While the notion of punishing machines for misdeeds has received some attention in science fiction, it seems worthwhile to take a brief philosophical look at this matter. This is because the future, or so some rather smart people claim, will see the rise of intelligent machines—machines that might take actions that would be considered misdeeds or crimes if committed by a human (such as the oft-predicted genocide).
In general, punishment is aimed at one of more of the following goals: retribution, rehabilitation, or deterrence. Each of these goals will be considered in turn in the context of machines.
Roughly put, punishment for the purpose of retribution is aimed at paying an agent back for wrongdoing. This can be seen as a form of balancing the books: the punishment inflicted on the agent is supposed to pay the debt it has incurred by its misdeed. Reparation can, to be a bit sloppy, be included under retaliation—at least in the sense of the repayment of a debt incurred by the commission of a misdeed.
While a machine can be damaged or destroyed, there is clearly the question about whether it can be the target of retribution. After all, while a human might kick her car for breaking down on her or smash his can opener for cutting his finger, it would be odd to consider this retributive punishment. This is because retribution would seem to require that a wrong has been done by an agent, which is different from the mere infliction of harm. Intuitively, a piece of glass can cut my foot, but it cannot wrong me.
If a machine can be an agent, which was discussed in an earlier essay, then it would seem to be able to do wrongful deeds and thus be a potential candidate for retribution. However, even if a machine had agency, there is still the question of whether or not retribution would really apply. After all, retribution requires more than just agency on the part of the target. It also seems to require that the target can suffer from the payback. On the face of it, a machine that could not suffer would not be subject to retribution—since retribution seems to be based on doing a “righteous wrong” to the target. To illustrate, suppose that an android injured a human, costing him his left eye. In retribution, the android’s left eye is removed. But, the android does not suffer—it does not feel any pain and is not bothered by the removal of its eye. As such, the retribution would be pointless—the books would not be balanced.
This could be countered by arguing that the target of the retribution need not suffer—what is required is merely the right sort of balancing of the books, so to speak. So, in the android case, removal of the android’s eye would suffice, even if the android did not suffer. This does have some appeal since retribution against humans does not always require that the human suffer. For example, a human might break another human’s iPad and have her iPad broken in turn, but not care at all. The requirements of retribution would seem to have been met, despite the lack of suffering.
Punishment for rehabilitation is intended to transform wrongdoers so that they will no longer be inclined to engage in the wrongful behavior that incurred the punishment. This differs from punishment aimed at deterrence—this aims at providing the target with a reason to not engage in the misdeed in the future. Rehabilitation is also aimed at the agent who did the misdeed, whereas punishment for the sake of deterrence often aims at affects others as well.
Obviously enough, a machine that lacks agency cannot be subject to rehabilitative punishment—it cannot “earn” such punishment by its misdeeds and, presumably, cannot have its behavioral inclinations corrected by such punishment.
To use an obvious example, if a computer crashes and destroys a file that a person had been working on for hours, punishing the computer in an attempt to rehabilitate it would be pointless. Not being an agent, it did not “earn” the punishment and punishment will not incline it to crash less in the future.
A machine that possesses agency could “earn” punishment by its misdeeds. It also seems possible to imagine a machine that could be rehabilitated by punishment. For example, one could imagine a robot dog that could be trained in the same way as a real dog—after leaking oil in the house or biting the robo-cat and being scolded, it would learn not to do those misdeeds again.
It could be argued that it would be better, both morally and practically, to build machines that would learn without punishment or to teach them without punishing them. After all, though organic beings seems to be wired in a way that requires that we be trained with pleasure and pain (as Aristotle would argue), there might be no reason that our machine creations would need to be the same way. But, perhaps, it is not just a matter of the organic—perhaps intelligence and agency require the capacity for pleasure and pain. Or perhaps not. Or it might simply be the only way that we know how to teach—we will be, by our nature, cruel teachers of our machine children.
Then again, we might be inclined to regard a machine that does misdeeds as being defective and in need of repair rather than punishment. If so, such machines would be “refurbished” or reprogrammed rather than rehabilitated by punishment. There are those who think the same of human beings—and this would raise the same sort of issues about how agents should be treated.
The purpose of deterrence is to motivate the agent who did the misdeed and/or other agents not to commit that deed. In the case of humans, people argue in favor of capital punishment because of its alleged deterrence value: if the state kills people for certain crimes, people are less likely to commit those crimes.
As with other forms of punishment, deterrence requires agency: the punished target must merit the punishment and the other targets must be capable of changing their actions in response to that punishment.
Deterrence, obviously enough, does not work in regards to non-agents. For example, if a computer crashes and wipes out a file a person has been laboring on for house, punishing it will not deter it. Smashing it in front of other computers will not deter them.
A machine that had agency could “earn” such punishment by its misdeeds and could, in theory, be deterred. The punishment could also deter other machines. For example, imagine a combat robot that performed poorly in its mission (or showed robo-cowardice). Punishing it could deter it from doing that again it could serve as a warning, and thus a deterrence, to other combat robots.
Punishment for the sake of deterrence raises the same sort of issues as punishment aimed at rehabilitation, such as the notion that it might be preferable to repair machines that engage in misdeeds rather than punishing them. The main differences are, of course, that deterrence is not aimed at making the target inclined to behave well, just to disincline it from behaving badly and that deterrence is also aimed at those who have not committed the misdeed.