Annie Liang

Abstract

This article examines how machine learning can contribute to economic models that explain behavior, rather than merely predicting outcomes. First, machine learning predictions can be used as benchmarks for economic models, for instance to quantify how far they are from the predictive limit. Second, algorithms can adversarially probe economic models to identify cases in which the theoretical predictions fail. Third, "hybrid" models can combine interpretable economic structure with flexible learning methods to leverage the strengths of both. Finally, large language models introduce qualitatively new possibilities, from simulating human responses to generating novel hypotheses. Throughout, I emphasize both promise and limits: while machine learning can uncover patterns that standard models miss, translating these algorithmic insights into interpretable and portable economic understanding typically still requires human judgment.