Living well with AI: Virtue,education,and artificial intelligence

Abstract

Artificial intelligence technologies have become a ubiquitous part of human life. This prompts us to ask, ‘how should we live well with artificial intelligence?’ Currently, the most prominent candidate answers to this question are principlist. According to these approaches, if you teach people some finite set of principles or convince them to adopt the right rules, people will be able to live and act well with artificial intelligence, even in an evolving and opaque moral world. We find the dominant principlist approaches to be ill-suited to providing forward-looking moral guidance regarding living well with artificial intelligence. We analyze some of the proposed principles to show that they oscillate between being too vague and too specific. We also argue that such rules are unlikely to be flexible enough to adapt to rapidly changing circumstances. By contrast, we argue for an Aristotelian virtue ethics approach to artificial intelligence ethics. Aristotelian virtue ethics provides a concrete and actionable guidance that is also flexible; thus, it is uniquely well placed to deal with the forward-looking and rapidly changing landscape of life with artificial intelligence. However, virtue ethics is agent-based rather than action-based. Using virtue ethics as a basis for living well with artificial intelligence requires ensuring that at least some virtuous agents also possess the relevant scientific and technical expertise. Since virtue ethics does not prescribe a set of rules, it requires exemplars who can serve as a model for those learning to be virtuous. Cultivating virtue is challenging, especially in the absence of moral sages. Despite this difficulty, we think the best option is to attempt what virtue ethics requires, even though no system of training can guarantee the production of virtuous agents. We end with two alternative visions – one from each of the two authors – about the practicality of such an approach.

Introduction

In 2014, Stanford ran a panel debate entitled ‘Does teaching ethics do any good?’ In the last two decades, universities have increasingly dedicated themselves to ethics – particularly, applied ethics – curricula (Darwish, 2015). Public attention to and concern about the rapidity of scientific and technological progress has motivated a growing emphasis on applied ethics, and particularly a desire for applied ethics courses in higher education (Darwish, 2015). However, a 2013 study in ethics in engineering showed that moral education did not improve either moral judgment or knowledge of responsible conduct in research (May and Luth, 2013). The ineffectiveness of ethics teaching is concerning, especially if courses in applied ethics are one conduit through which we expect the next generation to learn how we might live well with the artificial intelligence (AI) that increasingly inhabit our world. In particular, we need effective and flexible ethical training that can prepare future generations for living in a world in which novel ethical situations crop up with novel technologies and their applications and for designing AI systems that are likely to benefit rather than to harm.

AI tools are increasingly present in our lives. In this article, we consider the ways in which we live with the various autonomous tools, software, applications, and so on that are not directly programmed but rather modify themselves based on data during a training process, which we call ‘AI’. Discussions of how AI will transform the world abound, both in traditional media and online, particularly since the release of ChatGPT-3.5. At the heart of these discussions is the question of what it takes to live well with AI and how AI can be used to improve our lives without damaging the things we cherish. Given these questions, we must consider how to design, build, implement, govern, and regulate AI technologies. We must consider the normative and applied moral concepts needed to live well with AI, and how we might train future generations to ensure that they can engage with and deploy these concepts in their lives.

University ethics courses are one venue for training the next generation to act well. Increasingly, universities are requiring ethics courses – usually applied ethics – for those who might one day become the creators of novel AI. One dominant approach in applied ethics curricula is the principlist approach. ¹ According to this approach, living well with AI requires that we develop, follow, and instill in people norms and principles that will appropriately govern the way we live with AI. Many argue that if we simply find the right rules, follow them appropriately, and ensure they are passed on, we will live well with AI. This seems particularly problematic in the case of AI; rules may become quickly obsolete and tend to be reactive rather than anticipatory. In other words, in a rule-based system, humans are likely to be chasing after the right way to live, rather than providing a dynamic, living rubric as to what it means to live well with new technology. We think a more fruitful approach for training the next generation to live and act well with AI is Aristotelian virtue ethics.

We argue that Aristotelian virtue ethics provides all of the theoretical tools needed to guide our moral decision making in developing and implementing AI. Aristotelian virtue ethics is a useful guide to thinking about ethical action when the future is opaque, because it is a highly flexible system that is responsive to the ethically salient features of each situation. The reason that it is flexible is that virtue ethics is agent-based rather than action-based. Virtue ethics does not prescribe a set of rules, and it requires exemplars who can serve as a model for those learning to be virtuous. Virtue is similar to a skill, which is acquired through guidance from more experienced moral actors, practice, and feedback. The virtuous person has a skill set which allows them to determine the ethically salient features of a situation, identify the appropriate virtue required, and be motivated to take appropriate action.

However, implementing virtue ethical ideals is easier said than done. It is unclear whether or how soon virtue ethics can provide actionable moral guidance with respect to the development and implementation of AI technologies. Virtue ethics requires rigorous ethical training for agents and ethical exemplars who can provide illustrations of what it looks like to live well with AI. It also requires moral exemplars who have expertise in the technologies themselves, sufficient to help them correctly identify the ethically salient features of those technologies. There are few, if any, people who undergo such training and it is unclear whether there are any relevant ethical exemplars for us to follow. ² In short, we argue that the Aristotelian virtue ethics approach would provide the best basis for living well with AI, despite being difficult to implement.

We begin this article by considering virtue ethics, briefly. We show how it provides a better rubric for living well with AI than principlist or consequentialist approaches. Our view is that, despite the importance of a virtue ethics approach to living well with AI, doing this is incredibly difficult. In the final section of the article, each author provides an answer to the question of whether and how such a system might be implemented. Nicholas Smith argues that there is, in theory, little hope for implementing this virtue ethics approach. Darby Vickers argues, based on her experience in teaching virtue ethics, that there is some hope for educating students to live well with AI, even if it is extremely challenging.

Aristotle’s virtue ethics, briefly

In this section, we offer an account of Aristotle’s virtue ethics. ³ Aristotle argues that humans all ultimately aim at eudaimoina, ⁴ the highest human good; ‘eudaimonia’ is the exercise of the unique function or activity of humans (NE 1.7, 1098a12–20, esp. 16–17). ⁵ He identifies this unique human function as rational activity in accord with virtues of character (NE 1.7, 1098a12–20, esp. 16–17). ⁶ Virtues of character range from virtues still widely acknowledged, like ‘courage’, ‘generosity’, and ‘justice’ to the more obscure ‘megalopsychia’ (commonly rendered in English as ‘magnanimity’) and ‘megaloprepia’ (commonly rendered in English as ‘magnificence’). A virtue, in the Aristotelian sense, is a deeply rooted character trait which is stable across a variety of contexts and over the course of one’s life.

We acquire virtues by a process Aristotle calls ‘habituation’. This primarily involves the imitation of others more morally advanced than oneself, responding to novel moral situations, being corrected by one’s betters, and eventually coming to ‘see’ what a certain situation calls for (NE 2.1–2.4 1103a14–1105b18). For example, the brave person is one who, when confronted with a grave danger, sees whether, how, and why the circumstance calls for the danger to be confronted or avoided. The generous person is one who sees whether, how, and why money or other resources ought to be given. Mutatis mutandis for the other virtues. This is often condensed by saying, for example, that the generous person gives money ‘to the right person, in the right amount, at the right time, for the right reason, and in the right way’ (NE 2.9 1109a28–29). Overall, what habituation trains someone to do is to ‘see’ the salient ethical features of a situation and react accordingly. This involves not only a trained way of understanding typical situations but also a way of seeing that allows one to do the same thing for novel situations.

Habituation is a process by which a person becomes virtuous through a combination of instruction, practice, and feedback. For example, a child becomes generous by learning to share resources with others appropriately. At first, children are told or forced to share in appropriate circumstances, such as sharing a toy with a sibling after playing with it for an extended period of time. Then, when children share spontaneously, adults around them praise them for sharing. Once children learn to associate sharing with praise, they begin to share of their own volition, motivated by the pleasure that the praise induces and begin to share more frequently. As they do so, they receive feedback from parents, mentors, and peers about the circumstances in which it is appropriate to share. For example, sharing a pencil with a friend who lacks one is good, while sharing a prescription medicine with someone is inadvisable. As children mature, they realize that sharing – under appropriate conditions – is important in and of itself because they learn the importance of having resources for doing what one wants and undertaking good actions. This realization motivates them not only to share more but also to think carefully about the most appropriate ways to share. Once this realization occurs, agents share more spontaneously. As their character transforms, sharing becomes easier and more automatic. The final stage in the process is when these agents get pleasure from simply the act of being generous and their emotions and desires are calibrated such that a combination of their good instincts and judgment lead them to act in accordance with virtue.

Of course, habituation is not so linear as the above narrative makes it seem. For example, children spontaneously share with others, even before they are told to share and before praise reinforces sharing. ⁷ In addition, some children may be more inclined to share than others, either because of natural temperament or growing up in an environment that is either resource-rich or heavily interdependent. Mutatis mutandis for the other virtues. The purpose of the simplified illustration is to demonstrate that, in order to learn to be virtuous, one must undergo scaffolded, guided practice of virtuous actions.

During habituation, the actions an agent undertakes serve to shape the character of that agent. However, the agent undergoing habituation does not do virtuous actions in the same way as someone who is already virtuous. While anyone can undertake a virtuous action, only the virtuous person does so while fulfilling three additional conditions. The conditions are (NE 1105a26–32): (1) knowing the action is virtuous, (2) understanding why the action is virtuous and being motivated to do it for that reason, and (3) doing it from a firm and unchanging disposition. In other words, if we have virtues, we cannot lose them easily and, in general, we will be motivated appropriately and act appropriately. ⁸ Aristotle describes character development somewhat like athletic training. The more one undertakes virtuous actions, not mindlessly, but with deliberate practice (doing them knowingly and for the right reason), the more one conditions oneself to act virtuously in each circumstance. One’s character cannot be shaped by a single sort of virtuous act alone, just as one will not become strong through bicep curls alone. Rather, virtue requires that one calibrate the desires and emotional responses that guide one’s actions to the situation in which one finds oneself (NE 2.6: 1106a15–1107a26, esp. 1106b17–29).

What separates genuine virtues from more generic positive character traits is the presence of phronesis or ‘practical wisdom’. This is a kind of good judgment about what is genuinely good and how to go about attaining it, based on a combination of understanding the circumstances and responding with the correct virtue based on those circumstances (NE 6.5: 1140b5–8, 20–21; Vallor, 2016: 18–19). Aristotle says that ‘for virtue makes the goal correct and phronesis makes the means to that goal correct’ (NE 6.12 1144a8–9). The idea here is that Aristotle has already established the general end toward which all action is directed, that is, eudaimonia constituted by virtuous activity. Eudaimonia as a goal is almost always too vague to allow for successful practical reasoning about what one should do in any given situation. The virtues of character function to provide a finer-grained specification of one’s end as relevant to some particular situation in which one finds oneself.

For example, one part of a happy life is providing resources for others to enable them to act well and achieve in ways they might not be able to do otherwise. This is part of living a happy life because a happy human life involves living in a thriving community. So, generosity points individuals toward giving the appropriate amount, at the appropriate time, to the appropriate people. This is important so as to neither hoard resources for oneself nor to overshare and deplete one’s ability to continue to act well. All virtues require phronesis, because phronesis points to the virtue that needs to be deployed in a particular context. Thus, no one truly can exercise any virtue until they have also acquired phronesis.

A corollary to this is a view called the unity of the virtues, according to which one requires all the virtues (including phronesis) to be virtuous. Phronesis is only achievable if the person can use it to call into action the appropriate virtue for the circumstance in which one finds oneself. In other words, there is not a separate phronesis for each virtue (e.g. there is no phronesis of bravery separate from a phronesis of generosity), one has every virtue if they have any virtues and if one is lacking any virtue one has no virtues. The virtues rise and fall together. The virtues are not only similar in kind (they are all states of character that generate a type of correct action), but that they are not individual states of character but aspects of good character. The virtues are unified because virtues are not individual states of character. Each virtue is deployed in a particular situation, but good character overall involves having one’s emotions and desires tuned to the world in such a way that one is inclined away from the more distant vice in all cases (NE 2.6: 1106a15–1107a26, esp. 1106b17–29).

The virtues are not separable. While someone may, through practice and good instruction, gain proper emotional calibration in a single area of their life such that their emotions predispose them to act well, this is extraordinarily challenging to pull off. Imagine an emergency room (ER) doctor who is incredibly good at her job and taps into her well-calibrated emotions in the ER to be able to couple her instincts with her medical training to great effect. However, her instincts are only well-calibrated to the ER; when she comes home, she is easily irritated by her family, distracted, and immoderate. Clearly, this doctor is not holistically thriving; rather she is partitioning her life in such a way that she acts well in one sphere, but not in another. She is not acting well from a firm and unchanging character; we might say she is a good doctor, but she is not a virtuous person. Moreover, it is possible to imagine a scenario in which her partitioning of the various spheres of her life breaks down and her emotions prevent her from making a good decision in the ER.

The unity thesis is often abandoned in contemporary formulations of virtue ethics for several reasons. Some hold that the unity thesis makes virtue too hard to achieve. Others find it intuitive that people can have some virtues but not others (e.g. a soldier could be courageous without being generous). Foot (1983), Walker (1993), Hursthouse (1999), Badhwar (1996), and Watson (1984), for example, all reject essential pieces of the unity thesis. Walker and Foot hold that certain virtues may be incompatible, while Badhwar (1996: 308), for example, argues that virtues need only be unified ‘in a particular domain’ in an agent’s life and that possessing a virtue in one domain does not imply the existence of that virtue in another domain. ⁹ Russell (2009) and Annas (2011) defend weakened versions of it, according to which completely unified virtue is an ideal to which we aspire, or that the virtues are at least not completely separate.

While the unity thesis is often an unpopular or controversial thesis, we think it is essential for virtue ethics. It is true that the unity thesis suggests that it is extremely difficult to be a good person on the whole, that is, to be the kind of person about whom others correctly say ‘X is good’ full stop. However, being virtuous is difficult. It also strikes us as reflecting, correctly, that what any situation calls for will be determined by a complex mixture of all of one’s virtues (see, for example, Watson, 1984). Someone who has achieved complete virtue and moral situational sensitivity is rightly called a ‘phronimos’ or a ‘sage’. This kind of person is a model of, and a metric for, good action (NE 1113a31–33; Hursthouse, 1999: chapter 1; Vallor, 2016: 37).

A common commitment of both ancient and contemporary Aristotelain virtue ethics is that morality cannot be codified in a finite list of rules (Annas, 2004: 63–65; Hursthouse, 1999: 32–34; Vallor, 2016: 24–25). While there may be small handful of inviolable moral rules (e.g. do not murder, do not commit adultery), these are few and far between, and hardly offer comprehensive guidance. ¹⁰ Morality as an area of study does not admit of the specificity of mathematics or the sciences, and we should not expect to find a finite list of rules that constitute a complete decision procedure for ethics (NE 1.3 1094b11–1095a2). First, there are simply too many variables that must be accounted for when making any moral decision. You might think that one can get around this by proposing an indefinitely long list of rules that takes account of every possible situational variance. This is not a decision procedure that any actual human could construct or follow and therefore fails to provide comprehensive moral guidance. Ethics is, if nothing else, something that humans do; attempting to fill a purported gap in a moral theory by means of something impossible be done is a nonstarter. ¹¹

Furthermore, the guidance provided by the virtues and phronesis are strongly agent and situation relative. The virtues demand different things from people with different backgrounds, means, knowledge, and skills. If two people are presented with a circumstance calling for generosity and both have the means to give but one is a multimillionaire while the other has a minimum wage job, generosity might demand a higher degree of giving from the former than the latter. If a situation calls for courage, situations may differ as well. Imagine someone is trapped in a burning building and the fire department cannot arrive quickly. If one person on the scene knows to breach a door and has an axe and the other does not, inaction would speak badly of the former’s character but not the latter’s character.

Why virtue ethics

Virtue ethics has advantages over other normative systems regarding dealing with novel situations. An ethical system that can flexibly deal with novel situations is necessary in an age with rapidly changing technologies. The benefit of a virtue ethics approach is that there is at least some basis by which to judge what it would look like to live well with AI. Aristotle’s account of eudaimonia attempts to describe what it means for a human to live well in general – to live a life of rational activity in accordance with virtue. What it means to live well with AI is to consider how AI technologies might detract from or enhance our ability to engage in such activities and determine how to proceed accordingly. In other words, the virtues that we would need to live with AI are already well-known; the question is only how to apply them appropriately to current situations. ¹² Moreover, the way in which the proper use of such technologies would be determined depends on how the sage would use such technologies given the salient features of a situation. Virtue ethics, therefore, can guide designers and engineers of such technologies and inform all members of a society who interact with these technologies how to appropriately interact with them.

The future, and much of the present, is opaque. Determining what is at stake in a moral decision is often difficult or impossible. We are unable to know the minds and wills of other agents. Our understanding of the future is murky and is certainly not clear enough to make specific plans to guarantee specific moral outcomes. We cannot know what new technologies will be developed or adopted into wide use; we cannot predict the consequences of technology we cannot envision.

We, like Vallor (2016), hold that this opacity is particularly concerning when it comes to technological change. Vallor describes our current situation as one of ‘acute technosocial opacity’ (p. 6). For large chunks of our history, humans have had the luxury of assuming, reasonably confidently and reasonably correctly, that technology would continue to work much as it had in the past, or at least would change in broadly predictable ways. In the former case, the future would be sufficiently like the present such that whatever moral norms and rules we had would suffice to ensure that we could continue to live well in the face of technological change. Or, in the latter case, we would at least be able to foresee the impacts of certain technological changes and have sufficient time to prepare ourselves for new moral and social situations that updated technology would bring. However, current technological change is fast and unpredictable: ¹³ familiar technology changes, new and unexpected tools are introduced, and our moral and social life can quickly become something unfamiliar. In the face of this opacity, we cannot expect a moral system with fixed rules, especially ones that provide fairly detailed guidance or are tailored to present circumstances, to guide us in living well with changing technologies. Extant rules may have to be twisted and stretched to the breaking point to enable even a semblance of good moral decision making if we insist on a rule-based approach of some sort (Vallor, 2016: 6–7). Faced with this circumstance, we ought to consider whether virtue ethics, which never purported to reduce ethics to a finite list of rules to follow but nonetheless offers insight into living well, would enable us to cope with an opaque future.

Problems with rule-based approaches

Alternatives to virtue ethics tend to be either rule-based or based on calculi of ends. We confine ourselves in this article into addressing principlist approaches. Principlist approaches, which create governing rules or guidelines, are the dominant approaches to technology ethics. Rules are an ineffective way to guide action in novel or unpredictable situations, because any such formulation would be unlikely to address questions that arise from the specifics of the situation. Unexpected, or unexpectedly complex, problems can only be dealt with post hoc.

A good example of this comes from the case of gene-editing (and especially human gene-editing) in biomedical ethics. Despite the fact that there are general rules in the biomedical field (e.g. ‘do no harm’), specific rules must be created for each new innovation. Indeed, when Nobel Prize winner Jennifer Doudna discovered the way to undertake targeted gene-editing with CRISPR Cas-9, she realized that she needed to convene a new set of councils to govern the use of the new biotechnology. In her book, A Crack in Creation, she details a number of cases where she met with scientists and business people who wanted to use the new biotechnology for questionable purposes (Doudna and Sternberg, 2017: 184–188, 213–216). This convinced her that she must convene a summit to create a set of specific rules to govern the experimentation and innovation using CRISPR Cas-9.

The CRISPR meeting convened in January 2015 – along with subsequent summits – shows the drawbacks of the rule-based approach in the biomedical field as well as in governing technological advancements in AI. The original summits on which Doudna based her ideas – Asilomar I (1973) and II (1975) – were convened to determine rules on gene-editing during the initial recombinant DNA revolution. While CRISPR Cas-9 is much more powerful and accurate than recombinant DNA gene-editing that was discussed at the Asilomar summits, the technologies are similar insofar as they both alter genes. The rules that came out of the Asilomar conferences were highly specific to recombinant DNA and are an insufficient basis for regulating gene-editing more generally. Given that leaps forward in AI technologies differ much more widely than the two approaches to gene-editing mentioned above, it seems unlikely that any single rule-based approach could cover all cases, while providing sufficient specificity.

The principlist approach has been used in the AI ethics literature, often by explicitly citing inspiration from, or directly importing, principles laid out by Beauchamp and Childress in their seminal text Principles of Biomedical Ethics. A good example of this comes in Floridi and Cowls (2021) which argues that AI ethics can use Beauchamp and Childress’s four principles of biomedical ethics (Beneficence, Non-Maleficence, Respect for Autonomy, Justice) ¹⁴ and that AI ethics should add a fifth principle regarding transparency. ¹⁵

According to this view, the hard moral work, though maybe not the hard empirical or policy work, has either been done or isn’t particularly mysterious. We know what to do, even if we are still working out exactly how to do it or struggling to codify these otherwise clear principles and rules into official institutional rules or laws.

From an Aristotelian perspective, this is, at best, severely misguided. If these rules or principles are clear, concrete, and actionable, they are likely insensitive to the variety of novel situations that will inevitably arise. This is a central concern for anyone interested in the ethics of AI and machine learning. After all, living well with AI requires guidance on moral matters related to technology that changes rapidly and unpredictably, and often does things that are beyond our understanding. We should expect to encounter unexpected consequences and outcomes, and processes that operate in unfamiliar ways. Even if we can codify rules and pass them on or teach them effectively, we should worry about encountering situations in the future to which the rules and principles fail to apply. In short, principles frequently fail to provide concrete guidance, especially in novel circumstances.

Some principlist systems invoke virtues in addition to a set of guiding principles. In general, these principlist systems foreground (usually exceptionless or absolute) rules and use virtues to fill in gaps, adjudicate in cases of conflict, and guide application. ¹⁶ The rules and principles often appealed to in AI ethics make use of contested moral concepts; thus, applying these rules amounts to exercising good judgment. ¹⁷ This seems very much like saying that appropriately applying these rules is a matter of phronesis, in which case it is not good rules doing the work of helping us live well with AI, but practical wisdom. One might think that appealing to these rules and principles, as long as they are the correct ones or sufficiently detailed, amounts to backing into virtue ethics without trying. However, we are in danger of specifying our principles badly without phronesis as a guide. ¹⁸ Aristotle’s virtue ethics takes the opposite approach to principlist approaches that embrace virtues. For Aristotle, virtues, and specifically phronesis, serve as the ultimate guide to action. Rules, when they exist, are secondary; Aristotle mostly invokes rules when discussing the creation of laws for a state. ¹⁹ Laws, however, are established to promote virtue and they are not exceptionless, but must be adjusted based on circumstance. ²⁰ For this reason, virtue ethics by its nature provides a system that is more flexible and responsive to novel circumstances.

In addition, given the cultural and social norms in AI research, principlist systems may be a particularly poor fit for AI ethics. In order for principlist systems to constrain behavior, interested parties must agree to rules, have a culture that encourages obeying rules, and have some mechanism for enforcing those rules. Unless they do, rules are merely suggestions. A particular issue can be seen with the attempt to put in place some sort of ethical rules for the development of AI. In an open letter, a group of AI influencers and developers called for a 6-month moratorium on developing AI tools as or more powerful than ChatGPT-4. ²¹ The attempt at the moratorium was in parallel the moratorium enacted prior to Asilomar I in the 1970s; the Asilomar conferences put in place a set of rules that allowed research to proceed (Berg, 2008).

The biomedical field already relies on strong sets of rules that govern every action and experiment that is done. This culture of rule-following is lacking in computer technology fields. The call for a moratorium, so effective in DNA research, failed in AI research. This may be, in part, because many of the interested parties in AI position themselves as rule-breakers, innovators, and industry disruptors. Culturally, this sector of the technology industry has profited from a ‘move fast and break things’ mentality that is antithetical to – or unlikely to encourage – rule-following (Benjamin, 2019: 11–17). Thus, principlist approaches face two barriers: changing the culture of the technology community and creating and implementing a set of rules. Aristotelian virtue ethics has cultural change baked into it. It requires educating the next generation, especially those who will have an impact on the future of technology – those in design, governance, and so on – to shape the character of those who are making ethical decisions. This approach will likely provide a stronger foundation for envisioning a way to govern and live well with AI than a set of rules. ²²

A case study of problems with current rule-based approaches to AI ethics

Analyzing every extant AI ethics framework would be a prohibitively large task. Floridi (2019) estimated that, circa 2019, there were over 70 such frameworks being advanced by various groups, each with their own areas of interest, aims, specializations, agendas, and ways of specifying their principles. ²³ Here, we explore one paradigmatic example, a version of a principle of beneficence. The IEEE’s (2019) Ethically Aligned Design framework states that, taken collectively, the principles they have developed ‘Embody the highest ideals of human beneficence within human rights’ (p. 17). Somewhat more specifically, one of their seven principles states that ‘A/IS creators shall adopt increased human well-being as a primary success criterion for development’ (p. 21) with the recommendation that ‘A/IS should prioritize human well-being as an outcome in all system designs, using the best available and widely accepted well-being metrics as their reference point’ (p. 21).

The IEEE version of the principle of beneficence needs more detail in order to provide concrete guidance to those who develop or work with novel technologies. To fill in those details, the creators of the principles included specific guidelines for how their principles should be used. Below, we consider some paradigmatic examples.

The IEEE states that their principle of well-being should be interpreted with reference to the best available and widely accepted metrics of well-being. They hold that ‘There is now sufficient consensus among scientists that well-being can be reliably measured’ (p. 71) as well as the following positions:

• We encourage A/IS creators to consider the wide range of available indicators and select those most relevant and revealing for particular stages of the A/IS technology’s life cycle and the particular context for the technology’s use and evaluation. That is, measures of well-being that may be well-suited to wealthy, industrialized nations may be less applicable in low- and middle-income countries, and vice versa (p. 71).

• For A/IS to promote human well-being, the well-being metrics should be chosen in collaboration with the populations most affected by those systems – the A/IS stakeholders – including both the intended end users or beneficiaries and those groups whose lives might be unintentionally transformed by them. This selection process should be iterative and through a learning and continually improving process. In addition, ‘metrics of well-being’ should be treated as vehicles for learning and potential midcourse corrections. The effects of A/IS on human well-being should be monitored continuously throughout their life cycles, by A/IS creators and stakeholders, and both A/IS creators and stakeholders should be prepared to significantly modify, or even roll back, technology that is shown to reduce well-being, as defined by affected populations.

• Therefore, even though a product modification may increase well-being according to one indicator or set of A/IS stakeholders, it does not mean that this modification should automatically be adopted (pp. 79–80).

It is not clear that there is sufficient consensus among scientists that well-being can be reliably measured, and there is certainly no consensus among philosophers as to what well-being is, much less how to measure it. ²⁴ Furthermore, this is not an exhaustive list of ways that IEEE has elaborated their principles, merely a selection of some ways they have chosen to do so regarding beneficence.

Just as it would be a prohibitively large task to analyze every extant ethical framework for AI research and development, comprehensively analyzing every philosophical implication of the IEEE’s position on well-being alone would be a prohibitively large task, given that their Ethically Aligned Design is unwieldy and near 300 pages, excluding references and links to supporting studies and documentation. While their thoroughness is to be commended, we trust that our reader will be content if we can demonstrate our point with just a few examples.

The IEEE also makes policy and best practice recommendations, but these are vague and based on the assumption that those following these recommendations have a thorough understanding of what well-being consists in. For example, two principles for best practice are:

• A/IS creators should prioritize learning about well-being concepts, scientific learnings, research findings, and well-being metrics as potential determinants for how they create, deploy, market, and monitor their technologies, and ensuring their stakeholders learn the same (p. 73).

• A/IS creators should adjust their existing development, marketing, and assessment cycles to incorporate well-being concerns throughout their processes. This includes identification of an A/IS lead ombudsperson or officer; identification of stakeholders and end users; determination of possible uses, harm and risk assessment; robust stakeholder engagement; selection of well-being indicators; development of a well-being indicator measurement plan; and ongoing improvement of A/IS products and services throughout the lifecycle (p. 80).

Given our Aristotelian starting point, our view on these principles should be clear: even with the additional specifications and clarifications, principles cannot, on their own, ensure that we live well with AI. Even if these principles are useful for thinking through possible courses of action – for example, designating ombudspeople and determination of possible uses, harm and risk assessment – these will only promote well-being if the actions of those ombudspeople are virtuous ²⁵ and good determinations are made about how to habituate users to use technologies well, based on those assessments. Mutatis mutandis for the majority of these principles about best practices.

Moreover, such principles fail to innovate much beyond traditional best practice principles that govern engineering more generally. The main innovation is putting ‘A/IS’ in front of most of these ideas. If that is indeed the case, one might wonder why such a designation is necessary. Does it indicate that those innovating in AI fail to understand the principles apply to them or does it indicate that those innovating in AI fail to comply with standard engineering practices? In either case, it seems that such an elaborate set of principles are unlikely to solve these problems. Indeed, improving the character of those in the field is likely to have a far greater impact on our ability to live well with technology than a 300 page list of rules.

Exceptions to rules. ²⁶

In addition, there is a general issue with rules considered from an Aristotelian perspective. In NE 5.10, Aristotle explains that there are inevitably exceptions to rules that govern behavior. ²⁷ This arises from two circumstances: (1) legislators being unable to imagine future circumstances in which applying such a rule would render an injustice (NE 5.10 1137b13–27) and (2) cases where exceptions arise not from any deficiency in the rule, but from the nature of the object (NE 5.10 1137b27–3). Three types of exceptions might be one-off cases, cases which require a deviation to compensate for a historical trend of unfairness, or cases which are completely novel and unanticipated. These are precisely the sort of cases that manifest as technology changes. For example, algorithms tend to learn from and amplify historical bias in the data upon which they are trained. While we might make a principle to test algorithms sufficiently to root out algorithmic bias before algorithms are employed to generate decisions in the real world, there is unlikely to be a specific rule about how to handle cases in which algorithmic bias affected individuals before the bias was discovered. Moreover, each of these cases is likely to be unique to the particular way in which the decision affected the individual.

There are no exceptions to virtue ethics, because virtues merely lead one to act in the best possible way given the situation at hand. When this sort of action departs from standard rules, Aristotle says that the virtue responsible for this adjustment is decency. There are, of course, cases in which the manifestations of two specific virtues might clash. However, phronesis helps the sage solve the issue as to which virtue is more important to manifest in those cases. We argue that IEEE erred by being too detailed to govern new situations or too vague to be informative and applicable, not that the IEEE has made uniquely bad or misguided recommendations. Asilomar’s recommendations show similar flaws. Indeed, everyone’s principles have these flaws because these flaws are inevitable in attempting to govern the development and use of this unpredictable technology in opaque circumstances.

A case study of Kantian machine ethics. ²⁸

The principlist approaches that we have discussed thus far are approaches that consider how humans should make rules to live well with AIs. There are alternative approaches, particularly in machine ethics, that concern how to regulate AI behavior and interaction with humans. Machine ethics is the branch of AI ethics that concerns how to program AIs to make ethical decisions. Some of the principlist work in machine ethics uses a Kantian framework, suggesting, for example, that the categorical imperative can serve as a basis for moral decision-making frameworks in AIs (e.g. Ulgen, 2017; White, 2022). Ulgen (2017) considers how a Kantian framework might be applied both to weak AIs controlled or monitored by humans (such as autonomous weapons) and to artificial general intelligences; White (2022) focuses on designing Kantian artificial moral agents.

Kantian principlist approaches to machine ethics suffer from the same flaws described above, insofar as they veer between being too detailed to govern new situations and too vague to be informative and applicable. The categorical imperative prescribes that one act on a maxim that one could will to become a universal law; in other words, universalizing that maxim would not undermine the action taken by the agent. For example, one should not make a lying promise because if everyone made lying promises, it would undermine the purpose and value of promising. While the categorical imperative provides concrete guidance about some actions, it fails to do so about others. Indeed, Kant’s categorical imperative, laid out in the Groundwork for the Metaphysics of Morals, is merely a beginning to Kant’s moral theory. In addition to the categorical imperative, Kant lays out a complicated, multi-level procedure for how to interpret and apply the categorical imperative to our daily lives, including a set of virtues with which we should act. Even with this procedure, Kant (1991) acknowledges that judgment is often required and that there will always be exceptions that cannot find their resolution in the rules laid out (pp. 6, 411). ²⁹ Given that no machine has the ability, currently, to exercise judgment, such an approach is likely both misguided and premature. Even in the human–machine rule-generating approach Ulgen (2017) advocates, AIs lack the judgment necessary to correctly apply the generated rules.

Practicality: Can virtue ethics provide a basis for how to live well with AI?

It is prima facie important to discuss how one should use or live with AI in a way that actually provides guidance in novel situations. Given our Aristotelian perspective, our answer to the questions of the sort ‘what should we do or how should we live with AI?’ is ‘whatever the sage would do with AI’. While this is the answer that a virtue theorist in some sense should give, this is understandably unsatisfying to those who want an applied ethics of AI to provide concrete or specific guidance. Once there are moral sages with sufficient understanding of AI technologies, concrete and specific guidance comes from determining what the sage would do in a given situation. However, the implementation is complicated by the lack of concrete rules or guidance provided by a virtue ethics framework and by the lack of easily identifiable moral exemplars. Thus far, we have considered the theoretical justifications for an Aristotelian approach to virtue ethics; it is worth considering whether what we propose is practical.

We consider two different perspectives from which one can evaluate the practicality of a virtue ethics approach, each one written by one of the two authors of this article. The combination of the two endings is particularly important when considering how higher education can respond and adapt to the challenges to educational institutions posed by powerful AI tools. Smith’s ending deals with the more theoretical reasons why it will be very difficult to implement Aristotelian ethics in higher education settings. Institutional change is difficult and massive change would be needed in order to make a real impact. Smith’s ending is a sobering reminder that overhauling undergraduate education is not easy and that interventions at the undergraduate level may be too late to shape the character of students.

Vickers’ ending deals with what she sees in her ethics classroom as she has been teaching this material. Vickers’ ending is a ray of hope that even some limited forms of training might be helpful, even if it would be better to overhaul education to help students deal with a world of changing technology. Vickers has noticed, both among faculty and among students, that there is a desire to figure out what it would take to live well with AI and that faculty and students are willing to take certain radical steps to figure out what that would look like. ³⁰ Both authors hold that an Aristotelian virtue framework must be a part of this conversation as it is emerging and that we should take it seriously even if it is challenging to implement.

Nicholas Smith’s ending

We can answer our moral questions about the use of AI with old tools; we need no new principles or methods. However, an Aristotelian approach to AI is easier theorized than implemented. Finding virtuous people is a serious challenge, yet exemplars are vital for moral learning. Compounding this problem is that, among a vanishingly small population of virtuous people, we must ensure that at least some of them have the relevant technical knowledge. Under an Aristotelian framework, people would need to develop phronesis in order to make correct judgments about how AI research should proceed and to create and implement any principles that might aid in this process. Since we endorse the unity of the virtues, the acquisition of phronesis implies the acquisition of all virtues. In addition, we must ensure that those who have it are involved in developing, deploying, and making policy for AI.

A virtuous person who knows nothing at all about AI would, when presented with complex questions about the appropriate use of often inscrutable technology, be sufficiently self-aware to realize that they were the wrong person to ask. However, we cannot focus on exclusively training people to acquire technical knowledge, because even great technical knowledge of AI provides no guidance about how one ought to use it and toward what ends. A solution to this problem would require training virtue along with training the technical aspects of AI. Virtue ethics is highly sensitive to the capacities and knowledge of the agents involved in a given moral situation, so even completely virtuous agents might reasonably say, when presented with a moral question involving a complex new technology ‘Why are you asking me what to do? All I can say is that I don’t know enough. You need someone who knows how that works’. Developing virtue is hard work, as is becoming an expert in a new technology. Finding someone who has already done both is unlikely, and the educational demands of creating someone who meets that standard are immense.

If we want a virtuous person with a high level of technical knowledge, we’ll need to both habituate and educate them. Each of these are time-consuming and resource-intensive processes. In order to start a program to nurture virtuous people with sufficient technical knowledge, at a minimum, we would need at least one person who possesses both virtue and the relevant technical knowledge. This person would serve as an exemplar and provide guidance about using novel technology in novel situations, someone who can help habituate and guide other people, and so on. If a program along these lines can be developed, we can likely perpetuate it. This program would be easier to maintain than to start because once started, there should be increasingly good role models for new generations to imitate.

We might pursue both technical education and character development simultaneously, by trying to ensure that high-level technology and science educators are, if not fully virtuous, at least of sufficiently well-developed character that they can point their students in the right direction and be worth imitating. ³¹ We probably are not in a good position to identify morally upright technology educators than we are to successfully screen a pool of incoming technology students for those who are already properly habituated. It might be possible to develop course material that itself contributes to the development of virtue and could be used by even non-virtuous educators. ³² However, it is not clear that such one-size-fits-all course material could be sufficiently attentive to the nuances of virtue, especially in the cases in which it is administered by instructors who themselves may lack the requisite level of moral development.

This would help ensure the proper overlap in habituation and high-level scientific education. However, the combined technology and character education approach has other issues. If we rely on character development taking place during the process of higher education, the subjects of these efforts will overwhelmingly be people who have already developed their characters in a certain way. The standard Aristotelian view is that, after a certain point, character becomes fairly fixed. ³³ This should not be surprising; it is part of the definition of virtues that they are deeply rooted character traits that are neither gained nor lost easily. While the strictest forms of this view – that character just does not change after a certain point – seem false, it seems plausible that the amount of effort involved in changing the characters of those whose character is fixed is far greater than that involved in developing the character of a younger, more impressionable subject.

None of the above issues would be, on their own, decisive objections to educating people such that they are both virtuous and have the requisite technological knowledge and skills. However, when taken together, there are clearly substantial, perhaps collectively insurmountable hurdles. Still, embracing this approach is the best we can do. It is better to fail at this, having realized that we must aim at virtue if we are to live well with technology, than to accept defeat from the outset and settle for an approach to living with technology that is fundamentally insensitive to the problems with the currently dominant approaches.

Darby Vickers’ ending

Like Smith, I believe that there are serious challenges to overcome in the virtue education that we propose. However, I am more optimistic than he is that we can make a significant positive impact at the college level, contingent upon institutional changes. My optimism comes from three main sources: Aristotle’s concept of enkratia, recent research on brain plasticity in adulthood, and anecdotal personal experience teaching ethics courses. I address each of these three pieces in turn.

Aristotle (1986) understood – and communicates in the Nicomachean Ethics – that virtue is extraordinarily difficult to achieve. Indeed, if a virtuous person must have all of the virtues, they must have the virtue of megalapsychia. Only a truly great person can have the virtue of megalapsychia (sometimes translated as ‘magnanimity’) (NE 1124a1–4). The ancient Greek adjective ‘megalapsyche’ literally means ‘great souled’ and Aristotle says that the megalapsyche person is someone who is particularly suited for great deeds and expends her energy only on great causes because she is worthy and capable of achieving great things (NE 1123b15–30). The megalapsyche person is the best and most worthy (NE 1123b25–30). Only a select few can ever be megalapsyche for this reason. However, Aristotle’s schema allows that there is a way to undertake virtuous actions without being virtuous, namely enkratia.

The enkratic person has enough training in virtue to often be able to choose the virtuous actions but falls short of being virtuous because of a lack of the correct emotional calibration (NE 1145b1–3, 9–15). For Aristotle, as we mentioned above, being virtuous involves more than doing the virtuous actions, but understanding why they are virtuous, undertaking them for that reason, and doing them from a firm and unchanging character. In addition, the virtuous person feels pleasure from doing the virtuous action and that pleasure stems not from the action itself but from the fact that the action was virtuous. For example, imagine I find $100 on the road. Both the virtuous person and the enkratic person would try to find the owner or turn the money over to authorities to seek the owner. However, the virtuous person would feel no qualms about turning in that $100 and feel pleasure from doing it, even if $100 in their pocket would be nice, while the enkratic person might feel a desire to keep the money and feel pain or regret at losing the possibility of another $100 in their pocket. In addition, the enkratic person might not experience pleasure merely from doing the right thing.

Enkratia serves as both an intermediary step toward virtue and as a possible endpoint that creates respectable members of a society. For example, Aristotle argues that virtue requires the correct calibration of one’s emotions to the situation that one is in. There are some mental disorders, for example, that prevent one from being able to correctly calibrate one’s emotions to the situation, for example, chronic depression or psychothymia. Aristotle argues that individuals with these disorders can be enkratic, but not virtuous because they are unable to fully calibrate their emotions to the world (NE 7.5 1149a1–6).

Educating for enkratia remains challenging because it still requires developing the correct skills for determining what to do in various situations. The enkratic person is not a rule-follower, but is instead someone who is able to determine – at least on balance – how to act, even if her emotions are not appropriately calibrated to the situation. For this reason, no one-semester college class can develop enkratia; like with virtue enkratia needs to be something that students develop over time through the process of habituation. The rapid improvement of AI has increased the interest of departments and administrators, as well as the general populace, in ethics. I hope that we can generate enough momentum with this to provide ethics training in addition to incorporating this habituation and skill-building in ethical thinking throughout the curriculum, especially for those students who are training in technical fields.

An additional question is whether higher education is an appropriate place to start training in virtue. Smith is correct that Aristotle believes character is fixed after a certain point and that no amount of training will be able to change a fixed character (e.g. NE 3.5 1114a15–22, 1114b17–1115a3). In addition, Aristotle believes that a student without the proper upbringing can never be virtuous and that ethical training will not work on them (NE 1.3 1095a3–11, 1095b5–12). However, Aristotle envisioned that students would begin the theoretical aspects of their ethical training at about college age, because students needed to be sufficiently mature to benefit from such training (NE 1.3 1095a3–11). In my experience, this is a fertile age for ethical training; college students are quite receptive to Aristotle and are particularly interested in, for example, using his ideas on friendship to evaluate their own lives. Yet, it is true that most of the college students I teach have lacked the sort of environment that habituates students toward virtue, and Aristotle would say they are therefore the wrong candidates for becoming virtuous. I think, given recent understanding of brain plasticity, it is clear that it is challenging, but possible, to develop students for virtue – or at least enkratia – at a later time in their development than Aristotle envisioned.

Character is malleable at any age, but it becomes increasingly difficult to change as a person ages. Yet, recent research demonstrates that all sorts of traits, once thought to be fixed, including farsightedness caused by aging (presbyopia), are able to be altered through an intensive training process. ³⁴ Human brains are surprisingly plastic and retain the ability to make changes even after the preliminary bursts of neuronal growth and pruning that happen throughout childhood and periodically through adulthood, with a final major brain development in mid-to-late 20s. ³⁵ Even beyond this final period of brain growth, the brain continues to change and humans are able to harness that plasticity to learn new skills, form new habits, and innovate (Ericsson and Pool, 2016: 26–49). Some skills are hard to form after childhood (such as language learning and perfect pitch). ³⁶ Character does not appear to have a hard and fast developmental cut off. While character development may become more difficult as adults age, I hypothesize that the reason character changes less is because the average person does not want to put in sufficient effort to make changes. I hypothesize that teaching brain plasticity and growth mindset ³⁷ to students would help encourage them toward character development at a later stage, even if that change is more difficult.

One possible objection to this claim is as follows: even if character change is possible, ethics courses are notoriously bad at changing student behavior. Studies of students in ethics courses show that students do not behave better in general (May and Luth, 2013; Schwitzgebel, 2013), and any successes are modest (e.g. Antes et al., 2009 ³⁸ ; Schwitzgebel et al., 2020). Schwitzgebel concludes that ethics education is ineffective and may, indeed, make people worse (e.g. Schwitzgebel, 2009; Schwitzgebel and Rust, 2009, 2010). Indeed, Schwitzgebel thinks that ethics professors may act less well than those in other subdsiciplines of philosophy (e.g. Schwitzgebel, 2009).

Proponents of Moral Foundations Theory (MFT) provide an explanation for the lack of effectiveness of moral education. Jonathan Haidt (2013: 103–106) argues that students do not improve because we are trying to address the wrong part of them. He argues that our moral decisions are primarily emotional and we only reason about them after the fact (Haidt, 2013: 61–64). The model that moral foundations theorists use can be analogized to a rider on an elephant. The rider is the reasoning and the elephant is the emotions. Haidt explains this as the idea that the emotions lead our reasoning, at least for the most part, and much of what our reasoning can do for us is to provide ad hoc reasoning for our emotional decisions. As MFT proponents believe that emotions are the source of our moral reasoning, they think that ethics courses are worse than useless. Indeed, Haidt thinks they simply make people better at creating reasoning to justify their ethical decisions, particularly when those decisions put them at odds with other members of the community.

The account of MFT proponents fails to capture cases in which reasoning clearly influences ethical decisions. For example, lots of individuals change their views on ethical treatment of LGBTQ+ individuals not because of any emotional experience, but rather because they come to understand through reading or discussions with others (e.g. Sterelny, 2012: 155–160). This undermines the contention of the most die-hard MFT proponents, who argue that reasoning has little power to overcome emotional responses and ethical behavior. ³⁹ Aristotle’s view that acting virtuously (or viciously) requires a combination of desires, emotions, and reasoning seems closer to what actually happens when we make ethical decisions than the almost exclusively emotional reaction that is proposed by MFT views.

Moreover, while Schwitzgebel’s studies – which Haidt (2013: 103–106) and other MFT theorists cite – are interesting, these studies employ a limited idea of what it means to be ethical. Studies need to trace some concrete and measurable behavior, so they tend to use proxies for ethical behavior such as donating to charity or refraining from spending money on meat products (e.g. Schwitzgebel, 2013; Schwitzgebel et al., 2020). These are crude principlist standards that provide little, if any, indication about a person’s character. Take donating to charity as an example. Donating to charity may be an instance of generosity, but only if the person in question donates the correct amount of appropriately sourced resources to the correct cause, at the correct time. Studies which simply look at whether or not someone donated to charity (or even how much) cannot possibly capture the necessary ethical features of a situation; such a proxy is so crude as to be useless.

However, Schwitzgebel and the MFT theorists are correct that most ethics classes are indeed ineffective in training behavior. Changes in behavior cannot stem simply from learning principles or calculi or learning about various ethical theories. Rather, moral training can only happen through a combination of learning the skills to identify the salient ethical features of a situation, understanding what virtue is called upon, understanding how to act, correctly calibrating one’s emotions to the world such that they motivate one to act correctly, and receiving guidance and feedback. In other words, moral learning is a sort of apprentice learning, where students must have access to a learning environment that is seeded for them such that they can practice moral decision making (Sterelny, 2012: 165–171). Most ethics classes are not structured in this way.

My hope is that we can create an ethics curriculum that will move students toward enkratia. While this is optimistic, I see potential for two reasons. First, I have seen firsthand how students respond in the classroom to the sorts of teaching that we describe in this article. Second, part of the reason is that the rapid improvement of AI has increased the interest of departments and administrators, as well as the general populace, in ethics. I have hope that this may lead toward popular support for the sort of curricular changes that we advocate. I engage each of these reasons in turn.

I begin with my own experience. I have taught a number of introductory ethics classes using Aristotle, upper division seminars focusing entirely on Aristotle, and AI ethics classes where we consider Aristotelian ethics as well as alternative pictures about how ethical skills are acquired. The responses from students are impressive and, in some cases, transformative. While Smith is certainly right that habituation cannot happen in a single semester in a college class, I think that regular engagement with this material may provide students with a starting point for becoming enkratic. Indeed, they acquire from these courses a set of skills which they can practice in the moral decisions that they make in their day-to-day lives. In the last part of my upper division ethics classes, I have students write a reflection paper about whether and how they think that the class has impacted them. While I teach them about Schwitzgebel and Haidt’s perspectives, the majority of the students state that the classes caused them to think differently about decisions they were making and many of them claim that these classes changed their behavior.

In technology ethics courses, the most prevalent topic students brought up in reflections were internships and jobs. The courses that I teach are upper division courses that are one way for computer science and electrical engineering students to fulfill their ethics credit. University of San Diego, as a Catholic institution, requires that all students take a class designated as ‘Ethical Inquiry’. For this reason, the majority of students in these courses come from STEM disciplines and most of them have never taken an ethics class previously. One of the things that I have students do toward the end of the course is to write a reflection paper on what they gained from the course. In the reflection papers, I noticed something striking; several students were either considering or reconsidering where they wanted to work or intern based on the information from the class. A few students even told me that they regretted not taking the course earlier, because they would have liked to think more carefully before accepting post-graduation jobs where they weren’t sure how the organizations would consider ethics of AI. Some voiced their concerns about working for military contractors, weapons manufacturers, or big tech. Others were concerned more specifically about whether superiors would take them seriously if the students voiced ethical concerns. This seems like evidence not only against the MFT view of ethical change but also in favor of teaching ethics at the college level to build ethical decision-making skills.

In addition, I want to consider the possibility of public support for infusing ethics into the college curriculum. The rapid improvement and projected ubiquity of AI has increased the interest of departments and administrators, as well as the general populace, in ethics. Some of this interest is reactionary – people are afraid of misuse and they are hoping that by forcing students to take an ethics class they might end up with better behavior. There are organizations, like the engineering grand challenges, that are trying to create people with technical know-how and some sort of ability to anticipate long-term consequences. While I don’t think any of these solutions are good ones, I think that the inclination is right. Namely, we need to be training the next set of technically inclined students to be virtuous if possible, or at least to be something approaching enkratic. I have anecdotal evidence for such a position being persuasive to students and causing them to change how they think about their lives and sometimes alter their behavior (from a combination of student testimony and student evaluations).

In short, both authors propose that ethics training is the correct way to go about helping the next generation to live well with AI. We both acknowledge the difficulty in the proposed path. As universities regroup and reflect on their position in the face of ChatGPT and other AI tools, we hope that those in power take seriously both the possibilities and the challenges of an Aristotelian approach to living well with AI.