Towards a better understanding on mitigating algorithm aversion in forecasting: an experimental study

Abstract

Forecasts serve as the basis for a wide range of managerial decisions. With the potential of new data sources and new techniques for data analysis, human forecasters are increasingly interacting with algorithms. Although algorithms can show better forecasting performance than humans, forecasters do not always accept these algorithms and instead show aversion to them. Algorithm aversion has become a widely known phenomenon. Drawing on the seminal study of Dietvorst et al. (J Exp Psychol Gen 144(1):114–126, 2015), we extend the evidence on algorithm aversion by introducing three environmental variables from the management accounting literature. We argue that time pressure, “do your best” goals, and forecasters’ data input decision rights on the algorithms input mitigate algorithm aversion. To test our hypotheses, we conducted an experimental study with 1,840 participants overall. We found support for our hypothesis that time pressure mitigates algorithm aversion. We found evidence that the mitigation effect is based on forecasters’ loss of confidence in their own forecast when they are under time pressure. We found no support for our hypothesis on “do your best” goals or forecasters’ data input decision rights.

Decision-making in the era of AI support—How decision environment and individual decision preferences affect advice-taking in forecasts.

“…The influence of trust in an advice is repeatedly highlighted in advice-taking research and is one of the most examined factors in advice-taking literature (Gino & Schweitzer, 2008; Jung & Seiter, 2021; Logg et al, 2019, Önkal et al, 2009; Wærn & Ramberg, 1996). People have to decide if they trust an algorithm forecast and to what extent (Shin, 2022b).…”

“…With progressing digitalization, there will also be more digital or automated forecasts (Brynjolfsson & Mitchell, 2017; Collins et al, 2021; Mikalef & Gupta, 2021), which highlights the relevance of further research on the AI–human interaction (Maedche et al, 2019; Shin, 2022b). These inherent changes also have several implications for organizations as they are confronted with considering and potentially implementing a constant influx of new forecasting methods (Jung & Seiter, 2021). Therefore, it is of prime relevance for organizations to better understand how decisionmakers interact with forecast information and specifically, which factors influence the acceptance of a given forecast as the best forecast is worthless if it is not or solely inappropriately used.…”

“…Forecasts represent an instrument of corporate management intended to generate statements about the future and, consequently, are an important organizational success factor. In particular, they serve managers (decisionmakers) as a basis for their decisions (Jung & Seiter, 2021). Hence, they have a major influence on the strategic and operational target achievement of a company (Sanders & Manrodt, 2003).…”

Taking Advice from ChatGPT

“…Two core competing findings are algorithm aversion [29] (a tendency to disproportionately punish algorithms when they err) and algorithm appreciation [30] (a tendency to prefer algorithm advice prima facie). Numerous studies have explored mediating mechanisms, including task objectivity [31], perceived competence [32], human input [33], learning [34,35], and time pressure [36], among others [31]. One literature review categorizes these effects into algorithm characteristics (agency, performance, capabilities, and human involvement) and human characteristics (expertise and social distance) [37].…”

Algorithm Aversion and Other Causes of Bias in Decision Behavior

“…We speak of algorithm aversion when subjects decline the use of an algorithm even though it is clearly recognisable that their own decisions or those of experts are by no means more successful (for the usual definitions, see, for example, ). There is a considerable amount of research results available on measures which can mitigate algorithm aversion (see, for example, Hinsen et al 2022;Gubaydullina et al 2021;Jung and Seiter 2021;.…”