인공지능의 법적 지위와 불법행위책임에 관한 검토

Artificial intelligence (AI), which is expected to serve as a key driving force behind technological advancement and social transformation in the future, is fundamentally based on two core components: “Algorithms”, processes for analyzing and reasoning from external data, and “Deep learning”, a mechanism that enables self-improvement through experience. From a cognitive science perspective, AI can be categorized according to its level of intelligence into “Weak AI”, “Strong AI”, and “Super AI”, each characterized by autonomy, unpredictability, decentralization, and opacity. Since the emergence of Sophia—a humanoid AI robot—in 2017, debates have arisen over whether artificial intelligence possesses sentience and self-awareness. In connection with the concept of strong AI, a central legal issue is whether AI should be granted legal personality or the capacity to hold rights. Opinions diverge:some scholars advocate for recognizing AI’s legal personality in a comprehensive or limited manner, while others oppose such recognition. However, under the current legal framework, granting legal personality to AI appears premature. In particular, with respect to tort liability, it remains difficult to conceptualize AI as a subject of responsibility. Accordingly, AI itself cannot bear liability for damages. Rather, it is more appropriate to impose liability on natural or juridical persons—traditional holders of legal personality—pursuant to existing principles of tort liability. Nevertheless, since it is difficult to regard AI as a rights-bearing legal person, the conventional fault-based liability system cannot be readily applied to damages caused by AI. Therefore, it becomes necessary to examine possible legal grounds and mechanisms for attributing liability based on either fault liability or strict (no-fault) liability in cases of harm caused by AI. Regarding fault liability, analogizing doctrines such as supervisory liability, employer’s liability, or keepers’ liability for animals appears inappropriate. However, under the doctrine of liability of possessors of structures (Article 758 of the Korean Civil Code), the possessor of an AI robot classified as weak AI may, in principle, bear primary liability for damages arising therefrom. As for strict liability, the applicability of owners’ liability for structures may be considered, but since the burden of proving the existence of a defect lies with the victim, its application would likely remain limited. With respect to product liability, although software is not presently classified as a “product” under the Product Liability Act, it may be regarded as such if considered “embedded software”, thereby potentially subjecting the manufacturer to liability. Concerning the liability of vehicle operators, an analogy might be drawn to impose such liability on AI owners or developers; however, extending it to end users seems inappropriate under current circumstances. Moreover, the traditional fault-based liability system—requiring the victim to prove negligence—faces inherent limitations when applied to damages caused by AI. Consequently, new principles of liability attribution have been proposed, notably strict liability and benefit liability. Of these, benefit liability, which imposes responsibility on those who enjoy the benefits derived from AI systems, appears more appropriate. Therefore, legislative responses grounded in the concept of benefit liability merit serious consideration as a future direction of legal reform.