through rules that express the terms of virtues and vices (“v-rules”), such as “do what is courageous” or “do not what is unjust”. While the list of virtues that yield positive rules of action is relatively short, it is complemented by a significant number of vices that can be expressed as negative rules (e.g., “don’t be greedy”). With regards to a decision-procedure, Hursthouse writes “P.1. An action is right iff it is what a virtuous agent would characteristically (i.e. acting in character) do in the circumstances” (p. 28).²¹

Even if Hursthouse’s decision-procedure is useful for the algorithmic implementation of virtue ethics, it might not be necessary as such from the view of system design. While Hursthouse’s virtue ethics can offer action-guidance for humans who are conflicted about what they ought to do, the same conflict does not necessarily arise for AVAs with dispositional virtues. An AVA that serves as a lifeguard and saves someone from drowning does so, not because of the conscious deliberation of decisions they could have made, but because they are courageous. Even if the agent in question followed an algorithm that can be described as a decision-procedure, the central focus is not the procedure itself, but rather the way it enabled the agent to save the person in danger. In fact, the concept of a decision-procedure, and the presumed requirement of having one to implement a normative theory, are entrenched with assumptions of how human rationality works. It assumes that ethical behavior is conducted in a rather stepwise algorithmic fashion, which in turn disregard the role of affective dispositions and enforces the sort of “particular situation → particular action” analysis akin to deontology and consequentialism. Since virtue ethics seeks to unite both thinking (e.g., conscious deliberations that can take the form of a decision procedure) and feeling (e.g., attitudes, emotions, desires) under term “character”, it should thus not be reduced to a description of the former.²² Simply put, a character is not a decision-procedure.

Thus, instead of taking a “top–down” approach to virtue ethics through its thick concepts as seen from a human perspective, a productive path forward is to start from some functional interpretation of virtue ethics that would carry out at least some important aspects of the theory. To that end, we have outlined an approach to AVAs that emphasize a holistic conception of character (involving both non-affective deliberation and affective dispositions) and bottom-up learning using artificial neural networks.

3.2 Virtuous capacities: rationality, autonomy, and consciousness

The second set of challenges is in many ways a corollary of the first: to what extent do AVAs rely on “higher-order” forms of moral capacities? Lacking human-like rationality, subjective experience, and autonomy, one might question whether artificial agents can be attributed moral agency at all. While such concerns perturb the overall possibility for machine morality, we will focus on how it challenges the prospects of artificial virtue. On the basis that it is both unfeasible and ethically problematic to equip artificial agents with human-like morality, we will argue that the development of AVAs should instead be driven by functional capacities that are shaped by normative considerations of how and to what extent AI systems should be involved in human practices.

In the context of moral machines, there has been widespread debate regarding the sufficient and necessary conditions for moral agency.²³ Central to these discussions are rationality,²⁴ autonomy,²⁵ and consciousness²⁶ (from now on collectively referred to as RAC). The good life for Aristotle’s animale rationale is life lived in accord with reason, implying an ability to follow reason, e.g., for holding beliefs and performing actions. It also entails introspective capacities of conscious deliberation and rational inquiry. As such, it differs significantly from the mere “goal directed behavior” of rational agents as conceived in AI development (Russell and Norvig 2020). Perhaps even more central to morality is autonomy, as it forms a basis for discussions about free will, moral agency, and responsibility.²⁷ In the Kantian tradition, a person is autonomous only if her actions, choices, and self-imposed rules are without influence of factors inessential or external to herself (Kant 2008). However, such accounts of autonomy are very different from the functional autonomy found in AI systems, where it roughly refers to the ability of doing something independent from human control.²⁸ Furthermore, it seems difficult to have ideas of “good” and “bad” without the conscious experience of positively or negatively valenced states. But from a neuroscientific and computational point of view, consciousness remains more or less as elusive as it was when Descartes wrote cogito ergo sum.

Beyond long-standing metaphysical debates, e.g. between free will and determinism (autonomy), and body and mind (consciousness), the prospect of artificial RAC is also ethically problematic, as it might result in human suffering,²⁹ artificial suffering,³⁰ or artificial injustice.³¹ There is also a danger that RAC and similar terms can be used to reproduce ideas that have been, and still are, used to justify abuse and hierarchies of dominance.³²

It seems clear that we cannot, at least in a near-term, construct AMAs with human-like RAC,³³ and even if we could, we need a much deeper understanding of RAC to even properly assess whether we should. The present project is, however, not to create artificial humans, but to construct agents that are able to serve important roles in human practices. As pointed out by Coleman, while Aristotle’s human arete emphasizes a life of contemplation and wisdom, the quest towards AVAs ought to be guided by exploring the android arete (Coleman 2001). For instance, it is possible to model rationality that allows artificial agents to effectively pursue goals without necessarily relying on the meta-cognitive abilities of human rationality. Besides self-legislative autonomy, there are flexible ways to construe artificial autonomy that enable human operators to oversee, intervene, or share the control of the system to avoid unwanted consequences.³⁴ Additionally, learning systems can employ simple reward functions that functionally mimic aspects of the role subjective preferences have in human cognition, without the phenomenological experience of suffering.³⁵

More importantly, following the “normative approach” to artificial moral agency (Behdadi and Munthe 2020), we believe that the android arete should be guided and constrained by the normative discussions of how AI systems should engage in human practices that normally presuppose responsibility and moral agency. That is, rather than focusing on theoretical discussions on whether AI systems can have moral agency,³⁶ the development of AVAs should be led by the ethical and practical considerations that relate to their specific role, e.g., as doctor assistants, chauffeurs, or teachers. In turn, this allows us to shift focus from general questions about moral capacities based on human-like RAC, to particular issues about how and whether certain AI systems should be incorporated into specific ethical domains, what moral roles they could potentially excel at, and how agency and responsibility ought to be allocated in those circumstances.

Besides carrying out a role, it is also possible that the envisioned virtuous agents could still acquire a certain moral status. In the context of social robotics, Gamez et al. (2020) have argued that AVAs could claim membership to our moral community on the basis of two separate but consistent views on moral status: behaviorism (Danaher 2020) and the social-relational approach (Coeckelbergh 2010; Gunkel 2018). According to the former, an artificial agent has a moral status if they are functionally equivalent to other moral agents.³⁷ In the latter, the moral status of an artificial agent depends on the meaningful social relations we develop with it, such as reciprocal trust, duties, and responsibilities. In their experiment, Gamez et al. (2020) found that while individuals made weaker moral attributions to AIs in comparison to humans, they were still willing to view the AIs as having a moral character. Thus, even if AVAs lack the unique metaphysical qualities of human morality, if we are willing to describe them as having a character—based on their behavior and our relationship to them—it could be sufficient reason to welcome them to our moral community.

3.3 Virtues, moral exemplars and eudaimonia

The third challenge is to decide the role of virtues in the moral cognition and behavior of AVAs. From Homer to Benjamin Franklin, many different lists of virtues have seen the light of day, emphasizing different aspects of ethical life (MacIntyre 2013). Some lists have been more prominent than others, in particular the cardinal virtues; prudence, justice, fortitude and temperance.³⁸ This might suggest that one could feed an artificial virtuous agent with widely accepted virtues, or generic virtues suitable for machines.³⁹ However, this solution would only be an option if there was (i) a universally agreed-upon list of the most essential virtues, and (ii) a way to implement said list in a top-down fashion. As argued in 3.1, even if a list was attainable, the approach to virtue has to be bottom-up since the only way to reach context-sensitive generals are through particulars. We therefore agree with MacIntyre’s historical analysis, that virtues ought to be based in a particular time and place, emerging out of the community in which they are to be practiced (MacIntyre 2013).

Still, this leads us to the question: in what way should a virtuous agent learn bottom-up? Inspired by Hursthouse (1999) and Zagzebski (2010), previous work in artificial virtue have centered on imitation learning through the role of moral exemplars (Berberich and Diepold 2018; Govindarajulu et al. 2019). The moral exemplar approach offers several appeals. By mimicking excellent virtuous humans, we do not have to worry about what virtues they in fact end up with since they would replicate something that is already virtuous. Besides providing means of supervision and control, imitation learning would also solve the alignment problem; i.e., the challenge of aligning machine values with human values (Berberich and Diepold 2018).

Nevertheless, there are issues with the moral exemplar-focus; in particular the challenge of deciding who is a moral exemplar and why. After all, there could be severe disagreements about who is and who is not a virtuous person. According to Zagzebski (2010), exemplars can be recognized through the emotion of admiration, which allows us to map the semantic extension of moral terms to features of moral exemplars. Govindarajulu et al. (2019) have provided a rudimentary formalization of Zagzebski’s suggestion using deontic cognitive event calculus (DCEC). In their model, admiration is understood as “approving (of) someone else’s praiseworthy action” (p. 33), which depends on a primitive emotional notion of pleased or displeased based on whether an action led to some positive or negative utility. Using the utility of consequences to define emotions, however, their model seems to be driven by consequentialism rather than virtue ethics.⁴⁰

Besides moral exemplars, we suggest that there is an alternative source for moral evaluation appropriate for bottom-up virtuous agents to be found in the concept of eudaimonia (conventionally translated as “well-being” or “flourishing”). Instead of relying on moral exemplars or a list of anthropocentric virtues, Coleman (2001) has argued for an eudaimonist approach to artificial virtue, where “all of one’s actions aim at a single end—in Aristotle’s case, happiness (eudaimonia)—and virtues are those character traits which foster the achievement of this end” (p. 249). Besides avoiding circular definitions of virtue (e.g., “virtues are qualities of virtuous individuals”), eudaimonia can explain the nature of virtues in terms of a goal or value to strive towards.⁴¹

An eudaimonist virtue ethics offers several benefits for the prospect of AVAs. Essentially, it enables us to model virtues in terms of their relationship to eudaimonia. If eudaimonia is defined as “increase moral good X”, virtues would then be the traits that help the agent to increase X. In machine learning terms, eudaimonia can be seen as the reward function that informs the learning and refinement of virtues and virtuous action. In this way, the artificial agent will become virtuous in the sense that it develops the dispositions that enable it to effectively pursue a certain goal or increase a certain value (depending on whether eudaimonia is defined as a goal or value). Another strength is that, while learning through imitation is limited to mere behavior, an eudaimonist approach can encompass values both intrinsic (e.g., hedonistic pleasure and pain) and extrinsic (e.g., values that support human ends).⁴² Furthermore, in cases where it is hard to settle on a suitable moral exemplar, a functional eudaimonia offers a “top-down backdoor” to implement certain values or goals that are then attained through a bottom-up learning process.

But adopting a eudaimonist view raises the further question: what should be the eudaimonia of virtuous agents? While the content of eudaimonia can be defined as the goal that ought to be achieved, or the good that ought to be increased, we believe that the important function of eudaimonia is that it provides a moral direction for the virtuous agent; a measure to evaluate and refine its moral character and virtues. However, this omits the difficult task of pinning down the actual goal or values an AVA should have.⁴³ But there are good reasons to remain cautiously silent on the de facto content of an AVA’s eudaimonia. First, it allows us to use eudaimonia as a functional placeholder for a wide variety of values and ends, on the premise that they can be implemented in computational systems (which we will discuss in Sect. 4.3). For practical reasons, it might be suitable for different AVAs to have different types of eudaimonic content depending on their functional role. Second, it recognizes the ambiguity of human eudaimonia, especially as it remains unclear whether and to what extent artificial systems can apprehend the complexity of the former.⁴⁴ This echoes Hursthouse (1999), who views eudaimonia as a value-laden concept that is intentionally ambiguous to allow for interpretative headroom and disagreement.⁴⁵

3.4 Technical implementation

The remaining challenge is to move from the conceptual plane towards technical implementation. We do so by examining previous work in artificial virtue, focusing on what it can fruitfully provide for the development of AVAs.⁴⁶

Howard and Muntean (2017) have put together a web of conceptual foundations for the construction of artificial autonomous moral agents (AAMAs). Beyond moral functionalism and bottom-up learning, they conjecture a number of, in their words “incomplete and idealized analogies” (p. 137) between human cognition and machine learning that can guide the deployment of AAMAs through a combination of neural networks and evolutionary computing. However, the actual details of the implementation are missing, and only partial results of an experiment are provided where neural networks have learned to detect irregularities in moral data.⁴⁷ The biggest flaw with their project, however, is that it is practically infeasible,⁴⁸ and furthermore, it is not clear how their envisioned agents should be deployed in moral situations besides classification tasks explored by Guarini (2006).

In a similar vein, Berberich and Diepold (2018) have explored the technical underpinnings of artificial virtue based on connectionist methods. Most interestingly, they have described how reinforcement learning can be used to shape the moral reward function of virtuous agents in three ways: (i) through external feedback from the environment, (ii) internal feedback by means of self-reflection, and (iii) observation of moral exemplars. However, besides providing a broad outline of potential AVA features, they offer no finer details of how such an agent could be constructed.

Thornton et al. (2016) have incorporated principles from virtue ethics along with deontology and consequentialism into the design of automated vehicle control. In their hybrid model, deontology determines vehicle goals in terms of constraints, consequentialism in terms of costs, and virtue ethics is used to regulate the strength of the applied costs and rules depending on the vehicle’s “role morality”. “Role morality” refers to behaviors that are acceptable given the context of a particular professional setting.⁴⁹ For instance, it is acceptable for an ambulance to break traffic laws—e.g. by running a red light—if it transports a passenger with life-threatening conditions. Essentially, Thornton et al. (2016) shows how the moral character of an AI system can be defined with regards to their societal role, and how the character can be modeled using “virtue weights” that balance costs and constraints that enables them to perform their function.

4 Artificial virtuous agents

Based on the takeaways from our investigation so far, we now turn to the task of interpreting the outlined theory in terms of functionality, which in turn can guide the further development of AVAs. The aim is not to provide a detailed implementation per se, but rather to discuss suitable methods that can be combined to functionally carry out features of virtue ethics in a variety of moral environments. Essentially, the viability of the proposed framework rests on the assumptions that (i) the function of dispositional virtues can be carried out by artificial neural networks (Sect. 4.2), (ii) eudaimonia can be functionally interpreted as a reward function that drives the training of the virtue networks (Sect. 4.3), and (iii) modern learning methods can in various ways support the development of artificial phronesis (Sect. 4.4).⁵⁰

4.1 Artificial character

In virtue ethics, a moral character can be defined as the sum of an agent’s moral dispositions and habits (George 2017). In our functional model, it consists of several components: a set of stable yet dynamic dispositions (virtues), a reward function (eudaimonia), a learning system (phronesis), and relevant mechanisms for perception and action determined by its role (e.g. input-sensors, memory, and locomotion). The moral character thus denotes the entire character, encompassing both moral qualities and non-moral qualities that enable it to perform its role. Virtues are “stable yet dynamic” in the sense that they are fixed at a given moment but have the ability change over time. Importantly, instead of applying a decision-procedure to a particular situation, the virtuous character continuously interacts within an environment based on its internal states.

4.2 Artificial virtues and vices

Given our eudaimonist view, virtues are defined as the character traits an agent needs and nourishes to function well in light of its eudaimonia. We extend the weight analogy of Thornton et al. (2016) and the classification approach of Guarini (2006) and suggest that the functional aspect of virtues can be captured in the function of nodes in an artificial neural network. The essential role of virtues in this view is that they, based on some input from the environment, determine the action taken by the agent (e.g., its output). In the simplest case, a virtue can be modeled as a perceptron that determines whether an agent acts in one way or another given a certain input. A perceptron is a threshold function that takes an input $x$ and produces the output value $f\left( x \right) = 1$ if $w \times x + b > 0$, where $w$ is the weight (or set of weights) and $b$ is the bias. By either increasing or decreasing the weight through feedback, the perceptron is effectively a learning algorithm for binary classification. In a system of perceptrons, the nodes themselves represent virtues since there is only one neural unit for every virtue (as illustrated in Fig. 1). This solution is suitable if there are only two possible actions (e.g. “save” or “don’t save”), and the binary virtues need to find an appropriate balance between two distinct vices (e.g. “courage” as a balance between “reckless” and “coward”).⁵¹

Fig. 1

Illustration of a network with three virtuous perceptrons. Three types of inputs are parsed into three corresponding virtues weights, each holding a value between two extremes (in this case represented numerically from − 1 to + 1). Each perceptron produces one out of two possible actions depending on its weight. For instance, if another agent is in danger, an agent with a positive courage weight (> 0) will try to help the other agent even if it poses a risk

In more complex applications, a virtue can consist of a larger network of nodes that receive an input and output one of many possible actions. In this view, virtues can be seen as a “higher level” amalgamation that encompasses a large set of particular “lower level” units (see Fig. 2). More nodes enhance the ability to process more detailed information that can in turn be used to produce more fine-tuned actions.

Fig. 2

Illustration of a system with three virtuous networks, each with an input layer, two hidden layers, and an output layer. Compared to the perceptron, a deep neural network can deal with classification tasks that are not linearly separable

Deciding on what virtues to implement and the actions they ought to perform entirely depends on context and functionality of the AVA in question. It also informs the choice between static and dynamic virtues and weights, e.g., whether one implements the virtues and weights one prima facie believes to be suitable for the AVA, or let them learn independently in light of an eudaimonic reward (discussed in Sect. 4.3). In an environment with a fixed set of possible actions but a wide range of environmental inputs, it would be suitable to provide the agent with static virtues relating to the fixed set of actions, but with dynamic weights in order for them to learn the appropriate action given a specific input. In an environment where we already know that a particular input always ought to be followed by a particular action, it would be more appropriate to provide the virtue with a static weight.⁵² However, in highly dynamic and noisy environments with a potentially infinite number of possible actions, the agent might be limited to unsupervised learning and reinforcement learning in accord with a reward function.⁵³

In addition to the choice between static or dynamic, it is also important to consider how the system deals with the conflict problem, i.e., the issue that arises when two or more virtues suggest different actions.⁵⁴ For instance, in a social situation, compassion might tell us to remain silent while honesty urges us to convey some painful truth. One solution is to resolve conflicts through mere comparison of strength, i.e., given a particular situation, the right action depends on what the most dominant virtue tells the agent to do. If an agent is more fair than selfish, it will still give food to the begging other. In a computational setting, such conflicts can be resolved by simple arithmetic; if fairness weight = 0.6 and selfishness weight = 0.4, then fairness > selfishness. A related solution is to model virtues in a pre-given hierarchy of priorities. Another option is to train the input-parsing network of the virtuous agent so it learns to map inputs to the appropriate virtue-network. This could be achieved through a process of supervised learning, where the network is presented with many pre-labeled scenarios that are related to specific virtue-networks.⁵⁵ Yet another option is to model more sophisticated forms of hybrid virtues that can combine and parse different aspects of inputs that relate to different virtues. Essentially, as there are many practical ways to combine and resolve conflict between virtues, we do not believe the conflict problem raises any serious concerns for the prospect of AVAs.

4.3 Artificial eudaimonia

Approaching artificial eudaimonia, the main challenge is to provide AVAs with a definition of eudaimonia that can be practically implemented in computational systems. Towards a possible solution, we propose a functional distinction between eudaimonic type (e type) and eudaimonic value (e value). We define e type as the kind of value or goal the virtuous agent strives towards (i.e., the eudaimonic content as described in Sect. 3.3), and e value as the quantitative measure of how much e-type that has been attained. Importantly, e type and e value provide the basis for learning, as the dispositional virtues change in light of whether an action increased or decreased the e value. To be functionally implementable, an e type must represent a preference to increase e value given by some identifiable measure (e.g., a quality or quantity), and that e value can increase or decrease through the agent’s actions. An e type can for instance be a preference to increase praise and decrease blame received from others. In that case, the agent needs to have some way of receiving feedback on their actions, and to qualitatively recognize the feedback as either praise or blame. With an e type defined by praise/blame-feedback, the agent’s e value will then increase if it receives praise and decrease if it receives blame. Another way of illustrating a functional e type is through resources. We can imagine that there is some quantifiable resource that agents need to survive (e.g. food). A selfish e type could then be defined as a preference to maximize the possession of said resource, and a selfless e type could conversely be defined as a preference to give resources to others. The selfish agent then increases its e value through actions that increase the resource (e.g., begging or stealing), and the selfless agent increases e value by giving.

One might question whether and to what extent our rather simplistic conception of eudaimonia can, in some meaningful sense, capture the variety and complexity of human values. In particular, our proposal rests on the rather strong assumption that moral goods and goals can be described in quantifiable measures. Furthermore, since we leave the choice of e type to human developers, it also raises the question whether the implemented values will be justified in relation to the AVA’s role.⁵⁶ While we do not have any definite answers to these issues, we believe that parts of the solution lie in further developments and experimental studies, and that our model provides a starting point for such endeavors. Even if no current AI system can apprehend the full depth of human values, it does not exclude the possibility for there to be some moral domain where some quantifiable moral good can be legitimately increased through computational means.

4.4 Artificial phronesis

While phronesis (“practical wisdom”) is a rather ambiguous concept within the virtue theoretic tradition,⁵⁷ in our functional simplification, we take artificial phronesis to broadly refer to the learning an agent receives from experience. This interpretation is motivated by recognizing the central role learning plays in cognition, for moral and non-moral capacities alike (Annas 2011). We will address four learning aspects of artificial phronesis, namely what is learned, how it is learned, the source of learning, and the technical method used.⁵⁸

For what and how, we make a distinction between (1) learning what action leads to good and (2) learning what is good in itself.⁵⁹ Based on our conception of artificial eudaimonia, (1) can be seen as the instrumental means to increase e-value according to some e-type, whereas (2) refers to an ability to change and refine the teleological component itself (e-type). We identify three features of (1) that we believe are crucial for the development of artificial phronesis and describe how they can be acquired: (1a) virtuous action, (1b) understanding of situation, and (1c) understanding of outcome.

(1a) Virtuous action refers to the ability to perform virtuous action, i.e., to act courageously or fairly. Given the action-determining role of virtues in our model, the key element in the development of virtuous action is to fine-tune virtues in light of eudaimonic feedback. However, to do so, the agent needs (1b) understanding of situation and (1c) outcome.

(1b) Understanding of situation refers to the ability to comprehend a situation, i.e., to know what input relates to what virtue; for instance, whether a situation calls for courage or honesty. (1c) Understanding of outcome, on the other hand, is the ability to comprehend what was the actual result of a performed action. While virtuous action can be trained by means of reinforcement using eudaimonic feedback, we suggest that (1b) can be carried out by an input parsing network trained on labeled data of various moral situations, i.e., by means of supervised learning.⁶⁰ Similarly, (1c) can consist of a network trained to recognize an outcome in light of the relevant quantity or quality at stake as defined by its e type, e.g., by learning from a dataset of labeled reactions. For instance, in a simple environment driven by praise/blame reactions, the role of the outcome network is to accurately classify a reaction as being either “praise” or “blame”.

The combined role of these features can be illustrated in the following example. An agent receives input from the environment in terms of another agent in need of help. The input parsing network classifies the situation as involving courage, and therefore sends the input to the courage network. In turn, the courage network assesses the risk of the situation at hand and determines whether it calls for a certain action (e.g., depending on the network’s balance between recklessness and cowardice). The outcome of the performed action is then read by the outcome network as a new input, and parses it to the eudaimonic reward system, which evaluates whether the action increased or decreased e value. If e value increased, positive reinforcement is sent to the courage network, in effect teaching the courage network that the performed action was appropriate. By contrast, negative reinforcement will reduce the likelihood that the same action will be performed given a similar scenario in the future.

Beyond trial-and-error (through external feedback) and supervised learning (provided by developers), AVAs could also learn from internally generated feedback (Berberich and Diepold 2018). We describe two possible routes that could guide the construction of an internal learning system, namely retrospective and proactive reflection. Given that an agent has the ability to store mappings between input-action-outcome-reward, it could analyze that information to retrospectively reinforce certain actions. Identifying patterns in past behavior, retrospective reflection could in turn form the basis of more nuanced behavior in complex environments based on statistical considerations (e.g., by applying non-linear regression). Proactive reflection, on the other hand, could be achieved through internal simulation of possible scenarios, where learning feedback is based on trial-and-error of hypothetical input-virtue-action-outcome-reward mappings.⁶¹

Another potential source for learning is the behavior and experiences of others, regardless of whether they are exemplars or not.⁶² If an AVA can observe that the outcome of another agent’s action increased some identifiable moral good (defined by the observer’s e type), it could teach the observer to positively reinforce the same action.⁶³

Evolution offers yet another potential source for learning. This could be achieved through the use of evolutionary computation (Howard and Muntean 2017), or other randomized search methods. The main idea behind evolutionary algorithms is to find candidate solutions to optimization problems using processes inspired by biological reproduction and mutation, along with a fitness function that evaluates the quality of the solutions (Bäck et al. 1997). A possible application to the development of artificial phronesis could therefore be to (i) generate an initial population of agents with virtues and other suitable parameters set randomly, (ii) evaluate the fitness of every individual according to how much e-value they have attained, (iii) select the most virtuous individuals for reproduction and generate offspring using crossover and mutation, (iv) replace the least virtuous individuals with the new offspring, and repeat steps (ii)–(iv) until the population is sufficiently virtuous.

We have thus far only been concerned with (1) learning what action leads to good as opposed to (2) learning what is good in itself. The latter can be achieved by modelling dynamic e-types that have the ability to change. There is an intuitive appeal for such an endeavor, as the ability to change our (human) concept of eudaimonia provides an important basis for personal, social, and moral progress. Yet, it also presents a puzzling paradox—how can we, on the basis of our current set of values, assess whether another set of values are more appropriate?⁶⁴ Beyond such meta-theoretical issues, there are good reasons why dynamic e-types in the context of moral machines should be approached with caution, as it can potentially pave the way for nonalignment of human-AI values⁶⁵ and reward hacking.⁶⁶

Still, we believe that there are a few suitable venues to explore dynamic e-types. The first is through the use of moral exemplars (discussed in 4.5), and the second is by means of a metaheuristic at system level. In a multi-agent environment, there can be some potential “higher good” to be achieved at a system level that cannot simply be resolved at the level of individuals. For instance, in a “tragedy of the commons” situation, individuals act in their own self-interest even though the collective action of the many creates catastrophic problems for everyone (such as a systemic collapse). Using randomized search methods in a multi-agent simulation of virtuous agents, dynamic e-types could then be used to identify e-types that satisfy hedonistic needs of individuals while simultaneously ensuring the prosperity of the entire population at large.⁶⁷

Finally, we will briefly discuss technological methods that can be used to develop artificial phronesis. Methods in machine learning are conventionally divided into supervised, unsupervised, and reinforcement learning, each offering their own unique set of advantages and drawbacks (Russell and Norvig 2020). In the first, a function learns to map the correct input to output based on labeled training data; in the second, it learns to categorize and find patterns in unlabeled data on its own; in the third, an agent learns to make actions that maximize some cumulative reward through feedback. Agreeing with Berberich and Diepold (2018), we believe that reinforcement learning (RL) provides the most appealing approach for AVAs, as it, contrary to the other two, is based on dynamic interaction with an environment and thus supports a continuous process of learning from experience (so called “online learning”). We have already described how RL constitutes the basis for the eudaimonic reward system, where e type corresponds to the reward function of the RL agent, and e value is the measure of how much reward is attained. With that said, supervised and unsupervised methods can also be fruitfully integrated into our model. In particular, we have described how supervised learning can be used to train the input parsing and outcome networks of AVAs. Unsupervised learning such as cluster analysis could also find suitable applications in the form of anomaly detection, for instance, by helping the agent to identify deviations and outliers in the internally stored moral data.

4.5 Moral exemplars

To incorporate moral exemplars in our framework, we have to provide some good solutions to two challenges, namely (i) how to pick a suitable moral exemplar, and (ii) how to learn from them. From a practical point of view, and given the limited capacities of current AMAs (Cervantes et al. 2020), the most obvious moral exemplar for AVAs is the human exemplar. While the use of human exemplars raises its own set of issues (see Sect. 3.3), it is nevertheless the human developer who defines and implements the e-types, learning system, virtues, and evaluates agent performance. But we will briefly outline a possible approach to moral exemplars that can be useful in computational settings derived from our model. The conditions are the following:

One agent (X) takes another agent (Y) as a moral exemplar if

• i. X and Y have the same e type, and

• ii. Y has a higher e value than X.

The first condition means that X and Y strive towards the same set of values and goals. The second means that Y has been more successful in achieving the same set of values and goals, e.g., due to its virtuous behavior.⁶⁸ Since X wants to achieve the same thing as Y (same e type) and recognizes that Y has some means of achieving it more effectively (more e value), it is reasonable for X to take Y as a moral exemplar. Although these conditions are difficult to model in the interaction between AVAs and humans (it would require that humans have a formally defined e type), we believe it can be implemented in the interaction between different AVAs, (e.g., in multi-agent simulations).

Adopting the outlined approach to artificial moral exemplars, the second issue—regarding how an agent should learn from exemplars – depends on the information provided by the environment. If agents only have access to the external behavior of others, we are limited to learning through behavioral imitation. However, in that case it might be impossible for agents to determine whether they should adopt a moral exemplar at all, since they would not have access to the e type or e value of others.⁶⁹ Alternatively, if all internal aspects of AVAs were accessible, agents could adopt moral exemplars by simply copying relevant aspects of their character, e.g., the structure and weights of the virtue networks. If they, on the other hand, had access to the e type and e value of others but not the virtue networks as such, it would still provide sufficient reason, given the conditions outlined earlier, to adopt exemplars and learn from them through behavioral imitation.

4.6 Software architecture

The final challenge is to explicate the appropriate connections between the different components of the AVA.⁷⁰ Ideally, the connections should be drawn in a way that effectively exploits component functionality while leaving room for learning through continuous exploration. We believe that the functionality of the discussed features informs such a design, and conclude by a stepwise explanation of the core aspects of a generic architecture that can guide the development of AVAs (Fig. 3):

Fig. 3

Generic architecture of the proposed AVA. Input from the environment is classified by the input parsing network and sent to the appropriate virtue network (in the figure represented as V1–V5). Each virtue network represents a character trait that can produce a range of different actions. The outcome of the performed action is taken as a new input by the outcome network and evaluated by the eudaemonic reward function. This in turn informs the learning system whether a particular action of a particular virtue is to be positively or negatively reinforced

• 1. Input parsing network: input from the environment is classified by the input parsing network. Its main role is to transmit the environmental input to the appropriate virtue network. It corresponds to understanding of situation, i.e., to know what virtue applies to a particular situation. The network can be trained through supervised learning using labeled datasets of various ethical scenarios. One critical aspect of the network is to determine whether a situation calls for moral action or not,⁷¹ and if the latter, whether it can constitute a basis for learning through observation.

• 2. Virtue networks: the invoked virtue network classifies the input to determine the most appropriate action. The number of nodes and layers depend on environmental complexity and the number of possible actions. If the input can be linearly separable into one out of two actions, a single perceptron can carry out the classification, but in complicated cases, a deep neural network would be more suitable.

• 3. Action output: the action determined by the virtue network is executed by the agent. This could in principle be any type of action; from simple movements and communicative acts to longer sequences of skillful action.

• 4. Outcome network: new environmental input is classified by the outcome network, corresponding to understanding of outcome. Similar to the input parsing network, it can learn from labeled data of situation-outcomes, in particular in light of the relevant moral goods as defined by its e-type.

• 5. Eudaimonic reward: the eudaimonic reward function evaluates the classified outcome according to e type and current e value. If the e value is increased, it sends positive feedback to the learning system, and negative feedback if e value decreased.

• 6. Phonetic learning system: the learning system reinforces the relevant virtue according to the eudaimonic feedback. Another possible application of reinforcement learning is to send feedback to the input network. For instance, if the negative feedback was exceptionally high, it might suggest that the input parsing network transmitted the environmental input to an inappropriate virtue network; if exceptionally positive, the input parsing can be positively reinforced.

• 7. Observation and moral exemplars: two additional sources for learning can be implemented in the form of (i) observation of others and (ii) moral exemplars. In the first case, if another agent’s action increased or decreased some identifiable e type, the observing agent could positively or negatively reinforce the same action.⁷² In the second case, if an agent adopts a moral exemplar (given conditions described in 4.5), it could learn from it by either copying aspects of its character or by mimicking its behavior.

5 Discussion

We have described how AVAs can be constructed in a way that functionally carries out a number of core features of its theoretical counterpart, including virtuous action, learning from experience, and the pursuit of eudaimonia. We believe the development of both simpler and more advanced virtuous systems can be guided by the presented framework. In a minimal case, the training of an AVA can solely rely on the feedback of its own actions. More advanced agents could potentially learn from passively observing the behavior of other agents and moral exemplars, or by means of internal feedback systems (e.g. retrospective and proactive reflection).

Bauer (2020) has argued that a two-level utilitarian approach to AMAs is superior to a virtue-theoretic approach since it encompasses the essential features of the latter while realizing additional benefits. He gives four reasons to support his claim: (1) conditional rules can serve the function of dispositional traits, (2) utilitarian AMAs avoid reference to the intangible concept of ‘virtue’ that obfuscates the design of AMAs, (3) utilitarian AMAs can be pre-loaded with widely agreed-upon rules, such as human rights and legal codes, and (4) since utilitarian AMAs would follow moral rules, they would be ethically better than “typical human behavior” whereas AVAs, being modeled on human behavior, would not. We believe our work shows that Bauer is misguided on all four points. Responding to (1), the sole purpose of dispositional traits is not to yield rule-following behavior as such, but rather to produce moral behavior in context-sensitive situations where simple rule-following principles are not applicable. Answering (2), while we agree that “virtue” is in many ways an intangible concept, we have shown that it can be given a functional definition within our eudaimonic framework. Against (3), we believe that connectionist learning offers the most effective methods to implement generic rules into AI systems so that they are carried out appropriately. Furthermore, a functional e-type allows for top-down implementations of widely agreed-upon values, provided that such values can be formalized. Responding to (4), we agree that imitation learning from human exemplars is not sufficient, which served as reason to adopt a teleological approach to machine ethics.

Given how eudaimonic reward trains virtues in light of outcomes, one could argue that the presented model is simply “consequentialism with an extra layer”.⁷³ But that would miss the point. Although teleological virtue ethics relies on some definition of moral good, it emphasizes the learning and dispositional aspects of how certain goods could in fact be increased. That is, while a form of consequentialism drives learning, it is the character’s dispositional virtues that produce the actions.

More generally, we do not believe that a virtue-theoretic approach is superior to deontology or consequentialism in every regard, but rather that it draws our attention to important aspects of morality that are overlooked in the field of machine ethics. After all, an artificial entity would only be truly virtuous if it could follow moral rules and be sensitive to the consequences of its actions. Ultimately, we believe that a hybrid-approach to machine ethics is most suitable as it could potentially realize the benefits of the three grand theories. However, since virtue ethics digs deeper into what it is to be a moral agent, we believe it offers a sketch of what a hybrid system could look like. If AMAs were ever to possess a character or belong to our moral community, they would indeed share many aspects with an agent as perceived through the virtue-theoretic lens.

Nevertheless, a number of issues remain to be solved before we see the introduction of morally excellent AVAs in our everyday lives. Conceptual work is needed to resolve the conflicts between anthropocentric notions of morality and the formalization of such concepts in AI development, and technical work is needed to bring AMAs into the complex ethical environments of the real-world. For the prospect of AVAs presented in this work, one critical issue is the lack of explainability in neural networks, often referred to as the black box problem.⁷⁴ Although it is unclear whether the issue can ever be completely resolved, we believe it can be approached carefully.⁷⁵

So what kind of AVAs can and ought to be constructed using our framework? We will outline some potential applications that can serve as venues for future work in the short-, mid-, and long-term.

For now, the current stage of artificial virtue is prototypical and confined to well-defined software environments and limited robotic tasks. Following the classification approach (Guarini 2006), AVAs could be trained to solve a range of moral classification tasks, including action selection, situation reading, and outcome understanding. For instance, one interesting venue for future work is to explore technical solutions to the conflict problem (i.e., when two or more virtues suggest different actions). Another application is to implement AVAs in multi-agent systems to study cooperation among self-interested individuals⁷⁶ or other forms of complex social behavior.⁷⁷

In the mid-term, virtue ethics may be incorporated into a number of AI systems in real-world domains, particularly in complex environments where bottom-up learning offers the only route to moral sensitivity. Additional benefits can be achieved by integrating virtue ethics in human-inspired architectures (Cervantes et al. 2016). In that way, artificial virtue can be equally propelled by advancements in brain science as from new AI methods. Further developments in the mid-term could potentially explore more sophisticated models and methods for learning, reasoning, communication, social cognition, and computational autonomy, provided that such capacities are desirable and ethically justified in relation to the AVA’s specific role (Behdadi and Munthe 2020).

In the long-term, AVAs might, as argued by Gamez et al. (2020), be legitimate members of our moral community. Most optimistically, AVAs might not only be morally excellent, but become moral exemplars to humans by conveying forms of morality that are yet to be discovered in our everyday moral landscape. However, as we discussed in Sect. 3.2, such projects should be approached with utmost caution and care; the development of increasingly more sophisticated AI systems in the moral domain walks a risky path of potentially causing an explosion of suffering, for human and artificial beings alike.

6 Conclusion

We have broadly explored various philosophical and technical dimensions of virtue ethics and developed a comprehensive framework for the construction of artificial virtuous agents based on functionalism, bottom-up learning, and eudaimonia. To our knowledge, it is the first work that presents a roadmap to artificial virtue that is conceptually thorough yet technically feasible. Ultimately, we believe that it offers a promising path towards excellent moral machines and hope that our work will inspire further developments of artificial virtue.