When Algorithms Err: Differential Impact of Early vs. Late Errors on Users' Reliance on Algorithms скачать в хорошем качестве

When Algorithms Err: Differential Impact of Early vs. Late Errors on Users' Reliance on Algorithms

(Prepared for INFORMS CIST 2020 #INFORMS2020 #CIST2020) Abstract Errors are a natural part of the development and use of predictive algorithms, but such errors may discourage users from relying on algorithms even when doing so could lead to better decisions. In this paper, we conduct two experiments to demonstrate that people's reliance on a predictive algorithm following a substantial error is affected by (i) when the error occurs and (ii) how the algorithm is used in the decision making process. We find that, when the prediction tasks are fully delegated to an algorithm, the impact of an error on reliance depends on whether the error occurs early (i.e., when users first start using the algorithm) or late (i.e., after users have used the algorithm for an extended period). While an early error results in substantial and persistent reliance reduction, a late error affects reliance only temporarily and to a lesser extent. However, when users have more control over how to use the algorithm's predictions (as opposed to the case of complete delegation), the risk associated with early errors decreases and, as a result, error timing ceases to have a significant impact on their reliance. Our work advances the understanding of algorithm aversion and informs the practical design of algorithmic decision-making systems. For the paper and the presentation schedule, please go to: https://mlin.scheller.gatech.edu/Outs...