Virtuous by Design

Artificial agents will have to make moral decisions. How might we equip them to do so?

I’m Tim Gorichanaz, and this is Ports, a newsletter about design and ethics. You’ll find this week’s article below, followed by Ports of Call, links to things I’ve been reading and pondering this week.

Today, our built environment moves. In earlier eras, the things humans made stayed still until someone stirred them or the weather had its way. But over the past century, we’ve begun to inhabit a world of automation.

And now, from factory equipment and thermostats to digital interfaces and processes, the automation in our world is becoming more adaptable, complex and capable.

Screen from Howl’s Moving Castle, © 2004 Studio Ghibli

Humans have always dreamt of automation: ancient robots guarding the Buddha’s relics, myths of golems, the 18th-century chess-playing Mechanical Turk, the earliest chatbots of the 1960’s—and not to mention all the representations in sci-fi books and film.

At the end of 2022, ChatGPT captured the world’s imagination, reigniting dreams of automatons. We’re still a long way from household robots. But given that we spend most of our time (and make most of our livings) in the cognitive realm, there’s a lot that a disembodied artificial agent can do for—and to—us. As this new class of technology becomes integrated into more and more of the things we use, from Gmail to PowerPoint, it seems all of a sudden our world is populating with artificial agents.

Consequently and rightly, we’re hearing more and more about AI ethics—discussions of what is good and best when it comes to artificial agents. Complex automation is synonymous with decision-making, and many decisions are moral decisions. (I happen to believe that all decisions are moral decisions in the end, even if most matter very little.) Even a system like ChatGPT must decide whether and how to answer queries that involve medicine and private information, for instance—and ethical dilemmas lurk behind all these decisions. The issue, then, is how to equip artificial agents to make moral decisions.

Morality, at heart, is about humans getting along with each other in a world where we sometimes want different things. It’s almost an oxymoron, then, to speak of moral machines, but for the fact that we have no choice. This gives rise to what’s known as the alignment problem, or the challenge of creating artificial intelligence systems that are aligned with human values. How might this be possible, given that artificial agents don’t grow up and die, they don’t have relatives or lovers, and they can’t get injured or betrayed? Is it possible for something that doesn’t have firsthand stake in being a human organism could participate meaningfully in morality?

Virtue Ethics for the Digital Age

Discussions of machine morality have been guided by frameworks developed in ethics, the branch of philosophy responsible for doing such things. Broadly, there are three approaches: deontology, which focuses on rules and laws; consequentialism, which focuses on outcomes; and virtue ethics, which focuses on personal traits and skills.

Virtue ethics is the oldest of these approaches (though the name itself is rather new), and in recent years it’s seen a surge of interest as a guiding ethical theory for digital technology. If you’re interested in all the reasons and arguments, I’d suggest Shannon Vallor’s magisterial book Technology and the Virtues. Specific to artificial intelligence, Nicolas Berberich and Klaus Diepold also make the case in their paper “Virtuous Machine—Old Ethics for New Technology?”

To give a sketch of the framework, virtue ethics is about identifying the qualities that are conducive to living well, both individually and in community. To some extent, different communities may find different qualities virtuous, though there do seem to be some human universals. Such virtues include honesty, respect and kindness. Virtue ethics suggests that, rather than trying to calculate every possible outcome of your actions or running your life by an exhaustive rulebook, you should do your best to cultivate the virtues within yourself. You can do this by identifying admirable people in your community and striving to do things like they do, and by tuning in to your sense of personal flourishing and fulfillment and doing more of the things that make you feel that way. Virtue ethics is not about binary thinking (that’s good, that’s bad), but about continual striving and growth.

One of the benefits of virtue ethics is that it allows for flexibility and contextual judgment in resolving moral questions. To borrow an example from Martin Gibert’s paper “The Case for Virtuous Robots,” imagine you’re in a moral dilemma that forces you to choose between saving a child or an elderly person—letting the other die. A consequentialist approach might suggest saving the child, since they have more years of life ahead of them. A law-based approach might not be able to make a decision at all, if its rules include not harming any humans even through inaction. Virtue ethics, on the other hand, would suggest taking in contextual factors that are opaque to the other systems (local norms, outside-the-box possibilities, how much time you have to make a decision, etc.). In the end, it might mean simply making a decision, acknowledging that there was no far-and-away best answer because sometimes life is tragic.

Toward Virtuous Artificial Agents?

You might think virtue ethics sounds nice but is difficult to implement reliably—in human life, let alone a technical system. There’s no set of rules you might encode. But contemporary AI systems present the opportunity for implementation through learning, without having to explicitly code in particular virtues or laws.

To this end, there have been a few recent papers exploring the feasibility of virtue ethics for artificial agents, with promising results.

One strategy leverages the way humans learn morality: through identifying and emulating moral exemplars. Simply put, we learn what is good and right by looking to our community around us (parents, family, teachers, TV characters, etc.), noticing who we admire over time, and striving to inhabit their personas when making moral decisions. Gibert, in the paper I mentioned above, for example, discusses several considerations for implementing a system to train artificial agents on human-selected moral exemplars.

Another recent paper, “Artificial Virtuous Agents in a Multi-Agent Tragedy of the Commons” by Jakob Stenseke, offers a technical implementation and an experimental proof-of-concept for virtuous artificial agents. The implementation includes moral exemplars but also other components, such as a goal function of maximizing eudaemonia, a concept related to flourishing and lasting happiness. (The interesting thing about maximizing eudaemonia is that it involves deliberately, almost paradoxically, not maximizing any particular quality or goal. Virtue ethics is all about balance between excesses, the golden mean. So an AI system adhering to virtue ethics would not run the risk of, to use Nick Bostrom’s somewhat silly example, destroying the known universe on a quest to maximize paperclip production.) Stenseke’s experimental results show that the virtue-based system he designed is feasible and capable of moral learning and emulating moral exemplars.

There are certainly challenges ahead. How might this approach work with a chatbot, for example, which has limited access to the context and intent behind any given query? More broadly, it may be difficult for artificial agents to understand what makes an example an example—especially when there are deep nuances involved—and to extrapolate wisdom to be applied in analogous (but not identical) cases. There’s also the question of what it could mean for an artificial entity to maximize eudaemonia, when this concept is so tied to the human condition—the need for food and livelihood, and the fact that we all will die.

In the end, I wonder what we might decide against automating. From this moment in history, it looks like we're on a one-way path toward full automation (whatever that may mean). But I think full automation would take away too much of what it means to be human. We wouldn’t accept it. Being human means being inefficient.

I’m reminded of Ray Bradbury’s 1950 short story “The Veldt,” which portrays life in a fully-automated smarthome. “I feel like I don’t belong here,” the wife says to her husband. “You look as if you didn’t know what to do with yourself in this house, either. You're beginning to feel unnecessary too.” At that point, their choices are either to unplug the house or to disappear inside.

---

Ports of Call

• On Grading: Grades have only been around 150 years, even if they seem to us like a centerpiece of education. Today, many teachers are beginning to rethink this “centerpiece.” A recent article from NPR goes into some of these efforts and perspectives in higher ed.

• On AI Detectors: Along with ChatGPT have come fears about ChatGPT and strategies to thwart ChatGPT. In my world of higher ed, most of this centers on policing and preventing cheating. I’ll probably write more about this in a future post, but for now, beware that “most sites claiming to catch AI-written text fail spectacularly.” (I don’t know why they say “most”—it’s all.)

• On Taste: Is there such a thing as beauty? Is it in the eye of the beholder, or is there an objectivity to it? I think about this question a lot, which for me is part of the appeal of Christopher Alexander’s work. For now, I share this short essay on taste from polymath Paul Graham.

• A Race: The 100k I ran last week has a highlight reel up on YouTube. If you’d like to see 15 lovely minutes of trail running, check it out. And here’s a little photo of me during the race, just for fun: