Annie Liang

Abstract

Many predictions previously conducted by human experts (e.g., loan assessments and medical diagnoses) can now be automated. How should individuals choose between institutions where predictions are made by algorithms, and institutions where predictions are made by people? We propose a framework to examine one key distinction: Machine learning algorithms consider a fixed, standardized set of covariates for all individuals, whereas human evaluators adapt the choice of covariates to each person. Our framework defines and analyzes the advantage of this customization -- the value of context -- in environments with complex prediction problems. We show that unless the agent has prior reason to think that their individualized context is especially informative, then the benefit of more information generally outweighs the value of customization.