Is Lawful AI Ethical AI?

Abstract

Attempts to impose moral constraints on autonomous, artificial decision-making systems range from “human in the loop” requirements to specialized languages for machine-readable moral rules. Regardless of the approach, though, such proposals all face the challenge that moral standards are not universal. It is tempting to use lawfulness as a proxy for morality; unlike moral rules, laws are usually explicitly defined and recorded – and they are usually at least roughly compatible with local moral norms. However, lawfulness is a highly abstracted and, thus, imperfect substitute for morality, and it should be relied on only with appropriate caution. In this paper, we argue that law-abiding AI systems are a more achievable goal than moral ones. At the same time, we argue that it’s important to understand the multiple layers of abstraction, legal and algorithmic, that underlie even the simplest AI-enabled decisions. The ultimate output of such a system may be far removed from the original intention and may not comport with the moral principles to which it was meant to adhere. Therefore, caution is required lest we develop AI systems that are technically law-abiding but still enable amoral or immoral conduct.