Risky Reforms: A Sociotechnical Analysis of Algorithms as Tools for Social Change

Abstract

This thesis considers the relationship between efforts to address social problems using algorithms and the social impacts of these interventions. Despite widespread optimism about algorithms as tools to promote reform and improve society, there is often little rigorous analysis regarding how algorithmic interventions will lead to particular desired outcomes. In turn, many well-intentioned applications of algorithms have led to social harm. In this thesis, I focus on the use of “risk assessments” in the U.S. criminal justice as a notable example of machine learning algorithms being used as tools for social change. Treating these algorithmic interventions as sociotechnical and political reform efforts rather than primarily technical projects, I center my analyses of risk assessments around their social and political consequences. In Part I (Interaction), I introduce a new “algorithm-in-the-loop” framework for evaluating the impacts of algorithms in practice, using experiments to uncover unexpected behaviors that occur when people collaborate with risk assessments. In Part II (Risk and Response), I interrogate typical conceptions of risk and how to respond to it, developing a novel machine learning method to analyze structural factors of violence and to support non-punitive and public health-inspired violence prevention efforts. In Part III (Reform), I place these technical studies in the broader context of social and political reform, describing the limits of risk assessments as a tool for criminal justice reform and articulating a new mode of practice—“algorithmic realism”—that synthesizes computer science, law, STS, and political theory in order to equip computer scientists to work more rigorously in the service of social change. By expanding the scope of questions asked of risk assessments, this dissertation sheds new light on how risk assessments represent a “risky” strategy for achieving criminal justice reform. Through these analyses, I chart the beginnings of a more interdisciplinary and rigorous approach to evaluating and developing algorithms as tools for social change.