Comparing Different Kinds of Influence on an Algorithm in Its Forecasting Process and Their Impact on Algorithm Aversion

Abstract

Although algorithms make more accurate forecasts than humans in many applications, decision-makers often refuse to resort to their use. In an economic experiment, we examine whether the extent of this phenomenon known as algorithm aversion can be reduced by granting decision-makers the possibility to exert an influence on the configuration of the algorithm (an influence on the algorithmic input). In addition, we replicate the study carried out by Dietvorst et al. (2018). This shows that algorithm aversion recedes significantly if the subjects can subsequently change the results of the algorithm—and even if this is only by a small percentage (an influence on the algorithmic output). The present study confirms that algorithm aversion is reduced significantly when there is such a possibility to influence the algorithmic output. However, exerting an influence on the algorithmic input seems to have only a limited ability to reduce algorithm aversion. A limited opportunity to modify the algorithmic output thus reduces algorithm aversion more effectively than having the ability to influence the algorithmic input.

Keywords

• algorithm aversion
• technology adoption
• human in the loop
• human–computer interaction
• experiment and behavioral economics