Machines in Decision Making: Algorithm Aversion and Procedural Transparency

Maschinen in der Entscheidungsfindung

Computational algorithms are widely present in everyday life e.g., conducting web searches, proposing music, matching job offers to applicants or suggesting the fastest route to a destination. In recent years, algorithms are increasingly involved in (computer-assisted) business, governance or health care decisions, rendering the interaction between human- and machine-made decisions a field with promising benefit, but also potential conflict.

Despite the widespread (and sometimes unnoticed) operation of algorithms and machine-based decisions, the general acceptance of algorithmic decisions as well as the confidence in their forecasts is rather low, constituting the so-called algorithm aversion (e.g., Dietvorst, Simmons, & Massey, 2014). It is a robust finding that people tend to have a critical reception of decisions made by machines and are hesitant to make use of the technology to inform their own conclusions. Therefore, the proposed project is concerned with the factors influencing the usage and perception of algorithms as a tool to assist human decision making as well as the subjective evaluation of automated decisions by (potential) users. The project is set out (1) to test the influence of providing procedural information about the algorithm on its usage and evaluation in decision making and (2) to explore the relevance of the specific scenario in which the decision is made. Overall results from the project will provide new insights into the causes of algorithm aversion as well as possible ways to mitigate (potentially unjustified) reservations against the use of algorithms in decision making.

All studies will follow an experimental approach, systematically varying conditions within and/or between participants and primarily measuring quantitative metrics of decision evaluation in the respective conditions. For example, trust, acceptance and confidence with respect to the decision and the decider (human/machine) will be measured. Qualitative measures and participant feedback about their decisions will also be included to allow for a comprehensive perspective on the assessment of algorithm aversion, as well as its potential modulation. All studies will be run as online surveys. Sample size calculations as well as statistical analyses will be performed in accordance with the respective literature.

The goal of the present project is twofold: First, the project will investigate the basis of algorithm aversion and its underlying determinants by systematically varying procedural information about the way algorithms operate with a specific focus on the incorporation of new information. If the (implicit) assumption that algorithms cannot incorporate new information contributes to algorithm aversion, providing information that decreases this belief should increase the perceived applicability of an algorithm. Second, potential boundary conditions for accepting or rejecting algorithms in decision making will be tested by varying aspects of the decision scenario. The evaluation of algorithmic decisions in different scenarios will also provide insight into the generalizability of the findings.

The project’s results will contribute to the theoretical understanding of the evaluation of algorithmic decisions and will also provide first advice to practitioners confronted with the task of implementing and/or using algorithms in the course of the digitalization of society and the workplace.