生成AIはなぜ「安全に止まれない」のか ―― 停止条件不在という構造的問題 ――

This paper examines the safety limitations of generative AI from a structural perspective, focusing on the absence of normative stop conditions. Current discussions on AI safety often emphasize issues such as misinformation, harmful outputs, alignment failures, and system control. However, this paper argues that these issues share a deeper structural cause: the lack of a defined criterion for when generation should stop. Generative AI systems are capable of producing highly optimized outputs, but they do not possess internal, normative conditions for terminating generation. Stopping is typically imposed externally through token limits, time constraints, or filtering mechanisms, which function as technical controls rather than meaningful or ethical completion criteria. This paper conceptualizes generative AI as a system that cannot reduce judgment, and argues that its safety challenges stem not from excessive intelligence or intent, but from the structural absence of stop conditions. The paper builds upon prior philosophical work on stop conditions and decision reduction, and is intended as a conceptual foundation for further research in AI ethics and safety.