The Effect of Level of AI Transparency on Human-AI Teaming Performance Including Trust in Machine Learning Interface

Abstract

The objective of this study is to investigate the impact of various levels of transparency in a human-AI teaming task on human performance, including task performance, trust, and workload. A task simulator using real-time AI was developed and used to compare two different levels of information transparency in AI. A total of 20 participants participated in the experiment, and each participant was asked to play a Pictionary game by drawing given words while the AI presented a guess of the words, in a simple form of human-AI cooperation. The task performance was measured for two different levels of transparency for displaying the top 1 or top 5 objects that the AI recognized as being the most similar to the participant's drawing. During the experiment, task completion time, the number of errors, an eye movement profile, subjective workload, and subjective ratings on trust were collected and analyzed, along with this post-trial interview. Results revealed that participants paid more attention to information display under conditions of higher-transparency condition while ameliorating workload and increasing the level of trust in cooperating with AI. Interview results identified the importance of individual differences in HAT performance, and as suggestions in providing transparency along with explainability information. While the study includes limitations such as limited levels of transparency, it confirms the benefits of transparency and other human factors issues in HAT. It is expected that the study can serve as a basis for further studies to determine effective transparency in HAT.