Functional Trustworthiness of AI Systems by Statistically Valid Testing

Abstract

This position paper critiques the EU AI Act's conformity assessment procedures, arguing they fail to ensure genuine functional guarantees for AI systems. The authors advocate for reliable assessment of the statistical functional properties of an AI system as the mandatory nucleus of any conformity evaluation. The paper identifies three essential components: defining the technical distribution of the application domain, establishing risk-based minimum performance requirements, and conducting statistically valid testing using independent random samples. The framework provides a rigorous foundation for meaningful AI trustworthiness assessment in regulatory contexts.

Peer Review

Community peer review and discussion of this publication.