Author: Carmen Mazijn, Carina Prunkl, Andres Algaba
Publisher: Institute for Ethics in AI, Oxford University
Publication Year: 2022
Summary: The following paper discusses how artificial intelligence systems can generate, propagate, support, and automate bias in decision-making processes. Most notions of group fairness evaluate a model’s equality of outcomes by computing statistical metrics on the outputs. The authors contend that these output metrics face inherent challenges and present a complementary approach that aligns with the growing emphasis on equality of treatment. They generate a canonical set that shows the desired inputs for a model given a preferred output by Locating Unfairness through Canonical Inverse Design (LUCID). By repeatedly interrogating the decision-making process, the canonical set reveals the model’s internal logic and exposes potential unethical biases.