Toward a Secure Framework for Regulating Artificial Intelligence Systems

Summary

This paper proposes a technical framework for regulating high-risk AI systems, built on a concept called operational qualification and a temporal self-replacement test modeled on human operator certification. It introduces the operational qualification score as a measure of regulatory fitness and presents SFAIR, a tool for automated, recurrent, and secure examination and attestation of that fitness. The authors validate the approach on an open-source high-risk AI system and make SFAIR's source code publicly available.