Out of Control -- Why Alignment Needs Formal Control Theory (and an Alignment Control Stack)

Control theory offers a proven framework for AI alignment: The paper argues that formal optimal control theory—a well-established engineering discipline used to manage complex systems like aircraft and power grids—should be central to AI safety research, rather than relying solely on current AI safety approaches that may lack the generalization needed for frontier AI systems.

Current alignment methods have interoperability gaps: Existing AI safety and mechanistic interpretability research, while advancing formal methods, often fails to address how different alignment and control protocols can work together seamlessly—a critical gap the authors identify as essential for managing increasingly complex AI systems.

A hierarchical "Alignment Control Stack" is needed: The paper proposes organizing AI controls across multiple layers (from physical to socio-technical), with each layer having specific measurement and control characteristics. This layered approach would clarify what types of controls are possible at each level and how they interact.

Regulators need formal assurances, not just promises: The authors emphasize that governments and regulators will require rigorous, mathematically-grounded frameworks to verify that AI systems remain controllable and safe—something the proposed control theory approach could provide more credibly than current informal safety arguments.

Alignment requires understanding both potential and limitations: Rather than assuming AI systems can always be controlled, the paper advocates for honest analysis of what control methods can and cannot achieve at different system scales, which is essential for sustainable and trustworthy AI deployment.