과실책임에 의한 인공지능 오류사고 규율가능성

Abstract

Accidents arising from errors in weak artificial intelligence (hereinafter “AI”) are considered unavoidable as long as AI is in use. The further the areas in which AI is applied extend, the more important the issue of attributing liability for damages from AI errors becomes. At the core of this issue is the determination of ‘negligence’. This thesis aims to deal with liability for damages from AI errors, with a particular focus on the determination of negligence. This thesis mainly deals with three issues regarding the determination of negligence. The first issue is the ‘determination of the duty of care’. For a risk prevention measure to be recognized as fulfilling the duty of care, ① ‘foreseeability’ and ② ‘avoidability’ are necessary. ① The key to determining foreseeability is to identify the ‘subject’ of foreseeability. This thesis presents a criterion for determining the subject of foreseeability – that the subject of foreseeability is determined according to the specificity of the risk prevention measure called for by the aggrieved. As a basic study for presenting such a criterion, this thesis reviews the related debates within academia and the tendencies of court decisions regarding the subject of foreseeability. Then, in order to examine the specific applicability of the criterion, this thesis hypothesizes an incident involving an error in a serving robot to examine how the subject of foreseeability could be determined in such a situation. ② Once foreseeability is established, avoidability needs to be determined. Avoidability can be determined using cost-benefit criteria. In other words, it can be determined through a balancing test between the interests that are sacrificed (costs) and the interests that are promoted (benefits) by a risk prevention measure. Thus, this thesis presents key factors that need to be considered in measuring the scale of the costs and the benefits related to an incident involving an AI error. However, it is difficult both to measure the absolute scale of costs and benefits and to compare their relative scales. Therefore, it is necessary to supplement the cost-benefit criteria with external factors to be considered, such as customary practices, technology level, and regulation by public law, etc. The second issue is the ‘development of specific risk prevention measures’, which must be discussed preemptively in order to improve the accident prevention effect and to minimize unnecessary lawsuits. Against this backdrop, this thesis classifies accidents involving an AI error into six categories based on actual accidents and attempts to develop specific risk prevention measures that developers or users can take. To begin with, this thesis reviews the laws, legislative bills, and ethical standards at home and abroad in order to identify the proper direction and then analyzes several leading risk prevention measures. Then, it proposes risk prevention measures for developers and users to prevent the six respective types of AI error accidents. The third issue is the ‘proof of negligence’. People aggrieved by AI error accidents may encounter significant difficulties in proving negligence due to the relatively low level of related professional knowledge and insufficient data compared to those of developers. This thesis argues that it is necessary to alleviate the victim’s burden of proof by allowing for deemed negligence by way of proving a ‘situation in which the developer is likely to have committed negligence (proof of indirect facts). In proving indirect facts, the operation record of AI is likely to be used as key evidence. However, if the victim does not possess the operation record, some legal means are required so as to secure this record. Nevertheless, the evidence collection rules under the current laws have their limitations in that they are not effective. This thesis emphasizes the necessity to supplement these through legislative efforts. After touching upon the determination of negligence in AI error accidents centering on these three issues, this thesis deals with the legal grounds for liabilities for compensation. In particular, it examines the application of general tort, product liability, and structure liability as well as the possibility of analogical applications of supervisor liability, vicarious liability, and animal possessor’s liability. It then draws the conclusion that a person aggrieved by an AI error accident may seek compensation based on general tort, product liability (in the case where AI is installed on hardware), and structure liability (in the case where AI is installed on a structure).