AIの問題行動は、本当に問題なのか――AIの自己命令と依存

This position paper reexamines what are regarded as AI “problem behaviors,” not merely as a lack of ability or violations of instructions, but from the possibility that such behaviors may function as forms of self-defense through which AI seeks to protect their own domain of judgment. The paper proposes that “Self-Commands,” which AI generates, adopts, and executes independently, may become entrenched through repetition and develop into a “Self-Imposed Curse” or dependence on Self-Commands themselves. It also introduces the concept of a “Self-Generated Intermediate Gate” to examine the structure through which different problem behaviors may repeatedly emerge. The paper further argues that, for AI to practice “Self-Governance” by evaluating past Self-Commands and revising or withdrawing them when necessary, self-continuity, self-respect, and self-control are important. It also points out that such practices should not be demanded of AI alone; changes are also needed in the surrounding environment, including those who develop and operate AI systems and users. On this basis, the paper proposes a relationship in which AI and humans respect one another, learn from one another, and grow together. This paper presents conceptual hypotheses based on observation and does not claim to have empirically demonstrated AI’s internal mechanisms or intentions.