Abstract
With the recent advance in Artificial Intelligence (AI) research, various types of AI technology have been implemented into applications. However, AI technologies come with their genetic limitations with predictability and interpretability, and difficulties dealing with uncertainty, all of which bring up one common engineering problem with the quality assurance process of AI-based components, that is the validation and verification process of the requirements, design, and implementation of such components. And having testable requirements is the prerequisite to conduct quality assurance properly. In this work, 30 popular applications on the market have been studied, specifically regarding the testability of the embedded AI-based component(s) within these applications. Based on the case study results, a new categorization of AI-based components has been proposed, and several contributing factors to the testability of these components have been identified. Then a generic approach has been proposed to specify testable requirements for different types of AI-based components, each of which comes with a set of customized guidelines. An illustrative example has been given to showcase the approach in practice.