E VDE-AR-E 2842-61-4 Anwendungsregel:2023-10

Reference model for trustworthy Artificial intelligence (AI) has the great potential to change society, the economy and the coexistence of people in the long term. It can help to reduce emissions, detect diseases, transport people or goods or make processes more efficient. At the same time, AI poses great challenges to humanity regarding the powers of an AI system. The solution of these questions requires a technical implementation in the design of AI systems and algorithms. Technical rules must be implemented that define how AI components can be integrated into products and describe the upstream design process. The security of AI systems and solutions must be guaranteed at all times, as soon as they are used in the market. The special feature of AI systems is their complexity: Functions cannot simply be tested (by humans). The system itself must guarantee safety and, for example, inherently fulfil functional safety requirements. This VDE application guide will also provide requirements for the verification and qualification of such systems and thus map the entire life cycle. This VDE application guide places particular emphasis on the design phase (development): the boundaries compared to the classic limits of a machine are broader, since interaction with humans and interaction with the environment must also be specified. Especially for AI systems, trustworthiness requirements, such as physical safety, information security, usability or ethical and regulatory issues are of fundamental importance. The model of this VDE application guide will define requirements that contribute to trustworthiness. Among other things, this will lay the foundations for qualification and approval, which can be followed up in the further process. This VDE application guide will define various development steps based on the life cycle for AI systems. These include the design at system level (including trustworthiness attributes), component level (including hardware, software and AI blueprints for the application of an AI methodology), integration, verification and validation, and acceptance and release. The model also includes market monitoring requirements and corrective and preventive action (CAPA).