Measuring understandability of intelligent systems: Scale development and validation across three domains

Abstract

While modern intelligent systems using black-box algorithms have proved their usefulness in many areas, whether the systems’ decisions and intentions can be fully understood by human users is still a critical question. However, the measurement of system understandability is lacking, and it undermines the development of this direction. To fill in such a gap, we conducted three studies to construct a scale to measure the understandability in three intelligent systems. In Study 1, we developed the original scale items through document analysis and expert interviews. In Study 2, we exposed 307 participants to autonomous vehicle systems which provided different amounts of information in simulated takeover scenarios. The participants’ responses towards these systems were collected using the developed scale. Exploratory factors analysis found 4 factors (Explanation Comprehensiveness, Trustworthiness Calibration, Cognitive Accessibility, and Explanation Necessity), and they had significant correlation with important attitudinal behavioral outcomes including trust, usage intention, and satisfaction. In Study 3, we further validated the structural and criterion-related validity of the scale using a new sample of 347 participants interacting with medical and financial decision support systems. The results indicate that the developed scale is a reliable and effective tool for assessing understandability in different intelligent systems, with potential to significantly enhance the design of intelligent systems to be more user-friendly and comprehensible.