고성능 인공지능의 능력위험 규율: 이론적 기반

As frontier AI performance has rapidly improved, new dimensions of risk have emerged that are difficult to manage with existing regulatory frameworks. This study aims to redefine the nature of AI risks and establish a theoretical foundation for rational governance. This paper categorizes AI risks into context risks and capability risks based on AI's characteristics of autonomy, adaptability, and inscrutability. Context risks are those understood in relation to specific application domains and contexts, characterized by context dependence, use attribution, and predictability. In contrast, capability risks are those understood based on AI's potential capabilities regardless of usage context, characterized by context independence, capability attribution, and uncertainty combined with irreversibility. Capability risks arise through three pathways: misuse, autonomy, and systemic accumulation, though the systemic accumulation pathway has more characteristics of complexity risks than capability risks. This study analyzes the debate between alarmist (principled and technical) and dissenting (principled, technical, and prioritarian) positions and examines the development patterns of existing AI technologies. Through the multimodal expansion and reasoning capability improvements of large language models, the environmental interaction capabilities of agent AI, the real-world extension of physical AI, the proliferation of high-performance open-source models, and the acceleration of AGI development attempts, this research confirms that capability risks are uncertain but real risks. This paper argues that the distinctive characteristics of capability risks necessitate new approaches distinct from existing context risk governance systems and establishes theoretical foundations across three dimensions. As a social foundation, it utilizes Luhmann's risk sociology to present the importance of communication and the necessity of pluralistic governance. As an ethical foundation, it demonstrates that virtue ethics is more effective than deontological and consequentialist ethics under uncertainty and presents the need for both developer ethics and machine ethics. As a legal theoretical foundation, it presents methods for building effective preparedness to resolve the dilemma between uncertainty and irreversibility.