Algorithm Credulity: Human and Algorithmic Advice in Prediction Experiments

This study examines algorithm credulity by which people rely on faulty algorithmic advice without critical evaluation. Using a prediction task comparing human and algorithm advisors, we find that participants are more likely to follow the same deficient advice when issued by an algorithm than by a human. We show that algorithm credulity reduces expected earnings by 13%. To explain this finding, we posit that people are more likely to perceive as credible an unpredictable and deficient piece of advice when produced by an algorithm than by a human. Overall, our results imply that humans might be particularly susceptible to the influence of deficient algorithmic advice.