Traditionally, ethics is discussed with reference to human behavior. In this sense, ethics is the philosophical discipline concerned with what is meant when one says that a person acts well, and with the reasons why one says that a person has acted well.

Most of the ethical problems associated with artificial intelligence, although characteristic of it and of enormous social relevance, can be counted among the ethical problems raised by every technology with a major historical and social impact. Some issues related to artificial intelligence, however, touch the very foundations of the concept of ethics. Reflection on the possibility of developing artificial systems whose behavior is almost indistinguishable from that of a human being, together with actual technical developments in the field of artificial intelligence, has raised doubts about the scope of this discipline. Indeed, the question is whether it makes sense to imagine ethical norms and attributions of responsibility that concern intelligent systems in general, without distinguishing between artificial systems and those that are not. This question is essentially connected to the one (already discussed in C is for Chatbot) concerning the philosophical appropriateness of the thesis according to which certain artificial systems have the ability to think—a capacity presumably indispensable if one wishes to consider them morally responsible subjects for their behavior.

A question closely related to this concerns the concept of computational trust. The English-language literature makes much use of the term “trustworthy AI” in reference to the development of methods for checking that artificial intelligence systems do what they are expected to do. The concept of trust, however, has traditionally been applicable only to relationships between human beings. The fact that we speak of “betraying an individual’s trust,” in fact, suggests that there is an important ethical component in relationships of trust. Thus, the question of attributing moral responsibility to artificial intelligence systems arises once again. In fact, the reasons why some of these systems are so autonomous, efficient, and versatile are the very same reasons underlying their unpredictability and the difficulties encountered when attempting to provide a meaningful interpretation of the internal mechanisms that govern them. The philosophical problem concerning their ethical status is therefore inextricably intertwined with the technical problem concerning their predictability.

Those who harbor a certain skepticism about the idea that a computer can have consciousness should not, however, be under any illusions. Even if one wishes to conclude that artificial intelligence systems cannot be considered capable of thought, numerous highly significant ethical questions concerning them remain open. Indeed, although it makes little sense, if not outright no sense, to maintain that an artificial system incapable of thought can make moral or immoral decisions, the problems connected with the development, production, commercialization, and use of these systems remain.

Of particular interest, for example, are the problems related to the moral responsibility of the human beings who develop, train, or employ an intelligent artificial system with regard to that system’s actions. These problems are particularly thorny precisely because of the autonomy of these systems and because the behavior of such a system does not depend exclusively on the code written by the person who programmed it, but also depends to a large extent on the data on which it was trained.

Another family of ethical problems relating to artificial intelligence is closely connected to the dependence of certain intelligent systems’ behavior on the quality of the data on which they were trained: the problems associated with these systems’ potential biases and potential unfairness. The data on which these systems are trained may be incomplete or of low quality because of biases that have tainted its collection. Moreover, even if these data were collected using appropriate methods, the data themselves may concern a historical period during which the inequalities and biases that we seek to counteract in the present were stronger and more widespread. The use of inadequate data entails a high risk of producing systems that exhibit, in their behavior, the same biases and the same inequalities that taint the data on which they were trained.

Francesco A. Genco

(translated by MIRAI)