Defining Artificial Intelligence: Regulating a Moving Target

• In Event: Algorithms at Work: The Practice of Prediction - III: Reframing Algorithmic Accountability in Practice

Abstract

While policy and regulation for artificial intelligence (AI) systems must begin with a shared definition of artificial intelligence, definitions of AI vary widely across experts. Many scholarly and popular discourses share a vision of an artificial general intelligence, but “intelligent” algorithmic systems in existence today do not fulfill this prophecy in practice. Yet a broad range of applications, from the humble Roomba to the mighty IBM Watson, are labeled “AI”. Such a varied set of technologies considered to be “intelligent” reveals the relationality and situatedness of a community’s conception of what AI is, what it does in the world, and what it might someday be capable of doing. Here we investigate the diversity of mental models across communities of practice about what constitutes AI. We conduct field interviews at three major artificial intelligence conferences, an ethnography of expert policy makers, an analysis of discussion of artificial intelligence across three social media platforms, and an analysis of coverage of artificial intelligence in the news media. We find little consensus on what properties constitute artificial intelligence, even among experts, and instead find that people’s individual visions of AI are shaped by their own hopes, fears, and commitments. Given this definitional quagmire, we offer an alternative perspective. Drawing on Latour’s Actor-Network Theory and Hutchin’s concept of distributed cognition, we argue that a distinct “artificial intelligence” does not and will likely never exist. In our discussion, we argue that artificial intelligence cannot be distinguished from the broader label of “algorithmic systems”.

Authors

• Peaks Krafft

• Karen Y Huang, Harvard University

• Michael Katell, University of Washington (Information School)

• Meg Young