AI and Quality

The topic of reliability includes, for example, the accuracy and performance of the AI system. The key question here is whether correct and reproducible results are produced. The goal is for AI outputs to be correct compared to a known ground truth or defined target. Robustness additionally describes the extent to which the AI system continues to function reliably even with slightly altered or incomplete data. The application should still produce meaningful results in the presence of disturbances, unusual inputs or noise – or it should terminate in a controlled manner and provide appropriate information. Contingency plans and functional safety ultimately clarify what happens in the event of disruptions and whether the impact of such events is irreversible.

In the area of data quality, data protection and data governance, representativeness is a key factor. This concerns whether training data is sufficiently complete to avoid bias. Transparent provenance describes where the training data originates and whether it was collected responsibly. Data quality also includes the timeliness and accuracy of the data, meaning it must be clean, up-to-date and consistent. This includes not only raw data but also the quality of labels for both training and test data. Clear guidelines should exist, and methods such as redundant multi‑labelling should be used.

Protection of personal and proprietary data is essential: sensitive, internal information that is not publicly accessible and may provide a competitive advantage must be secured. The protection of individuals is equally important, particularly regarding identifiability, such as might occur in image processing. Data governance describes the extent to which there is a strategic framework of rules, processes, roles and policies that ensures the availability, quality, integrity and security of data throughout its entire lifecycle.

AI-specific cybersecurity and resilience against AI-specific attacks describe the extent to which the AI system is protected against manipulation, attacks or misuse. Technical, organisational and legal protective measures are essential to prevent deliberate misuse or accidental data breaches.

Non-discrimination is relevant from a quality perspective, ensuring that unjustified discrimination and bias are excluded. The AI application must not indirectly or directly disadvantage certain groups. There are strong links here to reliability and data quality: only when high-quality data is used to train an AI system can the results also meet the corresponding standard.

Human oversight and control include aspects such as human agency and supervision. Humans must retain full control over the AI system, be able to understand its decisions and must not be manipulated by the technology. The overarching principle is that humans decide whether and how AI is used, and must always have the ability to intervene. People must monitor and control AI systems during use and intervene when necessary. The objective is to prevent harm, discrimination or malfunctions, achieved through human monitoring, validation and verification of AI decisions. More on human oversight (for high-risk AI)