The Missing Link to Safe AI Homologation

Abstract

The release of the EU AI Act and Machinery Regulation provides the necessary foundation for the use of artificial intelligence in safety components. In this context, both regulations define specific requirements for quality management, risk management and technical documentation, mandatory for conformity assessment by a notified body. In this manuscript, we detail the challenges of AI-based safety-related systems during development and certification, as traditional safety assurance approaches fall short. Further, we illustrate the possible homologation procedures and the interplay of both legal instruments. Additionally, we outline the notion of the efficient assurance argument, including its purpose in certification, to provide a means for quick system assessment and its characteristics during development, in particularly, in designing and maintaining the system  in meeting the required tolerable risk level during its lifetime. Based on this, we detail the specific regulatory requirements from both regulations and analyze the characteristics of the efficient assurance argument in meeting these qualities. We conclude, that this type of assurance argument can be seen as the missing link to safe AI homologation, since it serves the engineering team during risk management, system design, verification, validation and testing and operations, as well as the notified body for type approval by demonstrating regulatory compliance.