Can computers cope with human races?

Abstract

In trying to apply a computer to a task that humans do, we often discover that it doesn't work. One common problem is that humans are able to deal with fuzzy concepts, but computers are not—they need precise representations and it is hard to represent a fuzzy concept in a precise way. However, if we look closer at such tasks, we often discover that the weakness actually lies not in the computer but in ourselves—that we didn't understand what we were doing in the first place. When faced with a problem of this sort, some people refuse to recognize the conceptual failure. Instead of seeking a better representation for the task, they thrash away at making the fuzzy scheme work, insisting that there is nothing wrong with the conceptual base. I will illustrate one such problem with a true story. The central theme is the fuzzy concept of racial and ethnic classification, as used by the U.S. government and a horde of other bureaucracies. These organizations have been carrying out elaborate statistical computations and making major policy decisions based on this concept for many years with problematical results. I begin with my first encounter with this scheme, some 25 years ago.