03/06/2023

Colloquium Series - Ben Green

Algorithmic Realism: Reimagining Data Science to Promote Social Justice

The field of data science faces an impasse. With an ability to uncover empirical patterns and predict future events, data science appears to be a valuable tool for improving society. Data scientists have therefore proposed using algorithms to inform decision-making in high-stakes settings such as courts, schools, and hospitals. However, humanists, social scientists, and affected communities have exposed how these same algorithms can discriminate against racial minorities, marginalize workers, and increase surveillance. In response, data scientists have developed new methods focused on ethics and fairness. Yet even these methods fail to prevent—and at times exacerbate— the most severe harms of algorithmic decision-making. Thus, as data scientists strive to responsibly contribute to the social good, the field confronts a fundamental question: how can data scientists develop algorithms that reliably improve society?

Algorithmic Realism breaks through this impasse by presenting a guide for using data science to promote social justice. I argue that this task requires developing a new methodology for algorithm design and evaluation. The current methodology—which I call “algorithmic formalism”—focuses on the formal mathematical attributes of algorithms. As a result, even algorithms that satisfy disciplinary standards for rigor can entrench unjust social conditions, impose narrow logics of efficiency across domains, and fail to achieve the desired real-world benefits. To remedy these issues, I propose an alternative approach. “Algorithmic realism” is a data science methodology that situates design and evaluation around the real-world impacts of algorithms. Algorithmic Realism describes how to transform data science from a methodology based on formal, mathematical models into a practical tool for addressing real-world problems. This approach reveals why many high-profile algorithms (such as pretrial risk assessments) are unnecessary or even detrimental tools for reform. However, algorithmic realism also suggests new strategies for how data scientists can contribute to a more equitable society.