Guide to Designing Trustworthy AI: The AI Test Catalog Explained

The 'AI Test Catalog' (KI-Prüfkatalog) serves as a crucial tool for building trust and ensuring quality in AI systems. It operationalizes quality requirements and existing AI guidelines, employing a risk-based approach to testing. The catalog is designed to be applicable across a range of AI performance capabilities and application areas, fitting within existing testing frameworks. Its structure is meticulously organized to guide users through the process of evaluating and ensuring the trustworthiness of AI applications. The initiative is supported by KI.NRW, the central hub for AI in North Rhine-Westphalia, aiming to accelerate AI transfer from research to industry and foster societal dialogue, always prioritizing human ethical principles in AI development. The 'Certified AI' project further supports this by developing standardized testing criteria, methods, and tools to guarantee technical reliability and responsible AI practices.

The AI Test Catalog evaluates AI systems based on several critical dimensions of trustworthiness. These dimensions are designed to cover the multifaceted aspects required for responsible and reliable AI deployment. The core dimensions include: Fairness, ensuring equitable treatment and outcomes; Autonomy and Control, defining the appropriate balance of human and AI decision-making; Transparency, enabling understanding of AI operations and decisions; Reliability, guaranteeing consistent and dependable performance; Security, protecting against threats and ensuring system integrity; and Data Protection, safeguarding personal and sensitive information. Each dimension is further broken down into specific risk areas, providing a granular approach to assessment and mitigation.

Autonomy and Control (AK) is a crucial dimension that addresses the appropriate distribution of tasks and decision-making authority between humans and AI systems. The guide details the description and objectives for this dimension, along with a thorough protection needs analysis. It delves into specific risk areas, such as the 'appropriate and responsible design of task distribution between humans and AI applications' (GE) and 'ensuring the informedness and empowerment of users and affected parties' (IB). For each risk area, the catalog outlines risk analysis and goal setting, criteria for achieving these goals, and a range of measures covering data, the AI component, system embedding, and operational procedures. The final assessment provides an overall evaluation of the autonomy and control aspects.

The Reliability (VE) dimension ensures that AI systems perform consistently and dependably under various conditions. The guide provides a description and objectives for reliability, followed by a protection needs analysis. It breaks down reliability into several risk areas: 'Reliability under normal conditions' (RE), 'Robustness' (RO), 'Catching errors at the model level' (AF), and 'Assessment of uncertainty' (UN). For each of these areas, the catalog outlines the risk analysis and goal setting, criteria for achieving the desired reliability, and specific measures. These measures are detailed for data, the AI component, system embedding, and operational practices. An overall assessment is then presented for each risk area and for the entire dimension.

Data Protection (DS) is a critical dimension focused on safeguarding sensitive information processed by AI systems. The guide provides a description and objectives for data protection, followed by a comprehensive protection needs analysis. It identifies key risk areas, such as 'Protection of personal data' (PD) and 'Protection of business-relevant information' (GI). For each risk area, the catalog outlines the risk analysis and goal setting, criteria for achieving data protection objectives, and specific measures. These measures are detailed for data handling, the AI component, system integration, and operational practices. The overall assessment evaluates the effectiveness of data protection measures.