Unravelling Responsible AI: An Umbrella Review

Abstract

The term ‘Responsible AI’ (RAI) has become widely adopted in various sectors, including industry, research, and policy, and has also entered general public discourse. There are significant similarities and overlap with terms such as Ethical AI and Responsible AI. As the terminology surrounding AI evolves, it is important to untangle the explicit and implicit meanings of RAI, its relationship with other relevant concepts, and the implications for the AI landscape. This paper examines the ways in which RAI has been defined and described in systematic reviews within academic research between 2013 and early 2024, by conducting an umbrella review. Five main questions are explored in these findings: 1)What is RAI? 2) What are the motivations behind RAI efforts? 3) What is its purpose? 4) What are the terms related to RAI and how they are related? and 5) What are the current challenges and future directions of RAI? This review highlights that despite the potential benefits of AI, there remain risks and concerns surrounding it. This in turn calls for a set of computational and human measures, as well as principles that span from risk mitigation to benefiting humans and addressing social problems. However, RAI is conflated, used interchangeably with, or comprises other terms, and ‘responsible’ and ‘responsibility’ are often used with different connotations. There is also an interesting contradiction between the proliferation of RAI frameworks, and the need for more actionable or operationalisable ones. We discuss the implications of these findings and offer recommendations in light of current challenges and future directions to elucidate the meanings and understandings of RAI.