Machine Ethics Before the Revolution · Part 2

An Application of Aristotelian Virtue Ethics to Artificial Intelligence

· 14 min read
philosophyai ethics

Written for Aristotle (PHIL 4089) at Columbia University, fall 2019. Part two of Machine Ethics Before the Revolution. Published as written, with typographical and citation corrections.

Abstract

Ever improving computing power brings us closer to a reality with morally cognizant artificial intelligence each day. We have already begun to encounter the moral dilemmas associated with AI: The classic “trolley problem” lends itself to self-driving cars, the implicit biases built into our cultural corpus are adopted by computer systems as they absorb our literature and music, and the very process by which these machines learn moral laws - analyzing our personal data - may be ethically compromising itself. In these scenarios the importance of a sound ethical system for artificial intelligence is manifest.

In this paper I argue for one such system built on Aristotelian virtue ethics. The emphasis placed on moderation and consideration for the mean as well as the proposed method of becoming virtuous via habituation in Aristotle’s Nicomachean Ethics mirrors conventional machine learning techniques. I then touch upon a possible formulation of machine-machine morality, whereby artificial intelligences build their own ethical system through a form of computational ethical discourse. Finally, I weigh the practical advantage of control with a humanity friendly system - i.e., virtue ethics which always places the good of humankind first - against the exploratory value in a system “freely formulated” by machines, for machines.

Introduction

The ever expanding opportunity granted by advancements in artificial intelligence (AI) presents researchers, engineers, ethicists and regulators alike with a collection of moral questions in need of answers. These questions qualify the potential of this boundless technology by reminding us that AI can be intentionally - and sometimes unintentionally - harnessed for both good and bad causes. What will this technology mean for human employment in the future? How can we safeguard against mistakes or malicious attempts to fool these computers? What preparations must we take to maintain control of these systems? Inversely, how can we assure fair and “humane” treatment of these intelligences, and is that even our responsibility as creators of this technology?

As the capabilities of these autonomous moral agents advance there is a sense of impending eclipse; a point at which the machines will indistinguishably replicate human consciousness and thought. Yet one facet these systems have yet to progress in, an aspect crucial to further development, is ethics. The questions posed above boil down to one inquiry: what sort of moral code can we design for AI? Below I outline an application of Aristotelian virtue ethics to these intelligent systems, as well as a form of machine-machine ethics based on discourse ethics. I then provide some commentary on the apparent benefits and drawbacks of each philosophy.

Aristotelian Ethics: AI Implementation

At the core of Aristotle’s ethical system are the concepts of eudaimonia and ergon, which are generally translated as “happiness” and “function” respectively. The happiness that Aristotle speaks of is not the modern notion of happiness that is semi-synonymous with joy; rather, it is more closely associated with fulfillment or completion of our capacities - as humans or otherwise. The relation between these concepts is made clear in book I.7 of the Nicomachean Ethics:

…just as for a flute player, a sculptor, every craftsman, and in general for whatever has some function and action, the good - the doing well - seems to lie in the function,1

In order to discover what eudaimonia looks like for a given agent, one must inspect the capacities that are unique to it. As Aristotle points out, the function of a lyre player is the capacity that distinguishes her from other human beings: the fact that she plays the lyre. Aristotle claims that humans are distinct from animals and plants because we are capable of using reason, and so our function - and subsequently our happiness - must lie in the practical use of our rational capacities.

Up to this point Aristotle’s system seems to fit: AI are capable of a number of things that human beings either cannot do (or do far less well), so it is by no means a stretch to envision machines that are capable of fully replicating the human capacity of reason. Furthermore, Aristotle states that “the human good turns out to be activity of the soul in accord with virtue and, if there are more virtues than one, then in accord with the best and most complete.”2 If eudaimonia, human happiness, the good end at which all of our actions aim is merely excellent rational activity, AI are capable of reaching human excellence as well. Right?

It is not that straightforward. For Aristotle being virtuous is more complex than just performing a function well. In book II of the Nicomachean Ethics Aristotle introduces a dichotomy of virtues: the virtues of thought and the virtues of character. We will now go through each and see how they apply to artificial intelligence.

First, the virtues of character (êthikê) since these are where the study of ethics is rooted. Aristotle makes it very clear in NE II.1 that the virtues of character are habituated: “From this it is also clear that none of the virtues of character comes about in us naturally, since nothing natural can be habituated to be otherwise”.3 This habituation is not necessarily positive or negative. Aristotle expands upon this idea when he states that “it is from the same things and through the same things that each virtue both comes about and is ruined.”4 So, via habituation, humans can gradually become more or less virtuous with respect to their character - in other words, more or less ethical. Habituation, as a psychological concept, refers to the decline of an emotional or physiological response to a repeated stimulus. As children we form “good” (or bad) habits based on punishment or praise from our parents, teachers, and other adults around us. In Aristotle’s eyes a child is brought up well if he or she likes what is good and dislikes what is bad. Aristotle’s ethical framework relies heavily on our control over our desires. Look no further than book II.3 for an example:

For someone who abstains from bodily pleasures and enjoys doing just this is temperate, whereas someone who is annoyed is intemperate, and someone who endures terrible things and enjoys doing so - or at least is not pained by it - is courageous, whereas someone who is pained is cowardly.5

This seems to pose an immediate problem for our application to AI since computers are seemingly incapable of feeling pleasure or pain, let alone liking or disliking certain things. To be in any of the virtuous states we must not only perform virtuous actions, but enjoy doing so. How are we supposed to reconcile the crucial role that desires and emotions play in Aristotelian ethics with the apparent lack of any computational equivalent?

Unsupervised learning, the first sub-genre of machine learning that we’ll look at, is the computational equivalent of habituation. This approach takes a very large unclassified and unlabeled dataset, and trains an algorithm to build organized clusters of entries by looking for natural patterns. In the case of the virtue of temperance this might look like the following example. We provide a program with a dataset of dietary habits over some period of time - what a person is eating and drinking, how frequently, and how much exercise they are getting - as well as a personal wellbeing survey detailing how the various people included in the dataset are feeling emotionally throughout the same period. An unsupervised learning approach would sort the entries (people) by their similarities and differences, and likely produce two clusters: a temperate collection containing those who enjoy eating and drinking the right amount, and an intemperate collection to contrast. If we then give this model the ability to make “choices” for itself - i.e., how much it will eat and drink - it can place itself into either one of these behavioral clusters. How do we prevent it from becoming intemperate? This is where our second category of machine learning comes in: reinforcement learning.

Reinforcement learning is the answer to Aristotle’s emphasis on the importance of teaching. As noted above, habituation is vital to the formulation of our character virtues. However, intemperance and injustice are also both products of habituation gone wrong. Teaching is vital at a young age to make sure that habits are formed correctly, and reinforcement learning provides the same goal-driven approach and positive affirmation for unsupervised clustering models that our parents and guardians provide for us while we develop. A reinforcement learning model will correctly distinguish between the temperate and intemperate cases just as the unsupervised model did, but it will also associate a positive reward factor with one of the two clusters and a negative reward cluster with the other. Through this process we could teach an autonomous moral agent the character virtues.

We must now turn to the virtues of thinking. These are far more approachable for AI than the character virtues, since these rely more on logical procedure and scientific knowledge. Already AI are capable of advanced scientific inquiry, assisting researchers with data mining and inference formulation alike.6 Aristotle provides us with a list of five intellectual virtues: technical expertise (skills), practical wisdom, scientific knowledge, understanding, and theoretical wisdom. Of these the first is the least important to our ethical inquiry. AI have mastered technical expertise, also known as craft knowledge, for a variety of skills: IBM’s Deep Blue computer beat world champion Garry Kasparov at chess in 1997, robots have been working in factories for decades doing specialized tasks far more efficiently than their human counterparts, and computers play an integral role in nearly every sector of the modern global economy. Scientific knowledge is all things that are necessary, eternal, and could not be otherwise. These are things that we do not deliberate about, and are derived through demonstration from first principles.7 Since the basic laws of logic and bivalence that human beings use to demonstrate arguments are the same as those underlying computational architecture, we can assert that computers are also capable of having scientific knowledge.

The road to intellectual virtue seems to stop here for machines. The concept of understanding, and subsequently theoretical wisdom (the synthesis of scientific knowledge and understanding), seems to be slightly out of reach since this would require that computers comprehend the starting points of their demonstrations. There are no concrete arguments against the possibility of a machine-created AI, and there is plenty of research in the field of meta-learning whereby machine learning is used to optimize and design better machine learning algorithms. Perhaps it is through this “learning to learn” that machines will progress towards understanding (nous), but I will leave that for another discussion. Since the end goal of our ethical system is to provide AI with a framework for making decisions when presented with a moral dilemma we will now turn to Aristotle’s thoughts on decision making.

Practical wisdom (phronesis), or excellence in deliberation, is one of the most important Aristotelian virtues. It is how rational agents determine what means will most effectively bring about a desired end, and ultimately leads to a decision on whether to take a given action or not. Computers have no trouble being theoretically virtuous,8 as they are excellent at deducing patterns in large datasets and drawing conclusions from preexisting propositions. In order to be practically virtuous the desiderative faculty and the rational faculty must be in agreement:

What assertion and denial are in the case of thought, that, in the case of desire, is precisely what pursuit and avoidance are. […] it follows that both the reason must be true and the desire must be correct, if indeed the deliberate choice is to be an excellent one, and the very things the one asserts, the other must pursue.9

So, an artificial intelligence capable of autonomous moral agency would first need to be trained using reinforcement learning to have “correct” desires - i.e., to desire to be just, temperate, and courageous (whatever that may mean). Then, when presented with enough moral dilemmas, it will gradually learn which means lead to which ends, as this is the nature of nearly all inference machines. Proper alignment between the character virtues built during the reinforcement learning stage and the virtues of thought and deliberation that have gradually been perfected by machine learning scientists over the past two decades will result in an autonomous moral agent capable of acting within an Aristotelian ethical system.

An Alternative to the Aristotelian Approach

One thing that is emphasized throughout the Nicomachean Ethics is the importance of exceptions. Each scenario we are presented with provides a fresh test of our deliberative faculty, and the science of ethics serves merely as a general framework that we can apply when faced with a dilemma. Machines are experts at discerning the minute differences between puzzles and finding patterns that can be used to predict the outcome of novel problems - they need minimal help on that front. As discussed above it is the character virtues that mankind must help with, since computers have no natural conception of pain and pleasure and are incapable of becoming habituated well on their own.

But what if this wasn’t the case? If machines were capable of defining their own character virtues - i.e., temperance would refer to not using too much network bandwidth or too much memory - would it be possible to build a form of machine-machine ethics? Could a network of computers forced to share crucial resources (in a sense forced to survive together) discover forms of pleasure and pain more fundamental to their nature? Or will machine ethics always be limited by their human programmers? In my paper titled An Application of Kantian Ethics to Artificial Intelligence10 I discuss a system built on Jürgen Habermas’s Discourse Ethics that might be capable of independently devising a moral code for “machinekind”.

Control versus Exploration

The Aristotelian approach is a safe one. By maintaining command over the reward factor tied to various actions we can habituate AI to behave well. This system is for those who fear the technology singularity - an apocalyptic event in which technological growth spirals out of control and AI supersedes humanity as the dominant force on the planet. But it is limited. It allows us to create machines that behave within our own ethical guidelines, but does not allow for any abstract exploration of morality. The discoveries that could be made by a computational network pursuing its own ethical system under different constraints to the ones humanity has faced might have far-reaching implications. The Aristotelian approach results in the spread of implicit bias from the programmer to the AI, but a machine-machine ethical system might help us eliminate our own biases from our legal system. Either way, AI is the future of technological and ethical discourse on this planet, and deserves more attention than I am able to provide in this short paper.

Works Cited

  • Aristotle. Nicomachean Ethics. Translated by C. D. C. Reeve. Indianapolis: Hackett, 2014. Cited by Bekker pagination as NE.
  • Lobo, Daniel, and Michael Levin. “Inferring Regulatory Networks from Experimental Morphological Phenotypes: A Computational Method Reverse-Engineers Planarian Regeneration.” PLOS Computational Biology 11, no. 6 (2015).

Footnotes

  1. NE 1097b23 ↩

  2. NE 1098a15 ↩

  3. NE 1103a18 ↩

  4. NE 1103b7 ↩

  5. NE 1104b5 ↩

  6. In 2015 an AI system at Tufts University reverse-engineered the regeneration mechanism of a worm. It produced a remarkably simple model of the mechanism for researchers to interpret, signalling that AI are well on their way to contributing to the more intuitive and creative aspects of research. ↩

  7. “Hence scientific knowledge is a state affording demonstrations” (NE 1139b31) ↩

  8. “In the case of thought that is theoretical, however, and neither practical nor productive, the good state and the bad state are truth and falsity” (NE 1139a28) ↩

  9. NE 1139a20 ↩

  10. An Application of Kantian Ethics to Artificial Intelligence ↩