Plenty of Blame to Go Around: A Qualitative Approach to Attribution of Moral Responsibility

We present a computational model of blame attribution. Recently Mao and Gratch, following Attribution theory, created a computational model that assigned blame to an agent for a negative occurrence. Their model made categorical judgments, and could only assign blame to a single agent. Our model extends this work, using QP theory to provide a continuous model for the parameters involved in attribution and directly capturing the constraints postulated by Attribution theory. This allows our model to infer relative amounts of blame in a situation in a manner that is consistently overall with relative amounts of blame attributed in a psychological experiment. Who is to Blame? Bad things happen, and an important capability of social agents is to understand who is responsible. From the affairs of nations to personal misfortunes, accountability is an important part of how we understand the world around us. But how does a person go from perceiving such a situation to making a judgment of blame? This question has been the topic of much research in social psychology. Driven by the need to create social agents that can interact with people for a variety of purposes (tutoring, entertainment, assistants), creating computational models to capture such judgments is receiving increased attention. Without an understanding of blame assignment, an agent cannot properly infer the implications of social interactions. This paper describes how Qualitative Process theory [Forbus 1984] can be used in such modeling. We briefly summarize the elements of Attribution theory that address blame judgments. We then discuss the Mao and Gratch computational model [Mao & Gratch 2005][Mao 2006]. We present an alternative model for attribution of blame based on QP theory, which we claim better represents the underlying theory as well as human data. We present an evaluation of our model using data collected by Mao, showing that our model captures that data better, and makes additional predictions