Some remarks on the ethical relevance of leaky abstractions

Abstract

Algorithms occupy a central place in the current debates on the ethics of computing. The term “algorithm”, however, is used to refer both to specific socio-technical systems and to the logical abstractions at the heart of these systems, and this ambiguity can mislead analyses of the ethical impact of real-world applications of AI systems. In this paper, I discuss two analytical failure modes that follow from this ambiguity — algorithm reductionism and algorithm fetishism —, arguing that they stem from the attempt to find an ideal vantage point for understanding an AI system. To mitigate the impact of reductionism and fetishism, I propose that we need to adopt analytical tools that look at various levels of abstraction rather than picking one of them as the correct description of a system. Such a multi-layered approach would contribute to the design of effective ethical and legal interventions in AI systems, thus contributing to the socially beneficial use of intelligent technologies.