Abstract:
This dissertation assesses how artificial neural networks (ANNs) and other machine learning systems should be devised, built, and implemented in US governmental organizations (i.e. public agencies). While it primarily focuses on ANNs given their current prevalence and accuracy, many of its conclusions are broadly applicable to other kinds of machine learning as well. It develops an analytical framework, drawn from diverse fields including law, behavioral psychology, public policy, and computer science, that public agency managers and analysts can utilize. The framework yields a series of principles based on my research methodology that I argue are the most relevant to public agencies. The qualitative methodology consists of an iterative approach based on archival research, peer review, expert interviews, and comparative analysis. Critically, this dissertation’s intent is not to provide the specific answers to all questions related to machine learning in public agencies. Given the speed at which this field changes, attempting to provide universally applicable answers would be difficult and short term at best. Rather, this framework focuses on principles which can help guide the user to the proper questions they need to ask for their particular use case. In that same vein, the normative principles it provides are procedurally focused in scope rather than focused on policy outcomes. In other words, this framework is meant to be equally applicable regardless of what one’s specific policy goals are.
