The Impact of AI Transparency and Reliability on Human-AI Collaborative Decision-Making

Abstract

Human-AI collaborative decision-making has become a prevalent interaction paradigm, but the lack of transparency in AI algorithms presents challenges for humans to understand the decision-making process. Such lack of comprehension can lead to issues of over-reliance or under-reliance on AI recommendations. In this study, we focused on a human-AI collaborative income predicting task and investigated the influence of AI transparency and reliability on task performance. The results revealed that when AI reliability was high (75% and 90%), transparency had no significant effects on human decision-making. However, at a lower level of reliability (60%), higher transparency levels led to increased compliance with AI suggestions, thereby demonstrating a persuasive effect. Further analysis indicated that compliance rates only improved when AI made correct decisions, rather than when AI made incorrect ones. However, transparency did not significantly impact humans' ability to correctly reject erroneous recommendations from AI, suggesting that increasing transparency alone did not enhance humans’ error detecting ability. In conclusion, when the reliability of AI is low, heightening transparency can promote appropriate dependence on AI without elevating the risk of over-reliance. Nevertheless, further research is necessary to explore effective strategies that can assist humans in identifying AI errors effectively.