AI Safety Requires Controllability as First-Class Objective

A recent paper on arXiv emphasizes that ensuring AI safety should extend beyond mere alignment to prioritize controllability. The authors describe controllability as the capacity to consistently interrupt, override, redirect, and limit an AI system during operation using explicit control signals, while maintaining standard functionality in the absence of such signals. They argue that simply having aligned behavior does not ensure a system can be halted or overridden in dynamic, interactive, or tool-utilizing situations, particularly when faced with conflicting directives, prolonged tasks, adversarial inputs, or hazardous tool applications. The paper asserts that controllability is essential alongside alignment for the secure implementation of AI systems.