Ethics and AI: a crash course

By Sabrina Martin


Walk in to any entry-level Ethics class and you’ll observe academic philosophers teaching their students three theories of what it means to be ethical:

  1. Utilitarianism: Does an action produce net good consequences?;

  2. Deontology: Does an action follow a moral rule? e.g. the Golden Rule: ‘Treat others how you want to be treated’;

  3. Virtue Ethics: Does an action contribute to virtue (whatever that means--blame Aristotle)?

Ethical practitioners, on the other hand, will be quick to tell you the world doesn’t work so neatly.

Some Philosophy

Let’s apply these theories to AI. A civil servant in the Department for Education is trying to decide whether or not to let a computer predict children who might be at risk of falling behind at school. We might then ask, ‘Does automating this process produce a net good?’ The answer seems to be, probably, yes: the more children that we can identify as being at risk, the more children we can help.

One potential problem is that there might be a few children who are falsely identified as being ‘at risk’ due to patchy data fed into the automated programme, but the overall net good produced still seems to be positive.

Let’s look at the flipside: what about children who are missed out of our ‘at risk’ lists, because of patchy data? The net benefits of the automation might still be positive, but individuals can fall through the cracks (and this is a common concern with automated processes)!

So it now seems that a theory of net good by itself won’t help us determine whether or not we should automate. So this first theory doesn’t seem to fully solve our ethical conundrums.

Now let’s ask whether or not automating the prediction of at-risk children follows the Golden Rule. This ethical framework seems to produce a resounding ‘no’ (for most of us at least). Intuitively, we would not want our kids being categorized and labelled according to impersonal data points.

We now have two conflicting answers from different ethical theories. One says, ‘yes, automate’ the other says ‘no, don’t’.

It’s no surprise to philosophers that utilitarianism and deontology often produce moral answers; some of philosophy’s greatest moral conundrums were invented to highlight this disparity. But when making a real-life policy decision, what do we do? If I’m a utilitarian and my boss is a deontologist, how do we decide whether or not it’s ethical to automate?

There are a few answers to this. First, we might say that ethics isn’t the only consideration at play, and we have to think about cost, resources, and efficiency. So the ethics gets relegated to a secondary consideration--or we pick the ethical theory that is most in-line our economic reasoning and justify our moral rationality post hoc. Secondly, however, we might want to make philosophical purists angry by devising a framework that lets us evaluate automation decisions in a more holistic way.

This brings us to the third moral theory: virtue ethics.

AI and Virtue: Strange Bedfellows

Virtue is a concept that has largely gone out of our non-religious vocabularies. In philosophy, the idea roughly signifies moral excellence and human flourishing. So it certainly seems like an odd question to ask about how a machine can contribute to that. But perhaps if a computer can identify children who are at risk of not being able to flourish to their full potential, we might be able to make the case that automation can contribute to virtue. Virtue ethics, however, puts a lot of emphasis on learning through experience and that’s something that a computer seems to take away.

Another characteristic that virtue ethics emphasises is an idea of ‘well-roundedness’: a person can only be virtuous once they are well developed all around. So, as opposed to utilitarianism and deontology, where it’s enough to make you a moral person if you do one thing morally, you can’t be moral according to virtue ethics by just doing one good thing. You have to repeatedly do things well and in good consciousness, and this cultivates a virtuous moral character.

At first glance, this doesn’t appear to apply at all to our case study of flagging at risk students.

So how could this apply to machines?

A new way of thinking about AI Ethics

For a long time, AI ethics tended to be utilitarian. Ethicists prescribed looking at the consequences that automation would bring about and we decide whether or not automation was ethical based on the outcomes.

It’s now becoming more of a fad to consider AI ethics deontologically. Asimov’s Three Laws of Robotics are a great pop culture example of how we might start to think about the role of ethical duties when it comes to AI. The problem with this is, as utilitarian philosopher Peter Singer points out, is that ‘no technology can yet replicate Asimov’s laws inside a machine’. We can’t program a computer program never to harm anyone.

Instead, I suggest that we begin to look at the process more holistically. As one of my OI colleagues (who also has a penchant for philosophy) put it, ‘it's not enough for a machine to perform its most immediate task well. It has to be designed in a virtuous spirit’.

Automation is a process. It involves inputs, the actual automation process, and outputs. Whereas utilitarianism and deontology look at decisions individually, I instead propose that we might want to take a process-based approach to AI ethics where we break the ultimate decision down into its constituent parts and evaluate them each. So, perhaps a virtue ethics of AI has to make us look at the process that surround AI rather than the consequences of the machine outputs, or the rules that govern or its behaviour. This gets us back to the human side of things -- which is what virtue ethics fundamentally recommends.