When Imprecision Improves Advice: Disclosing Algorithmic Error Probability to Increase Advice Taking from Algorithms

Abstract

Due to the increasing number and complexity of information and available data, people use algorithmic decision support systems to improve the quality of their decisions. However, research has shown that despite an increasing “algorithm appreciation”, people still abandon algorithmic decision support when they see it fail – even if they know that the alternative, the human advisor, is advising worse. As algorithms consist of statistical models and never mirror the reality completely, mistakes of algorithms are inevitable – and as described, the following algorithmic refusal posts a threat to the quality of subsequent decision-making. With human advisors, people use the confidence of the advisor, often operationalized as a probability or Likert-scale estimate, as an indicator for correctness. The question arises whether information about algorithmic insecurities, e.g., their error probabilities, can serve a similar goal: to increase the transparency for human users and eventually counteract algorithmic refusal. Therefore, the authors are currently developing a study design that compares how algorithms that indicate vs. do not indicate an estimated accuracy influence participants’ advice taking. To investigate the effect on algorithmic refusal, participants get wrong information from an algorithm that indicates vs. does not indicate its error probability and in a second round, participants can decide whether they will use it again. The poster will describe open questions and limitations and provide a basis for further discussion on the ongoing research.