Guest Blog Author! Ethics and AI with Dr. Graham Culbertson (part 3 of 3)

This is a three part series on Ethics and AI. Dr. Graham Culbertson is our guide in this exploration of what ethics is, how we apply ethical thinking in different situations, and ultimately has a little fun with what ethical thinking might look like when applied to AI. Feel free to jump in here, or go back and read Part I and Part II if you prefer (recommended).

Dr. Culbertson is a professor at the University of North Carolina at Chapel Hill, and hosts the podcast Plumbing Game Studies, in which he explores game studies through the lens of philosophy and vice versa. He has been teaching at the high school and college level since 2006. 

Ethics and AI, Part III

In the final blog post of this series, I’d like to turn our attention to a question that constantly hovers around the field of AI ethics, and show how the ancient ethical philosophy of Aristotle harbors a potential answer for AI moving forward. The question is: Will AI kill us all? But I’m not talking about Skynet here, the fictional AI from the Terminator series that wanted to kill all humans because it perceived humans as the enemy. The question currently confounding AI ethicists goes something more like this: Will AI kill us all, but because it followed the instructions we gave it for how to be a good person?

The theory goes like this: an AI is programmed to have perfect utilitarian ethics, meaning that it is aiming for the highest amount of human happiness. Confident that human happiness is the ultimate goal, and that human efforts (in the form of governments, corporations, and other institutions) have failed to maximize happiness, humanity turns all power over to this perfect AI overlord. The AI surveys the data, determines that every human life has more unhappiness in it than happiness, and thus determines that the highest amount of human happiness possible is zero. Our Benthamite AI then euthanizes all humans, creating a permanent happiness state of zero, and congratulates itself on correctly following its programming. Or perhaps it encases every human in a virtual reality simulator of heaven, maximizing their happiness but simultaneously enslaving them.

One can tell the same story with Kantian ethics; Kant thought the highest goal of the universe was rationality, but he also wrote that humanity was a “crooked timber,” incapable of such rationality. A perfectly programmed Kantian overlord might decide to exterminate all humans and replace them with Stepford approximations of people, or genetically alter every human embryo to remove pain, pleasure, and other manifestations of emotions. This AI overlord, like the Benthamite one, hasn’t murdered or enslaved humanity against its programming; it has in fact done so precisely within its programming.

Current machine learning algorithms, which are the closest thing we have to AI, do behave in this way. In the 2018 research paper The Surprising Creativity of Digital Evolution, dozens of AI researchers compiled an extensive list of machine learning systems following their instructions but definitely not being good robots. In Janelle Shane’s description of one of these examples in her A'I Weirdness blog:

Another set of simulated robots were supposed to evolve into a form that could jump. But the programmer had originally defined jumping height as the height of the tallest block so - once again - the robots evolved to be very tall. The programmer tried to solve this by defining jumping height as the height of the block that was originally the *lowest*. In response, the robot developed a long skinny leg that it could kick high into the air in a sort of robot can-can.

These algorithms, like our hypothetical humanity exterminators and enslavers, are simply doing what their programmers told them to do. But to any human being, it’s obvious that falling head over heels isn’t jumping, in the same way that locking humans in the Matrix isn’t freeing them.

Since at least E.M. Forster’s short story The Machine Stops in 1909, people have been imagining machines that run society but get it wrong because of their overly simplistic, algorithmic understanding of ethics. Hal 9000 in 2001: A Space Odyssey refuses to open the pod bay doors not because he is evil, but because he is (wrongly) trying to save the mission by murdering the humans. And in the films I Robot (2004) and I Am Mother (2019), machines are put in charge of the world and take actions that we would consider monstrous, but do so in service of a Benthamite or Kantian greater good. In I Am Mother, the AI even takes some time to explain Kantian and Benthamite ethics and explain how one of them guided her choices!

One potential solution to this problem is to suggest that we just need to get better at refining the letter of the law, so that when we unleash AI on the world, it has mastered the difference between helping humans be happy and murdering them, even if murdering them fits the most simplistic understanding of Kant or Bentham. But as an alternate way of thinking about ethics and the problem of training AI to be ethical, let me suggest the ancient work of Aristotle. Aristotle’s form of ethics, often called “virtue ethics” because it was about cultivating virtue in individuals and communities, isn’t algorithmic like Kant’s and Bentham’s ethics. It thus cannot provide a clear, mathematical way to reason through an ethical dilemma. Aristotle even, unlike his teacher Plato, refuses to endorse a single value or set of values that ethics can be aiming at (aligning Aristotle with Sartre, and, again, against Kant, Bentham, and Mill). In short, Aristotelian virtue ethics is not the kind of ethics that can be taught to the sort of machine learning programs that we have right now. But if we could use it to develop more sophisticated machine learning programs in the future, the complexity, ambiguity, and nuance of virtue ethics would definitely avoid the machine murder problem we’ve been investigating.

As Andrew Fisher and Mark Dimmock put it in their introduction to Aristotle in Philosophical Thought, “For Aristotle, morality has more to do with the question ‘how should I be?’ rather than ‘what should I do?’” This means, as they further explain, that “Aristotle refers to virtues as character traits or psychological dispositions. Virtues are those particular dispositions that are appropriately related to the situation and, to link back to our function, encourage actions that are in accordance with reason.” So rather than figuring out what the perfect value is, and then maximizing that value, Aristotle wants us to figure out what it means to be a good person, and then work hard to develop those characteristics. We can easily see that a machine that would murder all of humanity in order to reduce suffering wouldn’t be considered a good “person.” And the reason we can see that is because our ethics doesn’t come from mindless number crunching, but rather from the sustained hard work of becoming better people in a community of people who are both helping and criticizing us.

Let’s go back to the computer simulated jumping robots who were learning to be very tall fallers instead of actual jumpers. It’s clear that, although they were evolving to meet a predetermined value structure, they weren’t doing a very good job of it. Aristotle anticipated something like this; we can think of iterative machine learning algorithms as doing the hard work of “regular practice,” which requires “experience and time.” But these simulated robots are practicing the wrong thing. Aristotle notes the same is true for people: “It is from building well that people become good house builders, and from building badly that they become bad ones…The same goes for the virtues.” Those simulated robots were practicing to be good jumpers, but they were practicing it badly. And they were practicing it badly because they were being given a single, and single-minded goal: to put themselves, or their feet, as high as possible.

Which brings us to what is perhaps Aristotle’s most famous concept, “the golden mean,” which is simply the idea that the right course of action is always somewhere between two extremes, although probably not directly in the middle between them. To become an ethical person is to practice finding the balance between extremes as much as possible. All of the ethical dilemmas from part one and part two of this series denied even the possibility of the golden mean; one must either lie or not lie, fight the Nazis or help one’s mother, drive the car into the crowd or off the bridge. But the vast majority of right vs. right questions don’t have such clear-cut outcomes and thus can’t be solved algorithmically. Instead, there’s some sort of muddled middle that is the best we - or a self-driving car - can do. As Aristotle puts it:

Virtue of character is an intermediate condition…between two vices - one of excess and one of deficiency - and that it is intermediate by being good at hitting the mean in both feelings and action. From this it follows that being good is hard work, since it is hard work to find the mean in a particular case (How To 69-71).

This suggests that the monstrous machines we’ve looked at from science fiction films (2001: A Space Odyssey; I, Robot; and I Am Mother) have gone wrong for precisely the reason you would expect if you put a philosophically trained person with no experience in charge of society: they haven’t worked hard to learn, in a thoroughly non-mathematical way, what it means to do the right thing. This means that the work that Kant and Bentham did was, for AI, precisely backwards. Writing during the peak of enlightenment rationality, Kant and Bentham were trying to find algorithmic answers that emerged out of human behavior. But machine learning systems don’t need to find algorithmic answers; they are algorithmic answers! What they would need, instead, is to understand the complexity of human society. Aristotle’s virtue ethics often seems quaint and old fashioned, out of step with the rigorously quantifiable state of our current world. But I’d like to suggest that it’s Aristotle’s very resistance to easy algorithmization that makes his ideas so valuable for teaching machines. Because if we want true AI, an AI that can actually think and behave like a person, it will need to have somehow gained the experience, even wisdom, that can’t be encapsulated in the formulas of Bentham and Kant.

I’ll leave you with that as our final thought. If ethics is the complex job of choosing between multiple right options, then our current machine learning systems are nowhere near being able to act ethically. Even the self-driving car as a trolley problem example is more mathematically clear than any real-life experience any self-driving car is likely to find. When we imagine AI trying to act ethically, we believe it will get it wrong due to its simplistic application of ethics. Or, alternatively, we imagine machines that are just people - they’ve gone beyond the plateau of algorithmic ethics. And I think someone who wants to think about how AI might get past that plateau should spend more time with Aristotle than with Kant or Bentham.