Clarifying Decision Making Processes: Tools for Interdependency Modeling

Abstract

Tools for problem specification in AI Decision making are underdeveloped at present. I propose two new tools for this purpose; first, a model of AI Decision Making, which supports problem identification and mitigation. Second, a Bill of Assumptions for Data Production. Data is an important component of AI Decision Making Systems, and data is necessarily produced by making a series of assumptions. My Bill of Assumptions for Data Production is a new approach to communicating these assumptions that facilitates collaboration, data transparency, and reduction of harmful bias. I illustrate this new approach by developing a dataset that estimates the distribution of Government education spending in the US across income deciles. My dataset informs existing Distributional National Accounts (DINA), which are a primary measure of income inequality in the US (Piketty et al., 2018). My estimate shows Government education spending is more progressive than assumed in current DINA. Furthermore, I show that removing federal education funding to postsecondary institutions would produce substantial harm.