신뢰할 수 있는 인공지능의 이론적 고찰

There have been growing discussions on establishing desirable artificial intelligence (AI) governance to maximize benefits and minimize harm and risk of AI. Today, the concept of trustworthiness is at the heart of AI governance. However, unlike active practical discussions on the sub- components of trustworthiness, fundamental discussions are insufficient. This may make it difficult to systematically understand and critically reflect on the present and future of trustworthy AI governance and may hinder the creation of desirable legal policies. Based on the awareness of these problems, this paper examines the theories of AI trustworthiness and discusses their implications. In human society, the concept of trustworthiness focuses mainly on reducing fundamental uncertainty by exercising reciprocity based on good will. While this may play a significant role in traditional rule-based AI contexts, it is difficult to apply it to deep learning AI, where uncertainty related to results rather than intentions is mainly highlighted. Accordingly, it is necessary to establish a new trustworthiness theory based on the results. The establishment of a risk management system to reduce uncertainty is at the core of this result-based trustworthiness theory. In the case of technical risks, which have been traditionally discussed in the engineering field, various risk management methodologies have already been established. On the other hand, in the case of social risks created by deep learning AI, it is not easy to apply the existing quantitative risk management methodologies because they are related to human qualitative values. In response, two alternatives have been proposed: human-in-the-loop and technological advancement. Instead of acknowledging the supremacy of either one, we should have convergent thinking that embraces both.