Claims about algorithmic decision-making, which are often accompanied by subsequent calls for greater accountability and regulation (Diakopoulos, 2015; Pasquale, 2015), are part of a much longer continuum of concerns about the ways in which technology can be said to have politics. According to this view, technology is never neutral but always already embedded with certain biases, values, and assumptions. At least since Winner’s (1986) influential arguments about Moses’ low-hanging bridges, scholars have followed suit in advocating the necessity to consider the politics of artifacts, particularly by atending to the values in design and the moral import of those design choices.24 Introna and Wood (2004), for example, argue that facial recognition systems are political in the sense that these algorithmic forms of surveillance carry certain assumptions about designated risks in the very design of the system. As Introna and Wood explain, the politics is mostly implicit, “part of a mundane process of trying to solve practical problems” (2004: 179). Every artifact implicating human beings, including algorithms, always carries certain assumptions and values about how the world works. Take the design of an ATM: “if you are blind, in a wheelchair, have problem remembering, or are unable to enter a PIN, because of disability, then your interest in accessing your account can be excluded by the ATM design” (2004: 179). The point is not that designers of ATM machines consciously or deliberately exclude people in wheelchairs from using an ATM but that the ways in which systems are designed always involve certain exclusionary practices that sometimes appear to be a more or less coherent and intentional strategy despite nobody “authoring” it as such (Introna and Nissenbaum, 2000).
The question of intentionality—whether someone explicitly authors an artifact to function in a particular way or whether its functioning is emergent—gets even more complicated with respect to machine learning algorithms. These algorithms are able to improve performance over time based on feedback. In addition, machine learning algorithms “create an internal computer model of a given phenomenon that can be generalized to apply to new, never-before-seen examples of that phenomenon”
Th e M u l tip l icity of A lgo rith ms
35