Numeric Default Logic as a Framework for Ethical AI

Abstract

Machine ethics aims to produce moral behaviour in artificial intelligence (AI) systems, by equipping such systems with moral reasoning capacities. One approach within machine ethics is to use deontic logics, i.e. logics that model moral reasoning. Horty has proposed a particular branch of default logic as a potential basis for deontic logic. Default logic models many intuitively desirable features of moral reasoning, such as the possibility of moral conflicts, the idea that moral rules can be overridden, and the distinction between different types of moral reasons. However, when considered in the context of AI ethics, Horty’s approach has some shortcomings. Specifically, the traditional, binary valuation of propositions appears unable to capture realistic decision-making scenarios, in which moral reasons can have a variety of strengths. Therefore, this paper explores a numeric default reasoning system, which extends Horty’s default logic with numeric valuations for propositions and default rules. It is shown that this new system preserves several advantages of Horty’s logic, but can also uniquely model certain intuitively common patterns of moral reasoning, specifically aggregation of reasons. Further directions are suggested, including exploiting the neurosymbolic character of the reasoning system to facilitate moral learning.