The Decline of 'AI System': A Shift in AI Governance

The article highlights a critical disconnect between the way AI is governed and the way it is actually built and deployed. The term "AI system," initially conceived as a neutral abstraction for policy purposes, has failed to gain traction within the engineering community. This divergence stems from the fact that engineers focus on specific components like models, pipelines, and agents, while policymakers seek a broader, more easily definable entity. The OECD's earlier work on defining "AI systems" assumed clear boundaries and traceable responsibility, but modern AI deployments often involve complex chains of interconnected components that span multiple teams and organizations.

This mismatch poses significant challenges for accountability, risk management, and identity architecture. Organizations are being asked to inventory "AI systems" and assign owners, but the reality is that they operate fluid and dynamic networks of AI components. The lack of a clear definition of "AI system" creates ambiguity and hinders regulatory efforts. The article suggests that a more effective approach to AI governance would involve focusing on the interconnectedness of AI components and developing strategies for managing risk across the entire AI ecosystem.

The EU AI Act aims to regulate AI systems based on their risk level. The decline of the term "AI system" raises questions about the Act's effectiveness in addressing the complexities of modern AI deployments. A more nuanced approach that considers the interconnectedness of AI components and the potential for cascading risks may be necessary to ensure that AI is developed and deployed responsibly. The focus should shift from regulating individual "AI systems" to governing the entire AI ecosystem, promoting transparency, accountability, and ethical considerations at every stage of the AI lifecycle.