Obligations to assess: Recent trends in AI accountability regulations.

Abstract

Policymakers are increasingly turning toward assessments of social, economic, and ethical impacts as a governance model for automated decision systems in sensitive or regulated domains. In both the United States and the European Union, recently proposed legislation would require developers to assess the impacts of their systems for individuals, communities, and society, a notable step beyond the technical assessments that are familiar to the industry. This paper analyzes four examples of such legislation in order to illustrate how AI regulations are moving toward using accountability documentation to address common AI accountability concerns: identifying and documenting harms, public transparency, and anti-discrimination rules. We then offer some insights into how designers of automated decisions systems might prepare for and respond to such rules.