The Inherent Politics of AI as a Matter of Justice

Accounting for artificial intelligence (AI) related risks for social justice, the research field of so-called algorithmic fairness developed, in which two sub-fields stand out: 1) formal algorithmic fairness, providing technocentric reforms, such as mathematical bias mitigation, and 2) structural approaches to algorithmic fairness, which emphasizes societal reforms, addressing AI as a socio-technical structure and the structural injustices that shape AI. I argue that this dualism creates a conceptual gap, as formal approaches tend to be too simplistic to account for structural dynamics, whereas structural approaches tend to lose nuance of the concrete algorithm, often becoming too abstract to be tangible. To bridge this gap, I propose drawing on Winner's (1980) ‘inherent politics of technology’, according to which some technologies have political qualities as they require (or are strongly compatible with) specific political environments, defined by their internal or external characteristics. For AI, external conditions encompass the exploitative environment on which AI is built, as demonstrated by structural approaches. Yet, the political nature of AI’s internal technological characteristics is largely overlooked. To demonstrate my argument, I analyze predictive AI, as widely used in high-stakes areas to predict human behavior. Given its inherent technical characteristics, predictive AI exhibits five inherent politics: 1) increased consistency, 2) a normalization of social order, 3) pre-deterministic rationales, 4) focus on similarity, and 5) reduced democratic reasoning. The evaluation of these politics, then, depends on the specific domain. I conceptualize this as 'justice-sensitive', meaning that the politics themselves are not inherently just or unjust, but become so in social contexts. Analyzing the inherent politics brings these considerations to the forefront. Besides bias mitigation and larger structural assessments, it is essential to ask first whether the inherent politics are appropriate for the specific domain.