Walter Sinnott Armstrong on AI and Morality
13:58 Share

This is Philosophy Bites with me, David Edmonds. And me, Nigel Warburton. If you enjoy Philosophy Bites, please support us. We're unfunded and all donations will be gratefully received. For more details, go to www.philosophybites.com. Can AI, artificial intelligence, help us make practical ethical decisions?

The philosopher Walter Sinnott Armstrong thinks so, and he's been working with a data scientist and a computer scientist to try to build a system that will be of use to doctors faced with ethical dilemmas. Walter Sinnott Armstrong, welcome to Philosophy Bites. Thank you so much for having me. It's a joy.

We're going to discuss today how human morality can be introduced into AI. But I want to start with a very basic question because people seem to define AI in all sorts of different ways. What's your definition of artificial intelligence?

I think artificial intelligence should be defined very broadly. It occurs whenever a machine learns something, because learning involves intelligence. And in particular, often in AI systems, the machine is given a certain goal, and it learns new and better means to achieve that goal. That's when artificial intelligence occurs.

So it involves learning. A crucial component of AI is that the machine or the algorithm learns as it proceeds. Exactly. And also that it has a goal and it tries out different means to that goal, tests which means are working best, and then finds new and better ways to achieve those goals. So I want to program my AI with human morality. It sounds easy enough. If I'm a utilitarian, I just program it to maximize efficiency.

or minimise suffering. If I'm a Kantian, I programme Kantian ethics into the machine. What's wrong with that? Well, what's wrong with that is the Kantians won't like the utilitarians and the utilitarians won't like the Kantians. Who made you the dictator who got to tell everybody else which moral system should be built into the machine? The problem of choosing which principles to build into the machine is basic and difficult.

So if we can't do it that way, how else can we introduce ethics into AI? Well, some people say, let's scour the internet and look for all of the posts that people have put on the internet far and wide. Look at their values from what they buy. Look at the language they use in their emails and online chat rooms and so on. And then use the AI system to further

find out what a human would say. So instead of being top down, as it were, defining what principles we want, that's bottom up. We scour the internet, see how people actually behave or think morally, and build an algorithm based on those judgments and values that we find out there.

Exactly. And so if you look at something like ChatGPT, it predicts what a human would say in response to the question.

But the answer that it gives might not be right. It might not be justified. Sure, humans would say that, but that just shows the weakness of humans. I can see that because, for example, if you hoover up human judgment, you're also hoovering up all human biases, which may include being racist and sexist and all the other flaws that we have ethically.

The other problem is that you don't know why it's giving that answer. It gives that answer, but what are the reasons that are really driving it? A lot of these deep learning techniques that are used now are basically black boxes and you don't know what's going on behind the scenes. So top-down doesn't work, bottom-up doesn't seem to work. What's the alternative?

Well, both, and make them work together. We call it a hybrid. But the idea is that you need some principles, but you don't want to impose your own principles. So you ask people, what are the features that really matter to you in a moral situation?

which aspects of the agent or the victim in a particular action matter morally to you. And then we refine those features and put them together in conflicts. Let me get concrete.

Suppose that there's one kidney that's available because there was a car accident and the victim was a kidney donor, an organ donor. And there are two people who need the kidney. We want to ask which features of the patient matter to which of the two patients should get the kidney.

And people will say things like, well, how long they've been on a waiting list or how long they're going to live if they get the kidney or how many children they have at home or whether they have a criminal record, things like that. And we collect those features that matter to people and a few of our own, recheck them to make sure people agree that they matter, and then build conflicts. Well, what if patient A is 31 years old and patient B is 50 years old?

Well then some people would say we ought to favor patient A. But what if patient A has no children but patient B has two children that depend on them? Then you can help three people by helping the one with children. So you ask people enough of these conflicts and then the AI system can predict what they'll say in a brand new set that they've never seen before. When you say what they'll predict...

You mean what a consensus will predict, because you and I may differ. We may have all the same facts and we may still differ about who we think should get the kidney when there's only one kidney going. You may be in the majority. I may be in the minority. Does that mean we go with your choice?

No. In some of these medical situations, I think you should go with what the local values are. Now, the values in North Carolina, where I live, might be very different than the values in London, where you live. And so it's fine if one hospital has slightly different values than the other. It's not fine if in North Carolina they say blacks should never get the kidney. So there are limits.

But I think it's fine to have some variation that local communities don't always have exactly the same rules.

How do you spot what are the relevant limits? Why is race irrelevant, but some other considerations acceptable so that in North Carolina they can come to one conclusion, in London they can come to another? Well, I'd have to give you all the arguments for our racism is bad. And, you know, it's interesting, we talked about Kantians and utilitarians. They agree on that, that it's bad.

And when we do our surveys, we find that over 95% agree that race should not matter. And the ones that say it should matter often say, well, you should favor blacks because they've had so many other problems before that.

So we get an algorithm that produces certain results for North Carolina, but I live in North Carolina and I still don't agree with the result that the algorithm has thrown up. Where does that leave me? Well, that leaves you trying to convince other people. But that's going to be happening regardless of whether it's an AI that picks the patient or a doctor that picks the patient. These are difficult situations and there's never going to be an answer that satisfies everyone completely.

Should you weight it according to expertise? I imagine that when it comes to how kidneys are distributed, that the person on the street may have a very different view.

to the kidney surgeon, and it seems sensible to allow the kidney surgeon two votes or three votes compared to the one vote of the ordinary person on the street. Sure, that makes sense for the simple reason that the surgeon knows a lot more facts about kidney exchanges. When I went into this, I had no idea how many hours a week you spent on dialysis when you're waiting for a kidney transplant.

I thought it was maybe two or three. Turns out it can be six hours a day, six days a week. And that makes a big difference to whether you can take care of your kids and whether you can keep your job and so on. So if people are not aware of those facts and they would make different judgments if they knew those facts, then you want to favor the judgments of the people who know the facts.

And that is often going to be the experts in the hospital system. But we can also look to the general public and correct their mistakes by informing them, by seeing whether they stick with the judgments. When they're given better information, they better understand what the situation is. Then I don't see any reason to favor the expert over the person in the general population if they're both equally well-informed.

How in practice is the AI going to help in this particular case? So we feed on all the data and the AI will throw up a result which we hope everybody can then live by. Yes, we hope that everyone can live by it, but also it can confirm. So if, for example, a doctor says, I think it should go to Joe, the one kidney that's available should go to the patient Joe.

Doctors are not always completely confident. And if the AI says, yes, it should go to Joe, then that's confirmation and helps the doctor. But if the doctor says it should go to Joe and the AI says, no, it should go to Sally, now we realize maybe we need to look at this more carefully and ask some other people to input. So this system doesn't necessarily dictate the final answer, but it can be helpful in the process.

Do you think what would actually happen is that we would subcontract the decision-making to the AI? Because we see that if we're driving a car and our sat-nav tells us to go left and we would normally turn right, we quickly learn if we turn right, there's a big traffic jam and we should have followed what the AI said. I mean, in practice, aren't doctors going to look at what the AI determines and just go along with it?

So they might or might not. I don't know what doctors you know, but the doctors I know don't tend to just go along with things. I think what they're going to ask is why? Why is the AI picking patient A instead of patient B? Now if you did it bottom up and just searched the internet, you wouldn't be able to answer that question adequately. But notice that we have these features that are morally relevant features.

And the AI will be able to say, because of this feature and this feature. So it can give the doctor a reason. Now, at that point, the doctor can say, I don't think you're putting enough weight on these other reasons and can argue with it. And then we can go to a higher authority, an ethics board of some sort in the hospital. But the doctor might say, oh, you're right. I didn't think of that.

Because I think a lot of immorality or moral mistakes that occur in our society are things that people would recognize if it were pointed out to them. And so in those cases, at least, the doctor might say, this machine helped me avoid making an error.

There are far fewer kidneys around than people who need them. How far are we away from a new system where this AI can facilitate doctors making decisions about how to allocate kidneys? So AI is already used in some kidney transplant centers.

but not for moral purposes. And one reason we picked kidneys was we knew it was already being used to allocate kidneys in the most efficient way so as to serve the greater medical reasons. And now we're saying, well, if you can do that for the medical reasons, then put in the moral reasons as well.

How far are we from getting a hospital to actually use this system? Look, I don't want them to use the system right now. We've got our little system and it works pretty well, but we need a lot more refinements on it. We're talking 10 years in the future. But 10 years from now, I wouldn't be surprised.

And how widespread is the application of ethics in AI? We've talked particularly about kidneys. How far can this go? It can go a long way.

We're already looking at dementia and patient preferences in late-stage dementia and using data from the patient to predict what they would want. We're talking to people in the business school about using it for hiring in certain ways that will introduce moral considerations, fairness with regard to gender and race, for example, into hiring decisions.

We're in contact with the military at West Point in the U.S., but also the Special Operations Forces. And they're all interested in using this kind of technology, not because they think it's going to work, but because they think it's worth trying. Walter St. Armstrong, thank you very much indeed. Thank you very much for having me.

For more Philosophy Bites, go to www.philosophybites.com. You can also find details there of Philosophy Bites books and how to support us.