Fredda paused for a moment and looked out over the audience, and happened to catch Alvar Kresh’s eye in the first row. She was startled to see the anger in his face. What had happened? Kresh was a reasonable man. What could have angered him so? Had some piece of news come to him? That possibility put a knot in her stomach. But never mind. Not now. She had to go on with the lecture.
“At the beginning of my previous lecture, I asked, ‘What are robots for?’ There is a parallel question: ‘What are the Three Laws for?’ What purpose are they meant to serve? That question startled me when I first asked it of myself. It was too much like asking, ‘What are people for?’ or ‘What is the meaning of life?’ There are some questions so basic that they can have no answer. People just are. Life just is. They contain their own meaning. We must make of them what we can. But as with robots themselves, the Laws, I would remind you once again, are human inventions, and were most certainly designed with specific purposes in mind. We can say what the Three Laws are for. Let us explore the question.
“Each of the Laws is based on several underlying principles, some overt and some not immediately evident. The initial principles behind all three Laws derive from universal human morality. This is a demonstrable fact, but the mathematical transformations in positronic positional notation required to prove it are of course not what this audience wishes to hear about. There are many days when I don’t wish to hear about such things myself.”
That line got a bit of a laugh. Good. They were still with her, still willing to listen. Fredda glanced to her notes, took a slightly nervous sip from her water, and went on. “Suffice to say that such techniques can be used to generalize the Three Laws such that they will read as follows: One, robots must not be dangerous; two, they must be useful; and three, they must be as economical as possible.
“Further mathematical transformation into the notation used by sociological modelers will show that this hierarchy of basic precepts is identical to a subset of the norms of all moral human societies. We can extract the identical concepts from any of the standard mathematically idealized and generalized moral social codes used by sociological modelers. These concepts can even be cast into a notation wherein each higher law overrides the ones below it whenever two come into conflict: Do no harm, be useful to others, do not destroy yourself.
“In short, the Three Laws encapsulate some ideals of behavior that are at the core of human morality, ideals that humans reach for but never grasp. That all sounds very comfortable and reassuring, but there are flaws.
“First, of necessity, the Three Laws are set down, burned into the very core of the positronic brain, as mathematical absolutes, without any grey areas or room for interpretation. But life is full of grey areas, places where hard-and-fast rules can’t work well, and individual judgment must serve instead.
“Second, we humans live by far more than three laws. Turning again toward results produced by mathematical modeling, it can be shown that the Three Laws are equivalent to a very good first-order approximation of idealized moral human behavior. But they are only an approximation. They are too rigid, and too simple. They cannot cover anything like the full range of normal situations, let alone serve in unusual and unique circumstances where true independent judgment must serve. Any being constrained by the Three Laws will be unable to cope with a wide range of circumstances likely to occur during a lifetime of contact with the available universe. In other words, the Three Laws render a being incapable of surviving as a free individual. Relatively simple math can demonstrate that robots acting under the Three Laws, but without ultimate human control, will have a high probability of malfunctioning if exposed to human-style decision situations. In short, the Three Laws make robots unable to cope unaided in an environment populated with anything much beyond other robots.
“Without the ability to deal in grey areas, without the literally thousands of internalized laws and rules and guidelines and rules of thumb that guide human decision making, robots cannot make creative decisions or judgment calls even remotely as complex as those we make.
“Aside from judgment, there is the problem of interpretation. Imagine a situation where a criminal is firing a blaster at a police officer. It is a given that the police officer should defend him- or herself, even to the use of deadly force. Society entitles—even expects—the police officer to subdue or even kill his attacker, because society values its own protection, and the officer’s life, over the criminal’s life. Now imagine that the officer is accompanied by a robot. The robot will of course attempt to shield the policeman from the criminal—but will likewise attempt to protect the criminal from the policeman. It will almost certainly attempt to prevent the police officer from firing back at the criminal. The robot will attempt to prevent harm to either human. The robot might step into the police officer’s field of fire, or let the criminal escape, or attempt to disarm both combatants. It might attempt to shield each from the other’s fire, even if that results in its own destruction and the immediate resumption of the gun battle.
“Indeed, we have run any number of simulations of such encounters. Without the robot present, the police officer can most often defeat the criminal. With a robot along, here are the outcomes more likely than the police winning: death of police officer and criminal with destruction of robot; death of police officer and with destruction of robot; destruction of robot coupled with escape of criminal; death of criminal and/or police officer with robot surviving just long enough to malfunction due to massive First Law/First Law and First Law/Second Law conflicts. In short, inject a robot into such a situation, and the odds are superb you will end up with a disaster.
“Theoretically it is possible for a robot to judge the situation properly, and not mindlock over being complicit in the death of the criminal. It must be able to decide that both the immediate and long-term general good are served if the police officer wins, and that coming to the assistance or defense of a criminal prepared to take the life of a peace officer is ultimately self-defeating, because the offender will almost certainly attack society again in other ways, if he or she is permitted to survive. However, in practice, all but the most sophisticated robots, with the most finely tuned and balanced potentials of First Law, will have no hope at all of dealing appropriately with such a situation.
“All the laws and rules we live by are subject to such intricacies of interpretation. It is just that we humans are so skilled, so practiced, in threading our ways through these intricacies that we are unaware of them. The proper way to enter a room when a party is in progress in midafternoon, the correct mode of address to the remarried widow of one’s grandfather, the circumstances under which one mayor may not cite a source in scholarly research—we all know such things so well we are not even aware that we know them. Nor is such practiced knowledge limited to such trivial issues.
“For example, it is a universal of human law that murder is a crime. Yet self-defense is in all places a legitimate defense against the accusation of murder, negating the crime and condoning the act. Diminished capacity, insanity defenses, mitigating circumstances, the gradations of the crime of murder from manslaughter to premeditated murder—all these are so many shades of grey drawn on the black and white of the law against murder. As we have seen with my example of the policeman and the criminal, no such gradations appear in the rigidity of the First Law. There is no room for judgment, no way to account for circumstances or allow for flexibility. The closest substitute for flexibility a robot may have is an adjustment in the potential between the First, Second, and Third Laws, and even this is only possible over a limited range.