Strategic Interactions with an Algorithm Assistant: The Power of Data and Mechanism

Abstract

We present a laboratory experiment that examines individuals’ willingness to take advice from artificial intelligence (AI)-based algorithms. We further explored the availability of data and mechanism underpinning the algorithm as the two main drivers behind AI-based algorithm acceptance. We find that providing users with the data used by the algorithm can substantially reduce their willingness to take advice from AI (the substitution effect); whereas making them aware of the mechanism underpinning the algorithm does not further crowd out algorithm reliance. If any, participants’ AI appraisal is higher when both data and mechanism are provided compared to when only data is provided (the information disclosure effect). However, most participants are reluctant to take advice from the algorithm, and the degree of algorithm aversion depends on their cognitive capability and self-confidence. Our results shed light on conditions under which users trust algorithms, and thus have considerable market and economic implications.

Keywords: algorithm aversion; artificial intelligence; centipede games; information acquisition

JEL Classification: D03; D60; D80