October 21, 2010

In the October issue of Scientific American you'll find an article on what might first appear to be science fiction, "Robot Be Good," but which claims to be about science-soon-to-be fact. As the subtitle touts, "Autonomous machines will soon play a big role in our lives. It's time they learned how to behave ethically."

The authors are Michael Anderson and Susan Anderson, respectively, a computer scientist working on Artificial Intelligence and a philosopher. The article shows a picture of Nao, a cute white and brown robot with blue eyes, whom the Andersons proudly claim is the "first robot to have been programmed with an ethical principle." Nao sports a halo—inserted by the magazine editor rather than earned by the robot, we assume.

The Andersons seem innocent enough. The examples they use—a robot working in a nursing home who must decide to which patient it should yield the TV remote or how many times it should nag a patient to take his medicine—don't involve any great ethical quandaries. The problem, however, is that the Andersons simply don't see any great ethical quandaries.

For the Andersons, ethics is simply following out a programmed algorithm—a sequence of clearly-defined steps that achieve a desired result—something computers can do far more quickly and thoroughly than human beings. Long division is an algorithm; so is finding the square root of a number. If you doubt a computer's capacity to outperform a human being in this area, time yourself against your calculator in dividing 65,439 by 89 or figuring out the square root of 15. So, the Andersons reason, a computer would ultimately be more efficient and impartial in churning through the seemingly endless particulars involved in real-life ethics.

What's wrong with the Andersons' rosy picture? Let's begin with the obvious. A human programmer creates the algorithm. If you don't see the problem, line up your robots and have each programmed by a different person, say, Pope Benedict, Mick Jagger, the avid euthanasia and infanticide advocate Peter Singer, a devout Muslim who upholds Sharia, and Pee-Wee Herman.  The programming of each will reflect the ethical caliber and goals of each, and their respective Naos will simply choose according to their irreconcilable views.

The Andersons have some recognition of the obvious: "Assuming that it is possible to give ethical rules to robots, whose ethical rules should those be? After all, no one has yet been able to put forward a general set of ethical principles for real-live humans that is accepted universally." But their response is as weak and confused as their enthusiasm is strong and bold: "machines are typically created to function in specific, limited domains," and in such particular situations, "most ethicists would agree on what is ethically permissible and what is not." And if that's not enough to ease your mind, "In situations in which there is no such agreement, we believe that machines should not be allowed to make autonomous decisions at all."

Alas, their solution only causes more problems. It simply isn't true, for example, that while ethicists disagree about general ethical rules, they agree on what to do in particular situations. If anything, it's the reverse. They can agree on general rules ("Do unto others as you would have them do unto you") but disagree about their application to particular issues in particular circumstances ("Does euthanasia follow from the Golden Rule?").

Another problem is one of inconsistency. Right after the Andersons state the above, they outline their own procedure for generating ethical principles using techniques taken from Artificial Intelligence research. The robotic program samples previous ethical decisions made by human beings and "abstracts" a general ethical principle that is then "encoded into the robot's programming." The general principle then acts as a general ethical rule—the very thing that they said human beings can't seem to agree upon!

That means, of course, that the Andersons actually believe that their robot might ultimately replace human beings in making ethical rules. They hint as much, stating that while human ethical decisions are all too often skewed by emotions and self-interest, making us "less than ideal ethical decision makers," a robot has neither self-interest nor emotions, so that "properly trained machines might even behave more ethically than many human beings would, because they would be capable of making impartial decisions, something humans are not always very good at."

We've thus come upon the final problem, the most insidious one. We already live in a culture which has largely taken to describing the human mind in terms of a machine, the computer. We speak of ourselves as hardwired to do this or that, of having data in our heads, of programming or reprogramming ourselves. Since we see our minds as no different than this particular machine, then it is quite easy for us to assume that this machine can take the place of our mind. If we see our ethical decision-making as being a kind of algorithmic process, then it only makes sense to yield our decision-making to the most powerful machine in running such programs. If we think ethics is merely a matter of calculations then it had better be done by the fastest calculators.

This is an especially dire threat for the following reason. We live in an age where there is more and more ethical disagreement, and so there will be a greater and greater temptation to find a solution, any solution. Given our worshipful attitude toward technology in general and computers in particular, we'd be inclined to seek just the kind of solutions offered by the likes of the Andersons. That would mean—to return to a previous point—we'd gladly yield moral authority to "ethical" computer programs, which ultimately means yielding to the entirely invisible men and women who wrote the programs.