O ne way to view the puzzle of machine ethics is to consider how we might program computers that will themselves refrain from evil and perhaps promote good. Consider some steps along the way to that goal. Humans have many ways to be ethical or unethical by means of an artifact or tool; they can quell a senseless riot by broadcasting a speech on television or use a hammer to kill someone. We get closer to machine ethics when the tool is a computer that's programmed to effect good as a result of

more »

... he programmer's intentions. But to be ethical in a deeper sense-to be ethical in themselvesmachines must have something like practical reasoning that results in action that causes or avoids morally relevant harm or benefit. So, the central question of machine ethics asks whether the machine could exhibit a simulacrum of ethical deliberation. It will be no slight to the machine if all it achieves is a simulacrum. It could be that a great many humans do no better. Of course, philosophers have long disagreed about what constitutes proper ethical deliberation in humans. The utilitarian tradition holds that it's essentially arithmetic: we reach the right ethical conclusion by calculating the prospective utility for all individuals who will be affected by a set of possible actions and then choosing the action that promises to maximize total utility. But how we measure utility over disparate individuals and whether we can ever have enough information about future consequences are thorny problems for utilitarianism. The deontological tradition, on the other hand, holds that some actions ought or ought not be performed, regardless of how they might affect others. Deontology emphasizes complex reasoning about actions and their logical (as opposed to empirical) implications. It focuses on rules for action-how we know which rules to adopt, how we might build systems of rules, and how we know whether a prospective action falls under a rule. The most famous deon-tologist, Immanuel Kant (1724-1804), held that a procedure exists for generating the rules of actionnamely, the categorical imperative-and that one version of the categorical imperative works in a purely formal manner. Human practical reasoning primarily concerns the transformation between the consideration of facts and the ensuing action. To some extent, the transformation resembles a machine's state changes when it goes from a set of declarative units in a database to an output. There are other similarities, of course-humans can learn new facts that inform their reasoning about action, just as machines can incorporate feedback systems that influence their outputs. But human practical reasoning includes an intervening stage that machines (so far) seem to lack: the formation of normative claims about what is permissible, what one ought to do, what one is morally required to do, and the like. It's plausible that normative claims either are ethical rules themselves or entail such rules. These normative claims aren't independent of facts, and they don't necessarily lead humans to action. In fact, humans suffer from "weaknesses of the will," as Aristotle called them, that shouldn't be a problem for a machine: once it reaches a conclusion about what it ought or ought not to do, the output will follow automatically. But how will the machine reach the middle stage-the normative conclusions that connect facts to action through rules? I think this is the problem for machine practical reasoning. A rule-based ethical theory is a good candidate for the practical reasoning of machine ethics because it generates duties or rules for action, and rules are (for Rule-based ethical theories like Immanuel Kant's appear to be promising for machine ethics because they offer a computational structure for judgment.