A Maturity Model for Trustworthy AI Software Development

Abstract

Recently, AI software has been rapidly growing and is widely used in various industrial domains, such as finance, medicine, robotics, and autonomous driving. Unlike traditional software, in which developers need to define and implement specific functions and rules according to requirements, AI software learns these requirements by collecting and training relevant data. For this reason, if unintended biases exist in the training data, AI software can create fairness and safety issues. To address this challenge, we propose a maturity model for ensuring trustworthy and reliable AI software, known as AI-MM, by considering common AI processes and fairness-specific processes within a traditional maturity model, SPICE (ISO/IEC 15504). To verify the effectiveness of AI-MM, we applied this model to 13 real-world AI projects and provide a statistical assessment on them. The results show that AI-MM not only effectively measures the maturity levels of AI projects but also provides practical guidelines for enhancing maturity levels.