The Moral Status of Input and Output Discrimination

Abstract

A salient concern regarding artificial intelligence (AI) systems is that they may involve objectionable forms of discrimination. This concern is usually associated with the possibility that the input and, as a result, the output of such systems may be inaccurate. However, the concern is sometimes also present when the input and output are accurate (or when they reflect the best available evidence). The paper considers when this concern is warranted and what should be the response to it. It submits that while there may be reasons against actions that rely on accurate input and output, these reasons are not always present, and when there are, they may not be decisive. Therefore, discrimination in the relevant senses should not always be avoided. The argument for this conclusion is thus based on two premises: that discrimination should be avoided only when it is wrong (or bad) overall as opposed to when it is merely pro tanto wrong (or bad); and that discrimination in the senses that are often considered in the AI context is not necessarily (and thus not always) wrong (or bad) overall.

Keywords: Discrimination, Artificial Intelligence (AI), Input, Output, Moral Status

JEL Classification: I10, I14, K00, K20, K32, L38, Z18