How can AI regulation be effectively enforced? : comparing compliance mechanisms for AI regulation with a multiple-criteria decision analysis

Abstract

Newly emerging AI regulations need effective and innovative enforcement and compliance mechanisms to assure that fundamental and human rights are protected when using an AI system. This study compares four different compliance mechanisms namely ‘Real-Time and Automated Conformity Assessment’, ‘Standardization and Certification’, ‘Algorithmic Impact Assessment’ and ‘Algorithmic Auditing’ as well as three different assurers of compliance namely deployers, notified bodies and civil society organisations. With an MCDA, this research has shown that civil society-based compliance mechanisms are believed to be less effective, less feasible and more costly compared to all other compliance mechanisms. Second, external compliance mechanisms (by notified bodies) were rated to be more effective but also more difficult to implement compared to internal compliance mechanisms. Third, algorithmic auditing scored highest among all policy options. Fourth, despite its experimental nature, automated and real-time compliance mechanisms are not scored significantly lower than other compliance mechanisms.