Ethicists have been asking themselves a question over the last couple of years that seems to come right out of science-fiction. Is it possible to make moral machines, or in their lingo, autonomous moral agents -AMAs? Asking the question might have seemed silly not so long ago, or so speculative as risk obtaining tenure, but as the revolution in robotics has rolled forward it has become an issue necessary to grapple with and now.
Perhaps the most surprising thing to emerge out of the effort to think through what moral machines might mean has been less what it has revealed about our machines or their potential than what it has brought to relief in terms of our own very human moral behavior and reasoning. But I am getting ahead of myself, let me begin at the beginning.
Step back for a second and think about the largely automated systems that surround you right now, not in any hypothetical future with our version of Rosie from the Jetsons, or R2D2, but in the world in which we live today. The car you drove in this morning and the computer you are reading this on were brought to you in part by manufacturing robots. You may have stumbled across this piece via the suggestion of a search algorithm programmed but not really run by human hands. The electrical system supplying your the computer or tablet on which you are reading is largely automated, as are the actions of the algorithms that are now trying to grow your 401K or pension fund. And while we might not be able to say what this automation will look like ten or 20 years down the road we can nearly guarantee that barring some catastrophe there will only be more of it, and given the lag between software and hardware development this will be the case even should Moore’s Law completely stall out on us.
The question ethicists are really asking isn’t really should we build an ethical dimension into these types of automated systems, but if we can, and what doing so would mean for human status as moral agents. Would building AMAs mean a decline in human moral dignity? Personally, I can’t really say I have a dog in this fight, but hopefully that will help me to see the views of all sides with more clarity.
Of course, the most ethically fraught area in which AMAs are being debated is warfare. There seems to be an evolutionary pull to deploy machines in combat with greater and greater levels of autonomy. The reason is simple advanced nations want to minimize risks to their own soldiers and remote controlled weapons are at risk of jamming and being cut off from their controllers. Thus, machines need to be able to act independent of human puppeteers. A good case can be made, though I will not try to make it here, that building machines that can make autonomous decisions to actually kill human beings would constitute the crossing of a threshold humanity will have prefered having remained on this side of.
Still, while it’s typically a bad idea to attach one’s ethical precepts to what is technically possible, here it may serve us well, at least for now. The best case to be made against fully autonomous weapons is that artificial intelligence is nowhere near being able to make ethical decisions regarding use of force on the battlefield, especially on mixed battlefields with civilians- which is what most battlefields are today. As argued by Guarani and Bello in “Robotic Warfare: some challenges in moving from non-civilian to civilian theaters”, current current forms of AI is not up to the task of threat ascription, can’t do mental attribution, and are currently unable to tackle the problem of isotropy, a fancy word that just means “the relevance of anything to anything”.
This is a classical problem of error through correlation, something that seems to be inherent in not just these types of weapons systems, but the kinds of specious correlation made by big-data algorithms as well. It is the difficulty of gauging a human being’s intentions from the pattern of his or her behavior. Is that child holding a weapon or a toy? Is that woman running towards me because she wants to attack or is she merely frightened and wants to act as a shield between me and her kids? This difficulty in making accurate mental attributions doesn’t need to revolve around life and death situations. Did I click on that ad because I’m thinking about moving to light beer or because I thought the woman in the ad looked cute? Too strong a belief in correlation can start to look like belief in magic- thinking that drinking the lite beer will get me a reasonable facsimile of the girl smiling at me on the TV.
Algorithms and the machines they run are still pretty bad at this sort of attribution. Bad enough, at least right now, that there’s no way they can be deployed ethically on the battlefield. Unfortunately, however, the debate about ethical machines often gets lost in the necessary, but much less comprehensive fight over “killer robots”. This despite the fact that war will be only a small part of the ethical remit of our new and much more intelligent machines will actually involve robots acting as weapons.
One might think that Asimov already figured all this out with his Three Laws of Robotics; 1) A robot may not injure a human being or, through inaction, allow a human being to come to harm; 2) A robot must obey the orders given to it by human beings, except where such orders would conflict with the First Law; 3) A robot must protect its own existence as long as such protection does not conflict with the First or Second Law. The problem is the types of “robots” we are talking about bear little resemblance to the ones dreamed up by the great science-fiction writer, even if Asimov merely intended to explore the possible tensions between the laws.
Problems that robotics ethicists face today have more in common with the “Trolley Problem” when a machine has to make a decision between competing goods rather than contradicting imperatives. A self-driving car might be faced with a situation of hitting a pedestrian or hitting a school bus. Which should it chose? If your health care services are being managed by an algorithm what is it’s criteria for allocating scarce resources?
A approach to solving these types of problems might be to consistently use one of the major moral philosophies: Kantian, Utilitarian, Virtue ethics and the like as a guide for the types of decisions machines may make. Kant’s categorical imperative of “Acting only according to that maxim whereby you can at the same time will that it should become a universal law without contradiction,” or aiming for utilitarianism’s “greatest good of the greatest number” might seem like a good approach, but as Wendell Wallach and Colin Allen point out in their book Moral Machines, such seemingly logically straightforward rules are for all practical purposes incalculable.
Perhaps if we ever develop real quantum computers capable of scanning through every possible universe we would have machines that can solve these problems, but, for the moment we do not. Oddly enough, the moral philosophy most dependent on human context, Virtue ethics, seems to have the best chance of being substantiated in machines, but the problem is that virtues are highly culture dependent. We might get machines that make vastly different and opposed ethical decisions depending on the culture in which they are embedded.
And here we find perhaps the biggest question that been raised by trying to think through the problem of moral machines. For if machines are deemed incapable of using any of the moral systems we have devised to come to ethical decisions, we might wonder if human beings really can either? This is the case made by Anthony Beavers who argues that the very attempt to make machines moral might be leading us towards a sort of ethical nihilism.
I’m not so sure. Or rather, as long as our machines, whatever their degree of autonomy, merely reflect the will of their human programmers I think we’re safe declaring them moral actors but not moral agents, and this should save us from slipping down the slope of ethical nihilism Beavers is rightly concerned about. In other words, I think Wallach and Allen are wrong in asserting that intentionality doesn’t matter when judging an entity to be a moral agent or not. Or as they say:
“Functional equivalence of behavior is all that can possibly matter for the practical issues of designing AMA. “ (Emphasis theirs p. 68)
Wallach and Allen want to repeat the same behaviorists mistake made by Alan Turing and his assertion that consciousness didn’t matter when it came to the question of intelligence, a blind alley they follow so far as to propose their own version of the infamous test- what they call a Moral Turing Test or MTT. Yet it’s hard to see how we can grant full moral agency to any entity that has no idea what it is actually deciding, indeed that would just as earnestly do the exact opposite- such as deliberately target women and children- if its program told it to do so.
Still for the immediate future, and with the frightening and sad exception of robots on the battlefield, the role of moral machines seems less likely to have to do with making ethical decisions themselves than helping human beings make them. In Moral Machines Wallach and Allen discuss MedEthEx an amazing algorithm that helps healthcare professionals with decision making. It’s an exciting time for moral philosophers who will get to turn their ideas into all sorts of apps and algorithms that will hopefully aid moral decision making for individuals and groups.
Some might ask is leaning on algorithms to help one make more decisions a form of “cheating”? I’m not sure, but I don’t think so. A moral app reminds me of a story I heard once about George Washington that has stuck with me ever since. Washington from the time he was a small boy into old age carried a little pocket virtue guide with him full with all kinds of rules of thumb for how to act. Some might call that a crutch, but I just call it a tool, and human beings are the only animal that can’t thrive without its tools. The very moral systems we are taught as children might equally be thought of as tools of this sort. Washington only had a way of keeping them near at hand.
The trouble lies when one becomes too dependent on such “automation” and in the process loses touch with capabilities once the tool is removed. Yet this is a problem inherent in all technology. The ability of people to do long division atrophies because they always have a calculator at hand. I have a heck of a time making a fire without matches or a lighter and would have difficulty dealing with cold weather without clothes.
It’s certainly true that our capacity to be good people is more important to our humanity than our ability to do maths in our head or on scratch paper or even make fire or protect ourselves from the elements so we’ll need to be especially conscious as moral tools develop to still keep our natural ethic skills sharp.
This again applies more broadly than the issue of moral machines, but we also need to be able to better identify the assumptions that underlie the programs running in the instruments that surrounded us. It’s not so much “code or be coded” as the ability to unpack the beliefs which programs substantiate. Perhaps someday we’ll even have something akin to food labels on programs we use that will inform us exactly what such assumptions are. In other words, we have a job in front of us, for whether they help us or hinder us in our goal to be better people, machines remain, for the moment at least, mere agents, and a reflection of our own capacity for good and evil rather than moral actors of themselves.