Sex, Lies, and AI: Consumer Beliefs About the "Mental" Representations of Algorithmic Recommendations

The growing algorithmic assistance across industries, from diagnosing a disease in the healthcare industry to matchmaking in online dating settings, has received considerable attention over the past two decades. Yet, evidence on consumer reactions to algorithmic assistance is mixed. While research on algorithm aversion revealed a strong preference of individuals to favor human over algorithmic assistance (Dietvorst et al., 2015), recent work provides equal evidence for algorithm appreciation (Logg et al., 2019). Earlier reviews synthesizing the current stream of research on consumer reactions to algorithmic assistance is either based on qualitative reviews (e.g., Hasler et al., 2013; Janssen et al., 2019), at the risk to reach different conclusions examining the same stream of literature (Buecker et al., 2021), or more narrowly focusing on specific characteristics such as the extent of anthropomorphism in algorithmic assistance (Blut et al., 2021) or assessing more narrow outcomes such as trust (Schaefer et al., 2016). The current work provides a comprehensive review and meta-analysis to assess the overall effect size of algorithmic assistance across a large number of reported studies (211 effect sizes), testing a variety of moderating effects, and providing guidelines that otherwise lead to an overestimation of algorithm aversion.