Minimum levels of interpretability for artificial moral agents

As artificial intelligence (AI) models continue to scale up, they are becoming more capable and integrated into various forms of decision-making systems. For models involved in moral decision-making (MDM), also known as artificial moral agents (AMA), interpretability provides a way to trust and understand the agent’s internal reasoning mechanisms for effective use and error correction. In this paper, we bridge the technical approaches to interpretability with construction of AMAs to establish minimal safety requirements for deployed AMAs. We begin by providing an overview of AI interpretability in the context of MDM, thereby framing different levels of interpretability (or transparency) in relation to the different ways of constructing AMAs. Introducing the concept of the Minimum Level of Interpretability (MLI) and drawing on examples from the field, we explore two overarching questions: whether a lack of model transparency prevents trust and whether model transparency helps us sufficiently understand AMAs. Finally, we conclude by recommending specific MLIs for various types of agent constructions, aiming to facilitate their safe deployment in real-world scenarios.