Reducing algorithm aversion through experience

Blame the Machine? Insights From an Experiment on Algorithm Aversion and Blame Avoidance in Computer-Aided Human Resource Management

“…Supporting prior research on blame avoidance in other fields of decision research ( Vis and van Kersbergen, 2007 ; Bartling and Fischbacher, 2012 ), we specifically contribute to HRM scholarship by revealing that study participants tasked with dismissing staff tend to delegate to the algorithm but only under certain conditions related to their confidence in human and machine forecast, echoing prior findings on the essential role of confidence in human and machine forecast by Filiz et al (2021) . In contrast, pretesting the algorithm reduces the likelihood of delegating an HR decision to it.…”

“…The study of Dietvorst et al (2015) was conducted with MBA students from a United States university, while our study relies on experimental data raised with a sample of both young professionals and graduate students from Germany. We contribute to the generalizability of the findings regarding algorithm aversion internationally and in practice, complementing recent scholarship by, among others, Lee (2018) , Newman et al (2020) , Filiz et al (2021) , and Renier et al (2021) . Although our findings on algorithm aversion are statistically significant and robust, the effect sizes we observe are substantially smaller—i.e., only one third as large as in Dietvorst et al (2015) .…”

“…The experiment revealed that study participants discounted the perceived accuracy of a prediction presented as algorithm-based advice much more steeply compared to the case in which the very same prediction was presented as human-made. Medical, psychological, and financial studies also show that people generally prefer human judgment over machine-aided models of prediction and find human forecasts more trustworthy ( Lee, 2018 ; Filiz et al, 2021 ). Research on computer-aided decision processes by Dzindolet et al (2002) and Renier et al (2021) shows that decision makers tend to perceive observed error rates of algorithm-based models as disproportionately more negative compared to human estimate-based models presumably because machine-made errors are incongruent to the widely held idea of perfection in machine-made forecast ( Stangor and McMillan, 1992 ; Madhavan and Wiegmann, 2007 ; Boyd and Crawford, 2012 ; Constantiou and Kallinikos, 2015 ; Berger et al, 2021 ) and because negative information cues are psychologically more salient than positive information cues ( Rozin and Royzman, 2001 ).…”

“…However, recent empirical studies show that the degree of decision makers’ reaction to pretesting an algorithm may be conditional upon the quality of this pretesting experience. The perceived relative precision of the machine vis-à-vis human forecast precision may influence decision makers’ response ( Filiz et al, 2021 ), as well as the specific task characteristic, particularly if stakes are high with regards to both tangible and moral costs (e.g., in the form of mental burdens; Lee, 2018 ; Burton et al, 2020 ). While research into this nexus is yet inconclusive ( Langer and Landers, 2021 ), Renier et al (2021) reveal that algorithmic fallibility will trigger harsher and comparatively more negative psychological reactions compared to experiencing equivalent human error and that a high-stakes HR choice context may be particularly salient in determining the acceptability of using an “imperfect” algorithm.…”

“…This phenomenon of algorithm aversion is intriguing because it contradicts the classic assumptions of utility maximizing behavior. While Dietvorst et al (2015 , 2018 ) successfully replicated their initial experiment, other studies further explored the effects of algorithm familiarity and trust ( Berger et al, 2021 ; Filiz et al, 2021 ), algorithms’ characteristics such as their ability to learn ( Berger et al, 2021 ), and their perceived fairness ( Newman et al, 2020 ) as determinants of algorithm aversion, the underlying psychological mechanisms of algorithm aversion are still not explored sufficiently enough yet. Particularly, the effect of different choice types and valences on algorithm aversion is not well understood yet (but see initial findings by Newman et al, 2020 , and Renier et al, 2021 ).…”