Effects of Advisor Accuracy on the Utilization of Algorithmic and Human Advice

While algorithmic decision making is frequently superior to human judgement, it has been repeatedly found that experiencing imperfect forecasts goes along with a decreased willingness to entrust decisions to an algorithm, constituting the so-called algorithm aversion (Dietvorst et al., 2015). In two experiments, it was investigated whether perceiving an algorithmic advisor as imperfect is sufficient for algorithm aversion to occur. All studies systematically varied the advisor (algorithm or human) in a within-subjects Judge-Advisor System paradigm (e.g., Logg et al., 2019). In Experiment 1, information about the advisor accuracy (low, high or no information available) was systematically manipulated within participants, while Experiment 2 demonstrated the performance of an algorithmic or human advisor of identical, but non-perfect accuracy prior to the task. Results of both experiments showed that participants adhered more strongly to advice given by an algorithmic as compared with a human advisor and were more willing to incorporate advice from highly as compared with less accurate advisors. Importantly, there was no indication of algorithm aversion in any of the results, strongly suggesting that imperfect algorithmic forecasting is not the sole determinant of algorithm aversion.