Bias and Productivity in Humans and Machines

Abstract

Where should better learning technology (such as machine learning or AI) improve decisions? I develop a model of decision-making in which better learning technology is complementary with experimentation. Noisy, inconsistent decision-making introduces quasi-experimental variation into training datasets, which complements learning. The model makes heterogeneous predictions about when machine learning algorithms can improve human biases. These algorithms can remove human biases exhibited in historical training data, but only if the human training decisions are sufficiently noisy; otherwise, the algorithms will codify or exacerbate existing biases. Algorithms need only a small amount of noise to correct biases that cause large productivity distortions. As the amount of noise increases, the machine learning can correct both large and increasingly small productivity distortions. The theoretical conditions necessary to completely eliminate bias are extreme and unlikely to appear in real datasets. The model provides theoretical microfoundations for why learning from biased historical datasets may lead to a decrease (if not a full elimination) of bias, as has been documented in several empirical settings. The model makes heterogeneous predictions about the use of human expertise in machine learning. Expert-labeled training datasets may be suboptimal if experts are insufficiently noisy, as prior research suggests. I discuss implications for regulation, labor markets, and business strategy.

Keywords: machine learning, training data, decision algorithm, decision-making, human biases

JEL Classification: C44, C45, D80; O31, O33