Virtue and Artificial Moral Intelligence

As AI decision-making plays an increasing role in our daily lives, and makes more and more important contributions to how we manage our most vital interests, the question of to what extent algorithms can be made sensitive to the full scope of our moral concerns is of ever greater concern. My proposal is that we can identify a scale along which we can place the suitability of an AI system for virtue judgments, meaning appropriately evaluating instances of virtue (or vice). To do so, I harness the notion of a developmental pathway, where virtue theorists describe the acquisition of virtue as a task involving movement along a number of stages of increasing sensitivity to and autonomous control over the facets that go into virtuous action. This prominently includes not just behavioral capacities (i.e. a carpenter building a table, an AI succeeding at a sorting task) but also psychological capacities, such as perception, imagination, foresight, and so on. We need to have multiple levels of evaluation in order to make sense of virtue judgments: that of how the action in question fares in this particular case, and how the disposition that this action is an instance of fares in the larger context of the agent’s life. We need appropriate interpretation of the subject matter of our judgments (not just as arrays of formal symbols, but mapped onto tangible situations) in order to appropriately judge the relationship between an individual action and the disposition which it is an instance of, since we need a sensible understanding of how having concrete disposition of character is of substantial import on concrete actions of the respective kind, something that involves more than formal relationships but rather an appreciation of the referents of the numbers and relations that go into machine learning.